Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
This guide covers migrating Messages API code. If you use Claude Managed Agents, no changes beyond updating the model name are required.
Automate your migration with the Claude API skill. In Claude Code, run /claude-api migrate to invoke the bundled Claude API skill. It works for any target model on this page:
/claude-api migrate this project to claude-opus-5The skill applies the model ID swap and, as needed, breaking parameter changes, prefill replacement, and effort calibration for your target model across your code base, then produces a checklist of items to verify manually. It asks you to confirm the migration scope (entire working directory, a subdirectory, or a specific file list) before editing any files. The skill also detects Amazon Bedrock and Claude Platform on AWS clients and adjusts model ID formats and feature changes for those platforms.
Claude Fable 5 is Anthropic's most capable widely released model, generally available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Claude Mythos 5 shares the same capabilities and is offered in limited availability to approved customers in Project Glasswing.
The baseline settings shared by claude-fable-5 and claude-mythos-5:
thinking configuration is required. Both thinking: {type: "disabled"} and manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) return a 400 error.invalid_request_error. Organizations with a ZDR arrangement should contact their Anthropic account team to discuss data retention configuration. Alternatively, you can configure data retention per workspace. See Model-specific data retention requirements for per-platform details.Where the two models diverge:
stop_reason: "refusal". Claude Mythos 5 does not include these classifiers. See Refusals and fallback.Claude Mythos 5 is the access-gated successor to Claude Mythos Preview, the invitation-only research preview. Claude Fable 5 is the generally available model with the same capabilities, and the changes in this section apply equally to both targets.
Migration is mostly drop-in. Claude Mythos 5 and Claude Fable 5 use the same Messages API and the same tool use patterns as Claude Mythos Preview, and token counts are roughly unchanged because all three models use the same tokenizer. The key changes to check are the features that are no longer available (listed in the next section) and thinking output. If you migrate to Claude Fable 5, also plan for safety classifier refusals, which Claude Mythos Preview and Claude Mythos 5 do not have; see Refusals and fallback.
For the Claude Mythos Preview retirement timeline, see Model deprecations.
model = "claude-mythos-preview" # Before
model = "claude-mythos-5" # After
# Or, for the generally available model with the same capabilities:
model = "claude-fable-5" # AfterExtended thinking and thinking token budgets: Manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) is not supported on claude-mythos-5 or claude-fable-5 and returns a 400 error. Adaptive thinking is always on: the model determines when and how much to think on each request, and no thinking configuration is required. thinking: {type: "disabled"} returns an error. budget_tokens has no direct replacement: thinking is adaptive, and the effort parameter is a separate output-level control, not a thinking budget.
Before (Claude Mythos Preview):
client.messages.create(
model="claude-mythos-preview",
max_tokens=16000,
thinking={"type": "enabled", "budget_tokens": 10000},
messages=[{"role": "user", "content": "..."}],
)After (Claude Mythos 5):
client.messages.create(
model="claude-mythos-5",
max_tokens=16000,
messages=[{"role": "user", "content": "..."}],
)The change for Claude Fable 5 is identical, with claude-fable-5 as the model name.
Assistant prefill: Prefilling the assistant message is not supported on claude-mythos-5 or claude-fable-5 and returns a 400 error, the same as on Claude Mythos Preview. Use system prompt instructions instead.
Thinking output: On claude-mythos-5 and claude-fable-5, the raw chain of thought is never returned, but thinking blocks still carry readable summarized text when thinking.display is set to summarized. Pass thinking blocks back unchanged when continuing a conversation on the same model. See Thinking output on Claude Fable 5 and Claude Mythos 5.
claude-mythos-5 and claude-fable-5 use the same tokenizer as claude-mythos-preview (the tokenizer introduced with Claude Opus 4.7). Token counts are roughly unchanged when migrating from claude-mythos-preview. Compared with models before Claude Opus 4.7, the same content can tokenize to roughly 30% more tokens, varying by content and workload shape.
/v1/messages/count_tokens returns roughly unchanged values for claude-mythos-5 and claude-fable-5 compared with claude-mythos-preview. Re-baseline cost and latency on your own workloads.
claude-mythos-preview to claude-mythos-5, or to claude-fable-5 for the generally available model.thinking: {type: "enabled", budget_tokens: N}). Adaptive thinking is always on, and no thinking field is required.thinking: {type: "disabled"} configuration. Disabling thinking returns an error on claude-mythos-5 and claude-fable-5.budget_tokens. It has no direct replacement: thinking is adaptive, and the effort parameter is a separate output-level control, not a thinking budget.thinking field treats it as display text only and passes thinking blocks back unchanged when continuing on the same model. thinking.display defaults to "omitted" on claude-mythos-5 and claude-fable-5, the same as on Claude Mythos Preview; set display: "summarized" to receive readable summaries. See Thinking output on Claude Fable 5 and Claude Mythos 5.thinking and redacted_thinking blocks from prior assistant turns first. Thinking blocks from claude-mythos-5 and claude-fable-5 are tied to the model that produced them, and models other than Claude Fable 5 and Claude Mythos 5 silently ignore them. Stripping keeps cross-model requests minimal and uniform.stop_reason: "refusal" and read the stop_details.category field. Claude Fable 5 runs safety classifiers that Claude Mythos Preview and Claude Mythos 5 do not have. See Refusals and fallback.claude-mythos-preview.Claude Fable 5 and Claude Mythos 5 use the same Messages API and the same tool use patterns as Claude Opus 5, with the same 1M token context window by default and the same 128k max output tokens. The prefill and sampling-parameter restrictions, and the thinking display behavior, carry over from Claude Opus 5 unchanged. The changes to check are always-on thinking, pricing, Priority Tier, and data retention.
model = "claude-opus-5" # Before
model = "claude-fable-5" # After
# Or, for the Project Glasswing model with the same capabilities:
model = "claude-mythos-5" # AfterThinking can no longer be disabled: On Claude Opus 5, thinking is on by default and can be turned off with thinking: {type: "disabled"} at an effort level of high or below. On claude-fable-5 and claude-mythos-5, adaptive thinking is always on, and thinking: {type: "disabled"} returns a 400 error at any effort level. Remove the thinking: {type: "disabled"} configuration and use lower effort levels to control token spend instead.
Pricing: Claude Fable 5 and Claude Mythos 5 are priced at $10 USD per million input tokens and $50 USD per million output tokens, compared with $5 USD and $25 USD for Claude Opus 5. See Claude pricing.
Priority Tier: Priority Tier is not supported on Claude Opus 5, so no existing traffic is affected. If your organization has a Priority Tier commitment, Claude Fable 5 supports it; Claude Mythos 5 does not.
Data retention: Claude Fable 5 and Claude Mythos 5 require 30-day data retention and are not available under zero data retention (ZDR) arrangements; both are designated Covered Models. See Model-specific data retention requirements.
claude-opus-5 to claude-fable-5 (or claude-mythos-5).thinking: {type: "disabled"} configuration; it returns a 400 error on claude-fable-5 and claude-mythos-5. Use lower effort levels to control token spend instead, and revisit max_tokens for workloads that ran with thinking disabled on Claude Opus 5.If your code is on Claude Opus 4.7 or earlier, first apply the relevant Migrating to Claude Opus 5 from-section for the API-level changes from your current model, then the remaining delta in this section.
Migration is mostly drop-in. Claude Fable 5 and Claude Mythos 5 use the same Messages API and the same tool use patterns as Claude Opus 4.8, with the same 1M token context window by default and the same 128k max output tokens. Token counts are roughly unchanged because the models use the same tokenizer. The key changes to check are always-on adaptive thinking, thinking output, safety classifier refusals (Claude Fable 5 only), and pricing.
model = "claude-opus-4-8" # Before
model = "claude-fable-5" # After
# Or, for the Project Glasswing model with the same capabilities:
model = "claude-mythos-5" # AfterThe items in this section describe the API and behavior differences worth checking after you swap the model ID. Except where noted, they apply equally to claude-fable-5 and claude-mythos-5.
Adaptive thinking is always on: Adaptive thinking is the only thinking mode on claude-fable-5 and claude-mythos-5. The model determines when and how much to think on each request, and no thinking configuration is required. thinking: {type: "disabled"} returns an error. Use the effort parameter to control thinking depth.
The behavior change to check: on Claude Opus 4.8, requests without a thinking field run without thinking; on claude-fable-5 and claude-mythos-5, those same requests run with adaptive thinking. max_tokens remains a hard limit on total output, thinking plus response text, so revisit it for workloads that ran without thinking on Claude Opus 4.8. See Cost control.
Before (Claude Opus 4.8):
client.messages.create(
model="claude-opus-4-8",
max_tokens=16000,
thinking={"type": "adaptive"},
output_config={"effort": "high"},
messages=[{"role": "user", "content": "..."}],
)After (Claude Fable 5):
client.messages.create(
model="claude-fable-5",
max_tokens=16000,
output_config={"effort": "high"},
messages=[{"role": "user", "content": "..."}],
)The change for Claude Mythos 5 is identical, with claude-mythos-5 as the model name.
Extended thinking and thinking budgets (unchanged): Manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) is not supported on claude-fable-5 or claude-mythos-5 and returns a 400 error, the same as on Claude Opus 4.8. budget_tokens has no direct replacement: thinking is adaptive, and the effort parameter is a separate output-level control, not a thinking budget.
Assistant prefill (unchanged): Prefilling the assistant message is not supported on claude-fable-5 or claude-mythos-5 and returns a 400 error, the same as on Claude Opus 4.8. Use system prompt instructions instead.
Thinking output: On claude-fable-5 and claude-mythos-5, the raw chain of thought is never returned, but thinking blocks still carry readable summarized text when thinking.display is set to summarized. Pass thinking blocks back unchanged when continuing a conversation on the same model. See Thinking output on Claude Fable 5 and Claude Mythos 5.
Safety classifiers and the refusal stop reason (Claude Fable 5 only): claude-fable-5 runs safety classifiers on requests and during response generation. Claude Mythos 5 does not include these classifiers. When a classifier declines a request, the Messages API returns stop_reason: "refusal" as a successful HTTP 200 response, not an error. The stop_details.category field reports which classifier fired, with categories such as "cyber", "bio", and "reasoning_extraction", or null when the refusal maps to no named category. See the refusal category table for the full set.
You are not billed for the input tokens of a request refused before any output is generated. When a classifier fires mid-stream, the input and already-streamed output are billed; discard the partial output.
To re-run refused requests on another model automatically, pass the opt-in fallbacks parameter, which is in beta on the Claude API. The parameter is not available on the Message Batches API or on Amazon Bedrock, Google Cloud, and Microsoft Foundry; on those three platforms, run the retry client-side or use the SDK refusal-fallback middleware. See Refusals and fallback.
Start at high effort: The effort parameter default remains high. On Claude Opus 4.8, the recommendation for coding and high-autonomy work is to set xhigh explicitly. On claude-fable-5 and claude-mythos-5, use high as the default for most tasks and reserve xhigh for the most capability-sensitive workloads. Lower effort settings still perform well and often exceed xhigh performance on prior models. Reduce effort if a task completes but takes longer than necessary. See Prompting Claude Fable 5.
Lower prompt caching minimum: The minimum cacheable prompt length on claude-fable-5 and claude-mythos-5 is 512 tokens, lower than the 1,024 tokens on Claude Opus 4.8. Prompts that were too short to cache on Claude Opus 4.8 can now create cache entries, with no code changes required. See Prompt caching for per-model minimums.
claude-fable-5 and claude-mythos-5 require 30-day data retention; on the Claude API, requests to claude-fable-5 that do not meet this requirement return a 400 invalid_request_error. Claude Opus 4.8 remains available under ZDR. See Model-specific data retention requirements.claude-opus-4-8 to claude-fable-5 (or claude-mythos-5).thinking: {type: "disabled"} configuration. Disabling thinking returns an error on claude-fable-5 and claude-mythos-5, and requests without a thinking field run with adaptive thinking.claude-fable-5 and claude-mythos-5.thinking field treats it as display text only and passes thinking blocks back unchanged when continuing on the same model. thinking.display defaults to "omitted" on claude-fable-5 and claude-mythos-5, the same as on Claude Opus 4.8; set display: "summarized" to receive readable summaries. See Thinking output on Claude Fable 5 and Claude Mythos 5.thinking and redacted_thinking blocks from prior assistant turns first. Thinking blocks from claude-fable-5 and claude-mythos-5 are tied to the model that produced them, and models other than Claude Fable 5 and Claude Mythos 5 silently ignore them. Stripping keeps cross-model requests minimal and uniform. The exception is redeeming a fallback credit, which requires the request body echoed under that feature's exact rules.stop_reason: "refusal" and read the stop_details.category field. To re-run refused requests on another model automatically, consider the opt-in fallbacks parameter (beta). See Refusals and fallback.effort setting. Start at high for most tasks, including workloads that ran at xhigh on Claude Opus 4.8.claude-opus-4-8; per-token pricing differs.Claude Opus 5 is a step-change improvement over Claude Opus 4.8, strong on deep reasoning, agentic and long-horizon tasks, and test-time compute scaling. For behavioral differences and model-specific prompting patterns, see Prompting Claude Opus 5.
Claude Opus 5 is a drop-in upgrade for Claude Opus 4.8 at the same pricing of $5 per million input tokens and $25 per million output tokens; see Claude pricing. There are two breaking changes for code already running on Claude Opus 4.8, covered under Breaking changes below. Claude Opus 5 supports the same set of features as Claude Opus 4.8, including the 1M token context window (the default, with no beta header), 128k max output tokens, adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and server-side and client-side tools, with two exceptions: web fetch is not available on Claude Opus 5, and Priority Tier is not supported on Claude Opus 5. See each tool page for model availability.
This section covers the delta from Claude Opus 4.8 only. If your code is on Claude Opus 4.7 or earlier, use the sections below instead: Migrating to Claude Opus 5 from Claude Opus 4.7 or Migrating to Claude Opus 5 from Claude Opus 4.6 and earlier Opus models. They include this delta plus the breaking changes from earlier models (sampling parameters rejected, manual extended thinking rejected, prefill removed, new tokenizer).
# Opus migration
model = "claude-opus-4-8" # Before
model = "claude-opus-5" # Afterclaude-opus-5 is a fixed model ID with no date suffix, the same scheme as claude-opus-4-8 and claude-sonnet-5.
Thinking on by default: On Claude Opus 4.8, requests without a thinking field run without thinking; on Claude Opus 5, the same requests run with adaptive thinking. max_tokens remains a hard limit on total output, thinking plus response text, so revisit it for workloads that ran without thinking on Claude Opus 4.8. To preserve the old behavior, pass thinking: {type: "disabled"}, subject to the effort cap in the next item; note that with thinking disabled the model can occasionally emit tool calls as plain text or include internal XML tags in its visible output, so prefer lower effort levels with thinking enabled where you can, and see Running with thinking disabled for mitigations where you can't.
Disabling thinking is capped at high effort: You can still turn thinking off with thinking: {type: "disabled"}, but only at an effort level of high or below. A request that combines thinking: {type: "disabled"} with effort xhigh or max returns a 400 error. Claude Opus 4.8 accepts this combination, so audit requests that disable thinking before you migrate.
The check is enforced on each request: every request's effort and thinking configuration is validated independently, so a request that raises effort to xhigh or max while thinking is disabled is rejected even if earlier requests in the conversation were accepted.
Before (accepted on Claude Opus 4.8, rejected on Claude Opus 5):
client.messages.create(
model="claude-opus-4-8",
max_tokens=16000,
thinking={"type": "disabled"},
output_config={"effort": "xhigh"},
messages=[{"role": "user", "content": "..."}],
)After (Claude Opus 5), either remove the thinking field to re-enable thinking:
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
output_config={"effort": "xhigh"}, # thinking is on by default
messages=[{"role": "user", "content": "..."}],
)or keep thinking disabled and lower the effort:
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
thinking={"type": "disabled"},
output_config={"effort": "high"}, # or "medium", "low"
messages=[{"role": "user", "content": "..."}],
)These are not required but will improve your experience:
Test max effort for capability-critical work: Claude Opus 5 supports the full set of effort levels (low, medium, high, xhigh, max). Where maximum capability matters more than token spend, test max effort. It can deliver gains on the most demanding tasks but may show diminishing returns from increased token usage and can be prone to overthinking on simpler ones. If you run at xhigh or max effort, set a large max_tokens so the model has room to think and act; start at 64k tokens and tune from there.
Consider automatic fallbacks: Claude Opus 5 ships with cybersecurity safety classifiers whose cyber-category refusals can fall back to Claude Opus 4.8. To re-run refused requests on another model automatically, consider the fallbacks parameter with the "default" mode (fallbacks: "default"), which selects a recommended fallback model based on the refusal category instead of a hand-maintained model list. Server-side fallback is in beta; the "default" mode requires the server-side-fallback-2026-07-01 beta header. See Refusals and fallback.
Cache shorter prompts: The minimum cacheable prompt length on Claude Opus 5 is 512 tokens, down from 1,024 tokens on Claude Opus 4.8. Prompts that were too short to cache on Claude Opus 4.8 can now create cache entries, with no code changes required. See Prompt caching for per-model minimums.
Change tools mid-conversation (beta): You can add or remove tools between turns of a conversation without invalidating prompt cache hits on earlier turns. Send the beta header mid-conversation-tool-changes-2026-07-01. This is useful for agentic workloads that expose tools progressively or retire them as a task advances; without it, a changed tool list invalidates the cached prefix.
Re-tune length and verbosity prompts: Default visible responses and written deliverables run longer on Claude Opus 5 than on Claude Opus 4.8, and lowering effort reduces thinking volume without reliably shortening the visible response. Prompt explicitly for conciseness or a target length instead. See Response length and verbosity and Written deliverable length.
Remove carried-over verification instructions and constrain scope: Claude Opus 5 verifies its own work without being told to, so remove explicit verification or self-check instructions carried over from prompts tuned for earlier models; leaving them in causes over-verification. For narrow tasks, constrain the task scope explicitly. In multi-agent frameworks, give explicit guidance on which scenarios warrant delegation or cap the number of subagents, because Claude Opus 5 delegates more readily than earlier models. See Task scope and over-verification and Controlling subagent spawning.
claude-opus-4-8 to claude-opus-5.thinking field: they run with thinking on Claude Opus 5. Revisit max_tokens, which remains a hard limit on total output (thinking plus response text), or pass thinking: {type: "disabled"} at effort high or below to preserve the old behavior. If you disable thinking, review Running with thinking disabled for the output artifacts that can appear and their prompting mitigations.thinking: {type: "disabled"} with effort xhigh or max returns a 400 error, enforced on each request. Re-enable thinking or lower the effort to high or below.effort setting: run a fresh effort sweep on your own evals rather than carrying over a setting tuned for an earlier model. low and medium effort are worth testing as cost and latency controls, and test max effort where maximum capability matters more than token spend. If you run at xhigh or max effort, raise max_tokens to at least 64k as a starting point.stop_reason: "refusal", and consider fallbacks: "default" (beta) to re-run refused requests on a recommended fallback model automatically.Claude Opus 5 should have strong out-of-the-box performance on existing Claude Opus 4.7 prompts and evals, at the same pricing of $5 per million input tokens and $25 per million output tokens. It supports the same set of features as Claude Opus 4.7, including the 1M token context window, 128k max output tokens, adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and server-side and client-side tools, with two exceptions: web fetch is not available on Claude Opus 5, and Priority Tier is not supported on Claude Opus 5. It also adds mid-conversation system messages and publicly documents refusal stop details.
If your code is on Claude Opus 4.6 or earlier, use Migrating to Claude Opus 5 from Claude Opus 4.6 and earlier Opus models instead. That section includes breaking changes (sampling parameters rejected, manual extended thinking rejected, new tokenizer) that the upgrade from Claude Opus 4.7 alone does not cover.
# Opus migration
model = "claude-opus-4-7" # Before
model = "claude-opus-5" # AfterThinking on by default: On Claude Opus 4.7, requests without a thinking field run without thinking; on Claude Opus 5, the same requests run with adaptive thinking. max_tokens remains a hard limit on total output, thinking plus response text, so revisit it for workloads that ran without thinking on Claude Opus 4.7. To preserve the old behavior, pass thinking: {type: "disabled"}, subject to the effort cap in the next item; note that with thinking disabled the model can occasionally emit tool calls as plain text or include internal XML tags in its visible output, so prefer lower effort levels with thinking enabled where you can, and see Running with thinking disabled for mitigations where you can't.
Disabling thinking is capped at high effort: You can turn thinking off with thinking: {type: "disabled"}, but only at an effort level of high or below. A request that combines thinking: {type: "disabled"} with effort xhigh or max returns a 400 error. Claude Opus 4.7 accepts this combination, so audit requests that disable thinking before you migrate.
The check is enforced on each request: every request's effort and thinking configuration is validated independently, so a request that raises effort to xhigh or max while thinking is disabled is rejected even if earlier requests in the conversation were accepted.
Before (accepted on Claude Opus 4.7, rejected on Claude Opus 5):
client.messages.create(
model="claude-opus-4-7",
max_tokens=16000,
thinking={"type": "disabled"},
output_config={"effort": "xhigh"},
messages=[{"role": "user", "content": "..."}],
)After (Claude Opus 5), either remove the thinking field to run with thinking:
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
output_config={"effort": "xhigh"}, # thinking is on by default
messages=[{"role": "user", "content": "..."}],
)or keep thinking disabled and lower the effort:
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
thinking={"type": "disabled"},
output_config={"effort": "high"}, # or "medium", "low"
messages=[{"role": "user", "content": "..."}],
)The following items are not breaking changes; they describe behavior differences worth checking after you swap the model ID.
Sampling parameters (unchanged): Setting temperature, top_p, or top_k to a non-default value returns a 400 error on Claude Opus 5, the same as on Claude Opus 4.7. The SDK request types still define these fields for compatibility with earlier models, so code that sets them type-checks, but the API rejects the request server-side. If you removed these parameters when migrating to Opus 4.7, no further changes are needed.
Effort default is high: The effort parameter default on Claude Opus 5 is high on the Claude API and Claude Code. If you already set effort explicitly, your setting is unchanged.
Effort levels recalibrated: The token allocation behind each effort level changes on Claude Opus 5 compared to Claude Opus 4.7, and Claude Opus 5 supports the full set of effort levels (low, medium, high, xhigh, max). Run a fresh effort sweep on your own evals rather than carrying over a setting tuned for Claude Opus 4.7. low and medium effort are worth testing as cost and latency controls, and test max effort where maximum capability matters more than token spend. If you run at xhigh or max effort, set a large max_tokens so the model has room to think and act; start at 64k tokens and tune from there. See Effort.
1M context window is the default: Claude Opus 5 serves the full 1M token context window by default with no beta header and no long-context premium. If your client passes a context-window beta header for compatibility with older models, you can remove it on Claude Opus 5.
Mid-conversation system messages: Claude Opus 5 accepts role: "system" messages immediately after a user turn in the messages array (subject to placement rules). Use the top-level system field for instructions that apply from the start. Claude Opus 4.7 rejects role: "system" in messages with a 400 error. If you maintain code paths that rebuild the full message history to update instructions, you can simplify them and preserve prompt cache hits on earlier turns.
Refusal stop details: The stop_details object on refusal responses (available since Claude Opus 4.7) is now publicly documented. When the model declines a request, it identifies the category of refusal, in addition to the existing refusal stop reason. No beta header is required, and there is no opt-out. See Handling stop reasons.
Lower prompt caching minimum: The minimum cacheable prompt length on Claude Opus 5 is 512 tokens, lower than on Claude Opus 4.7. Prompts that were too short to cache on Claude Opus 4.7 can now create cache entries, with no code changes required. See Prompt caching for per-model minimums.
Fast mode: Claude Opus 5 supports fast mode (research preview); fast mode is not available on Claude Opus 4.7, where requests with speed: "fast" return an error. The speed: "fast" parameter and fast-mode-2026-02-01 beta header work unchanged on Claude Opus 5.
These are not required but will improve your experience:
Consider automatic fallbacks: Claude Opus 5 ships with cybersecurity safety classifiers whose cyber-category refusals can fall back to Claude Opus 4.8. To re-run refused requests on another model automatically, consider the fallbacks parameter with the "default" mode (fallbacks: "default"), which selects a recommended fallback model based on the refusal category instead of a hand-maintained model list. Server-side fallback is in beta; the "default" mode requires the server-side-fallback-2026-07-01 beta header. See Refusals and fallback.
Change tools mid-conversation (beta): You can add or remove tools between turns of a conversation without invalidating prompt cache hits on earlier turns. Send the beta header mid-conversation-tool-changes-2026-07-01. This is useful for agentic workloads that expose tools progressively or retire them as a task advances; without it, a changed tool list invalidates the cached prefix.
Re-tune length and verbosity prompts: Default visible responses and written deliverables run longer on Claude Opus 5 than on earlier Opus models, and lowering effort reduces thinking volume without reliably shortening the visible response. Prompt explicitly for conciseness or a target length instead. See Response length and verbosity and Written deliverable length.
Remove carried-over verification instructions and constrain scope: Claude Opus 5 verifies its own work without being told to, so remove explicit verification or self-check instructions carried over from prompts tuned for earlier models; leaving them in causes over-verification. For narrow tasks, constrain the task scope explicitly. In multi-agent frameworks, give explicit guidance on which scenarios warrant delegation or cap the number of subagents, because Claude Opus 5 delegates more readily than earlier models. See Task scope and over-verification and Controlling subagent spawning.
claude-opus-4-7 to claude-opus-5 (or update aliases).thinking field: they run with thinking on Claude Opus 5. Revisit max_tokens, which remains a hard limit on total output (thinking plus response text), or pass thinking: {type: "disabled"} at effort high or below to preserve the old behavior. If you disable thinking, review Running with thinking disabled for the output artifacts that can appear and their prompting mitigations.thinking: {type: "disabled"} with effort xhigh or max returns a 400 error, enforced on each request. Re-enable thinking or lower the effort to high or below.effort setting: run a fresh effort sweep on your own evals rather than carrying over a setting tuned for Claude Opus 4.7. Test low and medium effort as cost and latency controls, and max effort where maximum capability matters more than token spend. If you run at xhigh or max effort, raise max_tokens to at least 64k as a starting point.stop_details on refusals (available since Claude Opus 4.7; now publicly documented), and consider fallbacks: "default" (beta) to re-run refused requests on a recommended fallback model automatically.speed: "fast" and the fast-mode-2026-02-01 beta header work unchanged on Claude Opus 5.Claude Opus 5 should have strong out-of-the-box performance on existing Claude Opus 4.6 prompts and evals at the same pricing, but there are a handful of behavioral and API changes worth knowing about as you migrate. Most of these changes took effect in Claude Opus 4.7; two more, thinking on by default and an effort cap on disabling thinking, take effect on Claude Opus 5. All of them are covered below, so this section is complete for code coming straight from Claude Opus 4.6. Claude Opus 5 supports the same set of features as Claude Opus 4.6, including:
Two exceptions: web fetch is not available on Claude Opus 5, and Priority Tier is not supported on Claude Opus 5.
# Opus migration
model = "claude-opus-4-6" # Before
model = "claude-opus-5" # AfterExtended thinking removed: thinking: {type: "enabled", budget_tokens: N} is no longer supported on Claude Opus 4.7 or later models and returns a 400 error. Switch to adaptive thinking (thinking: {type: "adaptive"}) and use the effort parameter to control thinking depth. On Claude Opus 5, adaptive thinking is on by default: thinking: {type: "adaptive"} is valid and equivalent to omitting the thinking field entirely (see the next item).
Before (Claude Opus 4.6):
client.messages.create(
model="claude-opus-4-6",
max_tokens=16000,
thinking={"type": "enabled", "budget_tokens": 10000},
messages=[{"role": "user", "content": "..."}],
)After (Claude Opus 5):
client.messages.create(
model="claude-opus-5",
max_tokens=16000,
thinking={"type": "adaptive"},
output_config={"effort": "high"}, # or "max", "xhigh", "medium", "low"
messages=[{"role": "user", "content": "..."}],
)Adaptive thinking is steerable through prompting and the effort parameter; see Choosing an effort level.
Thinking on by default: On Claude Opus 4.6 and Claude Opus 4.7, requests without a thinking field run without thinking; on Claude Opus 5, the same requests run with adaptive thinking. max_tokens remains a hard limit on total output, thinking plus response text, so revisit it for workloads that ran without thinking. To preserve the old behavior, pass thinking: {type: "disabled"}, subject to the effort cap in the next item; note that with thinking disabled the model can occasionally emit tool calls as plain text or include internal XML tags in its visible output, so prefer lower effort levels with thinking enabled where you can, and see Running with thinking disabled for mitigations where you can't.
Disabling thinking is capped at high effort: You can turn thinking off with thinking: {type: "disabled"}, but only at an effort level of high or below. A request that combines thinking: {type: "disabled"} with effort xhigh or max returns a 400 error on Claude Opus 5, enforced on each request. Audit requests that disable thinking before you migrate: re-enable thinking or lower the effort to high or below.
Sampling parameters removed: Setting temperature, top_p, or top_k to any non-default value on Claude Opus 4.7 or later models, including Claude Opus 5, returns a 400 error. The safest migration path is to omit these parameters entirely from request payloads. Prompting is the recommended way to guide model behavior on Claude Opus 5. If you were using temperature = 0 for determinism, note that it never guaranteed identical outputs on prior models.
Thinking content omitted by default: Thinking blocks still appear in the response stream on Claude Opus 4.7 and later models, but their thinking field is empty unless you explicitly opt in. This is a silent change from Claude Opus 4.6, where the default was to return summarized thinking text. To restore summarized thinking content, set thinking.display to "summarized":
thinking = {
"type": "adaptive",
"display": "summarized",
}The default is "omitted" on Claude Opus 4.7 and later models. If your product streams reasoning to users, the new default appears as a long pause before output begins; set display: "summarized" to restore visible progress during thinking. See Controlling thinking display for details.
Updated token counting: Claude Opus 4.7 introduced a new tokenizer, which later Opus models, including Claude Opus 5, also use. It contributes to improved performance on a wide range of tasks, and it may use roughly 1x to 1.35x as many tokens when processing text compared to models before Claude Opus 4.7 (up to ~35% more, varying by content).
/v1/messages/count_tokens returns a different number of tokens for Claude Opus 5 than it did for Claude Opus 4.6. Token efficiency can vary by workload shape.
Prompting interventions, task_budget, and effort can help control costs and ensure appropriate token usage. These controls may trade off model intelligence. Update your max_tokens parameters to give additional headroom, including compaction triggers. Claude Opus 5 provides a 1M context window at standard API pricing with no long-context premium.
Prefill removal (carried over from Opus 4.6): Prefilling assistant messages returns a 400 error on Claude Opus 4.7 and later models, including Claude Opus 5. Use structured outputs, system prompt instructions, or output_config.format instead.
The effort parameter allows you to tune Claude's intelligence versus token spend, trading off capability for faster speed and lower costs. Claude Opus 5 supports the full set of effort levels and defaults to high. Run a fresh effort sweep on your own evals rather than carrying over a setting tuned for an earlier model:
max: Can deliver gains on the most demanding tasks but may show diminishing returns from increased token usage and can be prone to overthinking on simpler ones. Test it where maximum capability matters more than token spend.xhigh: Extended capability for long-running agentic and coding work that needs more depth than the default.high: The default. Balances token usage and intelligence for most tasks.medium: Cost-saving step-down from the default, worth testing as a cost and latency control.low: Most efficient. Reserve for short, scoped tasks and latency-sensitive workloads.If you run at xhigh or max effort, set a large max_tokens so the model has room to think and act; start at 64k tokens and tune from there. Effort is more important for this model than for any prior Opus. Experiment with it actively when you upgrade.
Claude Opus 4.7 introduced several behavioral differences from Claude Opus 4.6 that are not API breaking changes but may require prompt updates or scaffolding removal. They carry forward to Claude Opus 5, with the adjustments noted below.
Response length varies by use case: Claude Opus 4.7 calibrates response length to how complex it judges the task to be, rather than defaulting to a fixed verbosity. This usually means shorter answers on simple lookups and much longer ones on open-ended analysis.
If your product depends on a certain style or verbosity of output, you may need to tune your prompts. For example, to decrease verbosity, add: "Provide concise, focused responses. Skip non-essential context, and keep examples minimal." If you see specific kinds of over-explaining, add targeted instructions in your prompt to prevent them.
Positive examples showing how Claude can communicate with the appropriate level of concision tend to be more effective than negative examples or instructions that tell the model what not to do. On Claude Opus 5, default visible responses and written deliverables run longer than on earlier Opus models, and lowering effort reduces thinking volume without reliably shortening the visible response; prompt explicitly for conciseness or a target length. See Response length and verbosity.
More literal instruction following: Claude Opus 4.7 interprets prompts more literally and explicitly than Claude Opus 4.6, particularly at lower effort levels. It does not silently generalize an instruction from one item to another, and it does not infer requests you didn't make. The upside of this literalism is precision and less thrash. It generally performs better for API use cases with carefully tuned prompts, structured extraction, and pipelines where you want predictable behavior. A prompt and harness review may be especially helpful for migration to Claude Opus 5.
More direct tone: As with any new model, prose style on long-form writing may shift. Claude Opus 4.7 is more direct and opinionated, with less validation-forward phrasing and fewer emoji than Claude Opus 4.6's warmer style. If your product relies on a specific voice, re-evaluate style prompts against the new baseline.
Built-in progress updates in agentic traces: Claude Opus 4.7 provides more regular, higher-quality updates to the user throughout long agentic traces. If you've added scaffolding to force interim status messages ("After every 3 tool calls, summarize progress"), try removing it. If you find that the length or contents of Claude Opus 4.7's user-facing updates are not well-calibrated to your use case, explicitly describe what these updates should look like in the prompt and provide examples.
Subagent spawning changed: Claude Opus 4.7 tends to spawn fewer subagents by default than Claude Opus 4.6, while Claude Opus 5 delegates to subagents more readily than earlier models. The behavior is steerable through prompting in either direction; give explicit guidance around when subagents are desirable, or cap the number of subagents. See Controlling subagent spawning.
Stricter effort calibration: Meaningfully changing from Claude Opus 4.6, Claude Opus 4.7 respects effort levels strictly, especially at the low end. At low and medium, the model scopes its work to what was asked rather than doing more than requested.
This is good for latency and cost, but on moderately complex tasks running at low effort there is some risk of under-thinking. If you observe shallow reasoning on complex problems, raise effort to high or xhigh rather than prompting around it.
If you need to keep effort at low for latency, add targeted guidance: "This task involves multistep reasoning. Think carefully through the problem before responding." See Recommended effort levels for Claude Opus 4.7.
Fewer tool calls by default: Claude Opus 4.7 has a tendency to use tools less often than Claude Opus 4.6 and to use reasoning more. This produces better results in most cases.
To increase tool usage, raise the effort setting. high or xhigh effort settings show substantially more tool usage in agentic search and coding. You can also adjust your prompt to explicitly instruct the model about when and how to properly use its tools.
Real-time cybersecurity safeguards: Newly added in Claude Opus 4.7, requests that involve prohibited or high-risk topics may lead to refusals. For legitimate security work such as penetration testing, vulnerability research, or red-teaming, apply to the Cyber Verification Program to request reduced restrictions. See Safeguards, warnings, and appeals for background.
High-resolution image support: Claude Opus 4.7 is the first Claude model with high-resolution image support. Maximum image resolution is 2,576 pixels on the long edge, up from 1,568 pixels on prior models. This unlocks gains on vision-heavy workloads and is particularly valuable for computer use, screenshot understanding, and document analysis.
High-resolution support is automatic and requires no beta header or client-side opt-in. Two things to plan for:
max_tokens and cost expectations for image-heavy workloads, or downsample before sending if you do not need the additional fidelity.See High-resolution image support on Claude Opus 4.7 for details.
These are not required but will improve your experience:
Re-evaluate max_tokens: Because the same text produces a higher token count on Claude Opus 4.7 and later models, update your max_tokens parameters to give additional headroom, including compaction triggers. Prompting interventions, task_budget, and effort can help control costs and ensure appropriate token usage.
Audit token-count expectations: Any code path that estimates tokens client-side or assumes a fixed token-to-character ratio should be re-tested against Claude Opus 5. Use the Token counting endpoint to verify.
Adopt task budgets (beta): Claude Opus 4.7 introduces task budgets. These budgets let you inform Claude how many tokens it has for a full agentic loop, including thinking, tool calls, tool results, and final output. The model sees a running countdown and uses it to prioritize work and finish the task gracefully as the budget is consumed. To use, set the beta header task-budgets-2026-03-13 and add the following to your output config:
output_config = {
"effort": "high",
"task_budget": {"type": "tokens", "total": 128000},
}You may need to experiment with different task budgets for your use case. If the model is given a task budget that is too restrictive, it may complete the task less thoroughly, referencing its budget as the constraint.
For open-ended agentic tasks where quality matters more than speed, do not set a task budget. Reserve task budgets for workloads where you need the model to scope its work to a token allowance. The minimum value for a task budget is 20k tokens.
A task budget is not a hard cap; it's a suggestion that the model is aware of. It differs from max_tokens:
task_budget: an advisory cap across the full agentic loop. The model sees it and uses it to pace itself.max_tokens: a hard per-request ceiling on generated tokens. It is not passed to the model, so the model is not aware of it.Use task_budget when you want the model to self-moderate, and max_tokens as a hard ceiling to cap usage.
Set a large max_tokens at max or xhigh effort: If you are running Claude Opus 4.7 or a later model at max or xhigh effort, set a large max output token budget so the model has room to think and act across its subagents and tool calls. Start at 64k tokens and tune from there.
Downsample images if high resolution is unnecessary: Claude Opus 4.7 and later models support images up to 2576px / 3.75MP. High-res images use more tokens. If the additional image fidelity is unnecessary, downsample images before sending to Claude to avoid token-usage increases. See Images and vision.
Consider automatic fallbacks: Claude Opus 5 ships with cybersecurity safety classifiers whose cyber-category refusals can fall back to Claude Opus 4.8. To re-run refused requests on another model automatically, consider the fallbacks parameter with the "default" mode (fallbacks: "default"), which selects a recommended fallback model based on the refusal category instead of a hand-maintained model list. Server-side fallback is in beta; the "default" mode requires the server-side-fallback-2026-07-01 beta header. See Refusals and fallback.
Cache shorter prompts: The minimum cacheable prompt length on Claude Opus 5 is 512 tokens, lower than on earlier Opus models. Prompts that were too short to cache can now create cache entries, with no code changes required. See Prompt caching for per-model minimums.
Change tools mid-conversation (beta): You can add or remove tools between turns of a conversation without invalidating prompt cache hits on earlier turns. Send the beta header mid-conversation-tool-changes-2026-07-01. This is useful for agentic workloads that expose tools progressively or retire them as a task advances; without it, a changed tool list invalidates the cached prefix.
Remove carried-over verification instructions and constrain scope: Claude Opus 5 verifies its own work without being told to, so remove explicit verification or self-check instructions carried over from prompts tuned for earlier models; leaving them in causes over-verification. For narrow tasks, constrain the task scope explicitly. See Task scope and over-verification.
claude-opus-4-6 to claude-opus-5 (or update aliases).temperature, top_p, and top_k from request payloads.thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} plus the effort parameter, or remove the thinking field entirely; adaptive thinking is on by default on Claude Opus 5.thinking field: they run with thinking on Claude Opus 5. Revisit max_tokens, which remains a hard limit on total output (thinking plus response text), or pass thinking: {type: "disabled"} at effort high or below to preserve the old behavior.thinking: {type: "disabled"} with effort xhigh or max returns a 400 error, enforced on each request. Re-enable thinking or lower the effort to high or below.max_tokens to account for the updated tokenization.xhigh or max effort, raise max_tokens to at least 64k as a starting point.stop_reason: "refusal", and consider fallbacks: "default" (beta) to re-run refused requests on a recommended fallback model automatically.If you are migrating from Claude Opus 4.5, Opus 4.1 (deprecated), or an earlier model directly to Claude Opus 5, apply all of the changes earlier in this section plus the cumulative changes below, which took effect between Opus 4.5 and Opus 4.7. If you are migrating from Opus 4.6, the changes earlier in this section are all you need.
# Opus migration
model = "claude-opus-4-5" # Before
model = "claude-opus-5" # AfterPrefill removal is covered in the breaking changes for migrating from Claude Opus 4.6.
Tool parameter quoting: Claude Opus 4.6 and later models may produce slightly different JSON string escaping in tool call arguments (for example, different handling of Unicode escapes or forward slash escaping). If you parse tool call input as a raw string rather than using a JSON parser, verify your parsing logic. Standard JSON parsers (such as json.loads() or JSON.parse()) handle these differences automatically.
These changes improve your experience on Claude Opus 4.7 and later models. Items marked (required on Opus 4.7) were optional recommendations when Opus 4.6 launched but are now mandatory; the rest remain recommended.
Migrate to adaptive thinking (required on Opus 4.7): thinking: {type: "enabled", budget_tokens: N} returns a 400 error on Claude Opus 4.7 and later models. Switch to thinking: {type: "adaptive"} and use the effort parameter to control thinking depth; on Claude Opus 5, thinking: {type: "adaptive"} is equivalent to omitting the thinking field, which runs with adaptive thinking by default. See Thinking.
response = client.beta.messages.create(
model="claude-opus-4-5",
max_tokens=16000,
thinking={"type": "enabled", "budget_tokens": 32000},
betas=["interleaved-thinking-2025-05-14"],
messages=[{"role": "user", "content": "Your prompt here"}],
)Note that the migration also moves from client.beta.messages.create to client.messages.create. Adaptive thinking and effort are GA features and do not require the beta SDK namespace or any beta headers.
Remove effort beta header: The effort parameter is now GA. Remove betas=["effort-2025-11-24"] from your requests.
Remove fine-grained tool streaming beta header: Fine-grained tool streaming is now GA. Remove betas=["fine-grained-tool-streaming-2025-05-14"] from your requests.
Remove interleaved thinking beta header: Adaptive thinking automatically enables interleaved thinking on Claude Opus 4.7, Opus 4.6, and Sonnet 4.6. Remove betas=["interleaved-thinking-2025-05-14"] from your requests. The header is still functional on Sonnet 4.6 with manual extended thinking, but manual mode is deprecated.
Migrate to output_config.format: If using structured outputs, update output_format={...} to output_config={"format": {...}}. The old parameter remains functional but is deprecated and will be removed in a future model release.
If you're migrating from Opus 4.1 (deprecated) or earlier models directly to Claude Opus 5, apply all of the changes earlier in this section, plus the additional changes in this sub-section.
# From Opus 4.1
model = "claude-opus-4-1-20250805" # Before
model = "claude-opus-5" # After
# From Sonnet 3.7
model = "claude-3-7-sonnet-20250219" # Before
model = "claude-opus-5" # AfterRemove sampling parameters
This is a breaking change when migrating from Claude 3.x models.
Starting with Claude Opus 4.7, setting temperature, top_p, or top_k to any non-default value returns a 400 error. The safest migration path is to omit these parameters entirely from requests, and to use prompting to guide the model's behavior. If you were using temperature = 0 for determinism, note that it never guaranteed identical outputs.
# Before - This will error in Claude 4+ models
response = client.messages.create(
model="claude-3-7-sonnet-20250219",
temperature=0.7,
top_p=0.9, # Non-default sampling params return 400 on Opus 4.7
# ...
)
# After
response = client.messages.create(
model="claude-opus-5",
# ...
)Update tool versions
This is a breaking change when migrating from Claude 3.x models.
Update to the latest tool versions. Remove any code using the undo_edit command.
# Before
tools = [{"type": "text_editor_20250124", "name": "str_replace_editor"}]
# After
tools = [{"type": "text_editor_20250728", "name": "str_replace_based_edit_tool"}]text_editor_20250728 and str_replace_based_edit_tool. See Text editor tool documentation for details.code_execution_20260521. See Code execution tool documentation for migration instructions.Handle the refusal stop reason
Update your application to handle refusal stop reasons:
response = client.messages.create(...)
if response.stop_reason == "refusal":
# Handle refusal appropriately
passHandle the model_context_window_exceeded stop reason
Claude 4.5+ models return a model_context_window_exceeded stop reason when generation stops because of hitting the context window limit, rather than the requested max_tokens limit. Update your application to handle this new stop reason:
response = client.messages.create(...)
if response.stop_reason == "model_context_window_exceeded":
# Handle context window limit appropriately
passVerify tool parameter handling (trailing newlines)
Claude 4.5+ models preserve trailing newlines in tool call string parameters that were previously stripped. If your tools rely on exact string matching against tool call parameters, verify your logic handles trailing newlines correctly.
Update your prompts for behavioral changes
Claude 4+ models have a more concise, direct communication style and require explicit direction. Review prompting best practices for optimization guidance.
token-efficient-tools-2025-02-19 and output-128k-2025-02-19. All Claude 4+ models have built-in token-efficient tool use and these headers have no effect.claude-opus-5output_config.format insteadthinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} plus the effort parameter (returns 400 on Opus 4.7)effort-2025-11-24 beta header (effort is now GA)fine-grained-tool-streaming-2025-05-14 beta headerinterleaved-thinking-2025-05-14 beta header (adaptive thinking enables interleaved thinking automatically)output_format to output_config.format (if applicable)temperature, top_p, and top_k (non-default values return 400 on Opus 4.7)text_editor_20250728, code_execution_20260521)refusal stop reasonmodel_context_window_exceeded stop reasontoken-efficient-tools-2025-02-19, output-128k-2025-02-19)Claude Opus 5 and Claude Sonnet 5 share the same API surface: both run with adaptive thinking on by default, both default the effort parameter to high on the Claude API and Claude Code, both serve a 1M token context window by default with 128k max output tokens, and neither supports Priority Tier. Manual extended thinking and non-default sampling parameters return a 400 error on both models, as does assistant prefill.
model = "claude-sonnet-5" # Before
model = "claude-opus-5" # AfterPricing: Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. For Claude Sonnet 5, introductory pricing of $2/$10 per million input/output tokens is in effect through August 31, 2026, after which the standard pricing of $3/$15 takes effect. See Claude pricing for complete pricing.
Disabling thinking is capped at high effort: On Claude Sonnet 5, thinking: {type: "disabled"} is accepted at any effort level. On Claude Opus 5, it is accepted only at an effort level of high or below; a request that combines thinking: {type: "disabled"} with effort xhigh or max returns a 400 error, enforced on each request. Audit requests that disable thinking before you migrate.
Mid-conversation system messages: Claude Opus 5 accepts role: "system" messages immediately after a user turn in the messages array (subject to placement rules); Claude Sonnet 5 does not. If you maintain code paths that rebuild the full message history to update instructions, you can simplify them and preserve prompt cache hits on earlier turns.
Web fetch is not available: The web fetch tool is available on Claude Sonnet 5 but not on Claude Opus 5.
claude-sonnet-5 to claude-opus-5.thinking: {type: "disabled"} with effort xhigh or max returns a 400 error on Claude Opus 5. Re-enable thinking or lower the effort to high or below.Claude Sonnet 5 offers the best combination of speed and intelligence in the Claude model family. It builds on Claude Sonnet 4.6.
Claude Sonnet 5 is a drop-in upgrade for Claude Sonnet 4.6. Introductory pricing of $2/$10 USD per million input/output tokens is in effect through August 31, 2026, after which the standard pricing of $3/$15 USD per million input/output tokens will take effect; see Pricing for details. There are two breaking API changes for code already running on Claude Sonnet 4.6: manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) and sampling parameters (temperature, top_p, top_k) set to non-default values are no longer accepted and return a 400 error. Use adaptive thinking with the effort parameter instead. Claude Sonnet 5 supports the same set of features as Claude Sonnet 4.6, including the 1M token context window, adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and the full set of server-side and client-side tools. Priority Tier is not available on Claude Sonnet 5. Claude Sonnet 5 also uses a new tokenizer.
If your code is on Claude Sonnet 4.5 or earlier, also apply Migrating to Claude Sonnet 5 from Claude Sonnet 4.5 and earlier Sonnet models. Those steps include breaking changes (assistant message prefilling rejected, tool parameter JSON escaping differences) that this section alone does not cover.
# Sonnet migration
model = "claude-sonnet-4-6" # Before
model = "claude-sonnet-5" # AfterItems 4 and 5 in the following list are breaking changes. max_tokens remains a hard limit on total output (thinking plus response text), so revisit it for workloads that ran without thinking on Claude Sonnet 4.6.
New tokenizer: Claude Sonnet 5 uses a new tokenizer. The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6. The exact increase depends on the content. Requests, responses, and streaming events keep the same shape, and no code changes are required, but anything you measure or budget in tokens shifts: usage fields and token counting results for the same text are higher, the 1M token context window holds less text, and a max_tokens limit tuned for Claude Sonnet 4.6 may truncate equivalent output. Per-token pricing is unchanged, so the cost of an equivalent request can differ. Re-run token counting against Claude Sonnet 5 rather than reusing counts measured against earlier models.
128k max output tokens (unchanged): Claude Sonnet 5 supports up to 128k output tokens, the same as Claude Sonnet 4.6. Existing max_tokens values remain valid. Account for the new tokenizer when sizing them.
Assistant message prefilling (unchanged): Prefilling the assistant message returns a 400 error on Claude Sonnet 5, the same as on Claude Sonnet 4.6. If you removed prefill when migrating to Claude Sonnet 4.6, no further changes are needed. Use structured outputs, system prompt instructions, or output_config.format instead.
Adaptive thinking on by default: On Claude Sonnet 4.6, requests without a thinking field run without thinking; on Claude Sonnet 5, the same requests run with adaptive thinking. To turn thinking off, pass thinking: {type: "disabled"}. Manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) is not supported and returns a 400 error. Use the effort parameter (default high) to control thinking depth.
Adaptive thinking is on by default for Claude Sonnet 5. The thinking field is shown explicitly here to set display: "summarized"; if you omit thinking, Claude Sonnet 5 omits thinking content from the response by default. For per-model defaults, see Configurations each model rejects.
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-sonnet-5",
max_tokens=16000,
thinking={"type": "adaptive", "display": "summarized"},
output_config={"effort": "high"},
messages=[
{
"role": "user",
"content": "Are there an infinite number of prime numbers such that n mod 4 == 3?",
}
],
)
# The response contains summarized thinking blocks and text blocks
for block in response.content:
match block.type:
case "thinking":
print(f"\nThinking summary: {block.thinking}")
case "text":
print(f"\nResponse: {block.text}")Sampling parameters removed: Sampling parameters (temperature, top_p, top_k) set to a non-default value are not accepted and return a 400 error.
Cybersecurity safeguards: Claude Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards. Requests that involve prohibited or high-risk cybersecurity topics may be refused. Refusals return as a successful HTTP 200 response with stop_reason: "refusal", not an error. See Safeguards, warnings, and appeals for background.
claude-sonnet-4-6 to claude-sonnet-5.max_tokens limits sized close to your expected output length, and raise them up to the 128k maximum (unchanged from Claude Sonnet 4.6) where useful.thinking: {type: "enabled", budget_tokens: N} configuration (returns a 400 error). Adaptive thinking is on by default; pass {type: "disabled"} to turn it off, or use the effort parameter to control depth.temperature, top_p, and top_k parameters set to non-default values (they return a 400 error on Claude Sonnet 5).stop_reason: "refusal" if your workload may touch cybersecurity topics.max_tokens for workloads that previously ran without thinking.If you are migrating from Claude Sonnet 4.5 or an earlier Sonnet model directly to Claude Sonnet 5, apply the Migrating to Claude Sonnet 5 from Claude Sonnet 4.6 changes plus the changes in this section.
Claude Sonnet 5 defaults to an effort level of high, in contrast to Sonnet 4.5 which had no effort parameter. Consider adjusting the effort parameter as you migrate. If not explicitly set, you may experience higher latency with the default effort level.
Prefilling assistant messages is no longer supported
This is a breaking change when migrating from Sonnet 4.5 or earlier.
Prefilling assistant messages returns a 400 error on Claude Sonnet 4.6 and later models, including Claude Sonnet 5. Use structured outputs, system prompt instructions, or output_config.format instead.
Common prefill use cases and migrations:
Controlling output formatting (forcing JSON/YAML output): Use structured outputs or tools with enum fields for classification tasks.
Eliminating preambles (removing "Here is..." phrases): Add direct instructions in the system prompt: "Respond directly without preamble. Do not start with phrases like 'Here is...', 'Based on...', etc."
Avoiding bad refusals: Claude is much better at appropriate refusals now. Clear prompting in the user message without prefill should be sufficient.
Continuations (resuming interrupted responses): Move the continuation to the user message: "Your previous response was interrupted and ended with [previous_response]. Continue from where you left off."
Context hydration / role consistency (refreshing context in long conversations): Inject what were previously prefilled-assistant reminders into the user turn instead.
Tool parameter JSON escaping may differ
This is a breaking change when migrating from Sonnet 4.5 or earlier.
JSON string escaping in tool parameters may differ from previous models. Standard JSON parsers handle this automatically, but custom string-based parsing may need updates.
Extended thinking changes: budget_tokens configurations from Claude Sonnet 4.5 (thinking: {type: "enabled", budget_tokens: N}) are not supported on Claude Sonnet 5 and return a 400 error. Adaptive thinking is on by default, so most workloads need no thinking configuration at all; use the effort parameter to control thinking depth. If you ran Claude Sonnet 4.5 without extended thinking, pass thinking: {type: "disabled"} to preserve that behavior.
Remove sampling parameters
This is a breaking change when migrating from Claude 3.x models.
Sampling parameters (temperature, top_p, top_k) set to a non-default value return a 400 error on Claude Sonnet 5. Remove them from requests, and use prompting to guide the model's behavior instead.
Update tool versions
This is a breaking change when migrating from Claude 3.x models.
Update to the latest tool versions (text_editor_20250728, code_execution_20260521). Remove any code using the undo_edit command.
Handle the refusal stop reason
Update your application to handle refusal stop reasons.
Update your prompts for behavioral changes
Claude 4 models have a more concise, direct communication style. Review prompting best practices for optimization guidance.
Claude Haiku 4.5 and Claude Sonnet 5 differ more at the API level than adjacent models within one class: Claude Haiku 4.5 uses manual extended thinking (off by default), a 200k token context window, and up to 64k output tokens, while Claude Sonnet 5 runs with adaptive thinking on by default, serves a 1M token context window by default, and supports up to 128k output tokens.
model = "claude-haiku-4-5-20251001" # Before
model = "claude-sonnet-5" # AfterThinking configuration: Claude Haiku 4.5 supports manual extended thinking (thinking: {type: "enabled", budget_tokens: N}) and rejects thinking: {type: "adaptive"}. On Claude Sonnet 5, the support is reversed: adaptive thinking is on by default, and manual extended thinking returns a 400 error. Remove thinking: {type: "enabled", budget_tokens: N} configurations and rely on the default, or pass thinking: {type: "disabled"} to turn thinking off. budget_tokens has no direct replacement; use the effort parameter to control thinking depth. Effort is not available on Claude Haiku 4.5 and defaults to high on Claude Sonnet 5.
Sampling parameters removed: temperature and top_p work on Claude Haiku 4.5 (one at a time, not both). On Claude Sonnet 5, setting temperature, top_p, or top_k to a non-default value returns a 400 error. Remove these parameters and use prompting to guide the model's behavior.
Assistant prefill removed: Prefilling the assistant message works on Claude Haiku 4.5 but returns a 400 error on Claude Sonnet 5. Use structured outputs, system prompt instructions, or output_config.format instead.
Larger context window and output: Claude Sonnet 5 serves a 1M token context window by default, up from 200k tokens on Claude Haiku 4.5, and supports up to 128k output tokens, up from 64k. Claude Sonnet 5 also uses a different tokenizer, so re-run token counting rather than reusing counts measured against Claude Haiku 4.5.
Pricing: Claude Haiku 4.5 is priced at $1/$5 per million input/output tokens. For Claude Sonnet 5, introductory pricing of $2/$10 per million input/output tokens is in effect through August 31, 2026, after which the standard pricing of $3/$15 takes effect. See Claude pricing.
Cybersecurity safeguards: Claude Sonnet 5 has real-time cybersecurity safeguards. Requests that involve prohibited or high-risk cybersecurity topics may be refused, returned as a successful HTTP 200 response with stop_reason: "refusal". See Safeguards, warnings, and appeals for background.
claude-haiku-4-5-20251001 (or the claude-haiku-4-5 alias) to claude-sonnet-5.thinking: {type: "enabled", budget_tokens: N} configuration (returns a 400 error). Adaptive thinking is on by default; pass thinking: {type: "disabled"} to preserve no-thinking behavior, and revisit max_tokens for workloads that ran without thinking.high) to control thinking depth and token spend; it is not available on Claude Haiku 4.5, so no existing setting carries over.temperature and top_p settings (non-default values return a 400 error on Claude Sonnet 5).max_tokens limits, which you can raise up to the 128k maximum.stop_reason: "refusal" if your workload may touch cybersecurity topics.Claude Haiku 4.5 is the fastest and most intelligent Haiku model with near-frontier performance, delivering premium model quality for interactive applications and high-volume processing.
For a complete overview of capabilities, see the models overview.
For Claude Haiku 4.5 pricing, see Claude pricing.
For significant performance improvements on coding and reasoning tasks, consider enabling extended thinking with thinking: {type: "enabled", budget_tokens: N}.
Extended thinking impacts prompt caching efficiency.
Extended thinking is deprecated in Claude 4.6 models and removed in Claude Opus 4.7. If using newer models, use adaptive thinking instead.
Update your model name:
# From Haiku 3.5
model = "claude-3-5-haiku-20241022" # Before
model = "claude-haiku-4-5-20251001" # AfterReview new rate limits: Haiku 4.5 has separate rate limits from Haiku 3.5. See Rate limits documentation for details.
Explore new capabilities: See the models overview for details on context awareness, increased output capacity (64k tokens), higher intelligence, and improved speed.
These breaking changes apply when migrating from Claude 3.x Haiku models.
Update sampling parameters
This is a breaking change when migrating from Claude 3.x models.
Use only temperature OR top_p, not both. Setting both returns a 400 error on Claude Haiku 4.5.
Update tool versions
This is a breaking change when migrating from Claude 3.x models.
Update to the latest tool versions (text_editor_20250728, code_execution_20250825). Remove any code using the undo_edit command.
Handle the refusal stop reason
Update your application to handle refusal stop reasons.
Update your prompts for behavioral changes
Claude 4 models have a more concise, direct communication style. Review prompting best practices for optimization guidance.
claude-haiku-4-5-20251001text_editor_20250728, code_execution_20250825); legacy versions are not supportedundo_edit command (if applicable)temperature OR top_p, not both (setting both returns a 400 error)refusal stop reason in your applicationWas this page helpful?