Opus migration

 
 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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/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](https://platform.claude.com/docs/en/about-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](https://platform.claude.com/docs/en/build-with-claude/context-windows) (the default, with no beta header), [128k max output tokens](https://platform.claude.com/docs/en/models/overview), [adaptive thinking](https://platform.claude.com/docs/en/build-with-claude/thinking), [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching), [batch processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing), the [Files API](https://platform.claude.com/docs/en/build-with-claude/files), [PDF support](https://platform.claude.com/docs/en/build-with-claude/pdf-support), [vision](https://platform.claude.com/docs/en/build-with-claude/vision), and server-side and client-side [tools](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview), with two exceptions: [web fetch](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool) is not available on Claude Opus 5, and [Priority Tier](https://platform.claude.com/docs/en/api/service-tiers#supported-models) is not supported on Claude Opus 5. See each tool page for model availability.
+Claude Opus 5 is a drop-in upgrade for Claude Opus 4.8 at the same pricing of $5 USD per million input tokens and $25 USD per million output tokens; see [Claude pricing](https://platform.claude.com/docs/en/about-claude/pricing). There are two breaking changes for code already running on Claude Opus 4.8, covered under [Breaking changes](https://platform.claude.com/docs/en/models/opus-5/migration-guide#breaking-changes). Claude Opus 5 supports the same set of features as Claude Opus 4.8, including the [1M token context window](https://platform.claude.com/docs/en/build-with-claude/context-windows) (the default, with no beta header), [128k max output tokens](https://platform.claude.com/docs/en/models/overview), [adaptive thinking](https://platform.claude.com/docs/en/build-with-claude/thinking), [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching), [batch processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing), the [Files API](https://platform.claude.com/docs/en/build-with-claude/files), [PDF support](https://platform.claude.com/docs/en/build-with-claude/pdf-support), [vision](https://platform.claude.com/docs/en/build-with-claude/vision), and server-side and client-side [tools](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview), with two exceptions: [web fetch](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool) is not available on Claude Opus 5, and [Priority Tier](https://platform.claude.com/docs/en/api/service-tiers#supported-models) is not supported on Claude Opus 5. See each tool page for model availability.
 
 ## Migrating to Claude Opus 5 from Claude Opus 4.8
 
 <Note>
-  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](https://platform.claude.com/docs/en/models/opus-5/migration-guide#migrating-from-claude-opus-47) or [Migrating to Claude Opus 5 from Claude Opus 4.6 and earlier Opus models](https://platform.claude.com/docs/en/models/opus-5/migration-guide#migrating-from-claude-opus-46). They include this delta plus the breaking changes from earlier models (sampling parameters rejected, manual extended thinking rejected, prefill removed, new tokenizer).
+  This section covers the delta from Claude Opus 4.8 only. If your code is on Claude Opus 4.7 or earlier, use these sections instead: [Migrating to Claude Opus 5 from Claude Opus 4.7](https://platform.claude.com/docs/en/models/opus-5/migration-guide#migrating-from-claude-opus-47) or [Migrating to Claude Opus 5 from Claude Opus 4.6 and earlier Opus models](https://platform.claude.com/docs/en/models/opus-5/migration-guide#migrating-from-claude-opus-46). They include this delta plus the breaking changes from earlier models (sampling parameters rejected, manual extended thinking rejected, prefill removed, new tokenizer).
 </Note>
 
 ### Update your model name

 
 ### Breaking changes
 
-1. **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](https://platform.claude.com/docs/en/build-with-claude/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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
+1. **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](https://platform.claude.com/docs/en/build-with-claude/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. Thinking tokens are billed as output tokens even when the thinking text is not returned to you, so although per-token pricing is unchanged, a workload that ran without thinking on Claude Opus 4.8 can produce more output tokens per request on Claude Opus 5; see [Cost control](https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost#cost-control). 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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
 
+   The response shape changes with it. With thinking on, a response can begin with one or more `thinking` blocks before the first `text` block, and because `thinking.display` defaults to `"omitted"` on Claude Opus 5, those blocks arrive with an empty `thinking` field alongside their `signature`. Code that reads the reply by position, such as `content[0].text` or a stream handler that treats the first `content_block_start` event as text, breaks on these responses. Select content blocks by their `type` field instead: read `text` from the blocks whose `type` is `"text"`, and branch on the block type when handling stream events. To receive readable thinking summaries instead of an empty `thinking` field, set `display: "summarized"`; see [Controlling thinking display](https://platform.claude.com/docs/en/build-with-claude/thinking#controlling-thinking-display).
+
+   If you run a tool-use loop, pass the `thinking` blocks from each assistant response back to the API complete and unmodified when you return tool results, including blocks whose `thinking` field is empty. Echo the assistant message as received rather than filtering its content blocks by type or rebuilding it: the API rejects edited, reordered, or partially dropped thinking blocks with a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
+
 2. **Disabling thinking is capped at `high` effort:** You can still turn thinking off with `thinking: {type: "disabled"}`, but only at an [effort](https://platform.claude.com/docs/en/build-with-claude/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.

 
 * Update the model name from `claude-opus-4-8` to `claude-opus-5`.
 * Review workloads that ran without a `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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for the output artifacts that can appear and their prompting mitigations.
+* Update response parsing that reads content by position, such as `content[0].text` or a stream handler that assumes the first content block is text: with thinking on, `thinking` blocks arrive before `text` blocks. Select content blocks by `type` instead.
+* If you run a tool-use loop, pass `thinking` blocks back complete and unmodified when you return tool results; modified blocks return a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
+* Verify any code that parses the `thinking` field treats it as display text only. `thinking.display` defaults to `"omitted"` on Claude Opus 5, the same as on Claude Opus 4.8, so thinking blocks arrive with an empty `thinking` field; set `display: "summarized"` to receive readable summaries. See [Controlling thinking display](https://platform.claude.com/docs/en/build-with-claude/thinking#controlling-thinking-display).
 * Audit requests that disable thinking: `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.
 * Re-evaluate your `effort` setting: run a fresh [effort](https://platform.claude.com/docs/en/build-with-claude/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.
 * Review prompts near the caching minimum: prompts of 512 tokens or more can now create cache entries, down from 1,024 tokens on Claude Opus 4.8.

 * For agentic workloads, consider [task budgets](https://platform.claude.com/docs/en/build-with-claude/task-budgets) (beta) and mid-conversation tool changes (beta).
 * Re-tune length and verbosity prompts: default visible responses and written deliverables run longer on Claude Opus 5, 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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#response-length-and-verbosity) and [Written deliverable length](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#written-deliverable-length).
 * Remove verification and self-check instructions carried over from prompts tuned for earlier models (they cause over-verification on Claude Opus 5), constrain task scope explicitly for narrow tasks, and in multi-agent frameworks steer or cap subagent delegation. See [Task scope and over-verification](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#task-scope-and-over-verification) and [Controlling subagent spawning](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#controlling-subagent-spawning).
-* Re-baseline cost and latency on your own workloads.
+* Re-baseline cost and latency on your own workloads. Per-token pricing is unchanged from Claude Opus 4.8, but thinking tokens are billed as output tokens, so workloads that ran without thinking can produce more output tokens per request.
 
 ## Migrating to Claude Opus 5 from Claude Opus 4.7
 
-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](https://platform.claude.com/docs/en/build-with-claude/context-windows), [128k max output tokens](https://platform.claude.com/docs/en/models/overview), [adaptive thinking](https://platform.claude.com/docs/en/build-with-claude/thinking), [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching), [batch processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing), the [Files API](https://platform.claude.com/docs/en/build-with-claude/files), [PDF support](https://platform.claude.com/docs/en/build-with-claude/pdf-support), [vision](https://platform.claude.com/docs/en/build-with-claude/vision), and server-side and client-side [tools](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview), with two exceptions: [web fetch](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool) is not available on Claude Opus 5, and [Priority Tier](https://platform.claude.com/docs/en/api/service-tiers#supported-models) is not supported on Claude Opus 5. It also adds [mid-conversation system messages](https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages) and publicly documents [refusal stop details](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback#refusal-response). On the Claude API and Google Cloud, Claude Opus 5 also supports [computer use](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool) as the stable `computer_toolset_20260801` toolset and the [browser use tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/browser-use-tool) for tasks inside webpages, neither of which Claude Opus 4.7 supports; existing integrations on the earlier `computer_20251124` version continue to work unchanged on both models. To upgrade an existing integration, see [Migrate from `computer_20251124`](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool#migrate-from-computer-20251124).
+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 USD per million input tokens and $25 USD per million output tokens. It supports the same set of features as Claude Opus 4.7, including the [1M token context window](https://platform.claude.com/docs/en/build-with-claude/context-windows), [128k max output tokens](https://platform.claude.com/docs/en/models/overview), [adaptive thinking](https://platform.claude.com/docs/en/build-with-claude/thinking), [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching), [batch processing](https://platform.claude.com/docs/en/build-with-claude/batch-processing), the [Files API](https://platform.claude.com/docs/en/build-with-claude/files), [PDF support](https://platform.claude.com/docs/en/build-with-claude/pdf-support), [vision](https://platform.claude.com/docs/en/build-with-claude/vision), and server-side and client-side [tools](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview), with two exceptions: [web fetch](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-fetch-tool) is not available on Claude Opus 5, and [Priority Tier](https://platform.claude.com/docs/en/api/service-tiers#supported-models) is not supported on Claude Opus 5. It also adds [mid-conversation system messages](https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages) and publicly documents [refusal stop details](https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback#refusal-response). On the Claude API and Google Cloud, Claude Opus 5 also supports [computer use](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool) as the stable `computer_toolset_20260801` toolset and the [browser use tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/browser-use-tool) for tasks inside webpages, neither of which Claude Opus 4.7 supports; existing integrations on the earlier `computer_20251124` version continue to work unchanged on both models. To upgrade an existing integration, see [Migrate from `computer_20251124`](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool#migrate-from-computer-20251124).
 
 <Note>
   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](https://platform.claude.com/docs/en/models/opus-5/migration-guide#migrating-from-claude-opus-46) 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.

 
 ### Breaking changes
 
-1. **Thinking 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](https://platform.claude.com/docs/en/build-with-claude/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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
+1. **Thinking 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](https://platform.claude.com/docs/en/build-with-claude/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. Thinking tokens are billed as output tokens even when the thinking text is not returned to you, so although per-token pricing is unchanged, a workload that ran without thinking on Claude Opus 4.7 can produce more output tokens per request on Claude Opus 5; see [Cost control](https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost#cost-control). 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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
 
+   The response shape changes with it. With thinking on, a response can begin with one or more `thinking` blocks before the first `text` block, and because `thinking.display` defaults to `"omitted"` on Claude Opus 5, those blocks arrive with an empty `thinking` field alongside their `signature`. Code that reads the reply by position, such as `content[0].text` or a stream handler that treats the first `content_block_start` event as text, breaks on these responses. Select content blocks by their `type` field instead: read `text` from the blocks whose `type` is `"text"`, and branch on the block type when handling stream events. To receive readable thinking summaries instead of an empty `thinking` field, set `display: "summarized"`; see [Controlling thinking display](https://platform.claude.com/docs/en/build-with-claude/thinking#controlling-thinking-display).
+
+   If you run a tool-use loop, pass the `thinking` blocks from each assistant response back to the API complete and unmodified when you return tool results, including blocks whose `thinking` field is empty. Echo the assistant message as received rather than filtering its content blocks by type or rebuilding it: the API rejects edited, reordered, or partially dropped thinking blocks with a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
+
 2. **Disabling thinking is capped at `high` effort:** You can turn thinking off with `thinking: {type: "disabled"}`, but only at an [effort](https://platform.claude.com/docs/en/build-with-claude/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.

 
 * Update model name from `claude-opus-4-7` to `claude-opus-5` (or update aliases).
 * Review workloads that ran without a `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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for the output artifacts that can appear and their prompting mitigations.
+* Update response parsing that reads content by position, such as `content[0].text` or a stream handler that assumes the first content block is text: with thinking on, `thinking` blocks arrive before `text` blocks. Select content blocks by `type` instead.
+* If you run a tool-use loop, pass `thinking` blocks back complete and unmodified when you return tool results; modified blocks return a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
+* Verify any code that parses the `thinking` field treats it as display text only. `thinking.display` defaults to `"omitted"` on Claude Opus 5, the same as on Claude Opus 4.7, so thinking blocks arrive with an empty `thinking` field; set `display: "summarized"` to receive readable summaries. See [Controlling thinking display](https://platform.claude.com/docs/en/build-with-claude/thinking#controlling-thinking-display).
 * Audit requests that disable thinking: `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.
 * If you removed sampling parameters during the Opus 4.7 migration, no action is needed. If you re-added them with a 400-retry path, remove that retry path.
 * Re-evaluate your `effort` setting: run a fresh [effort](https://platform.claude.com/docs/en/build-with-claude/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.

 * If you used fast mode on Claude Opus 4.7, no request changes are needed beyond the model ID: `speed: "fast"` and the `fast-mode-2026-02-01` beta header work unchanged on Claude Opus 5.
 * For agentic workloads, consider [task budgets](https://platform.claude.com/docs/en/build-with-claude/task-budgets) (beta) and mid-conversation tool changes (beta).
 * Re-tune length and verbosity prompts, and remove verification and self-check instructions carried over from prompts tuned for earlier models.
-* Re-baseline cost and latency at your chosen effort level.
+* Re-baseline cost and latency at your chosen effort level. Per-token pricing is unchanged from Claude Opus 4.7, but thinking tokens are billed as output tokens, so workloads that ran without thinking can produce more output tokens per request.
 
 ## Migrating to Claude Opus 5 from Claude Opus 4.6 and earlier Opus models
 
-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:
+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 in this section, so it 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:
 
 * [1M token context window](https://platform.claude.com/docs/en/build-with-claude/context-windows) at standard API pricing with no long-context premium
 * [128k max output tokens](https://platform.claude.com/docs/en/models/overview)

 
    Adaptive thinking is steerable through prompting and the [effort parameter](https://platform.claude.com/docs/en/build-with-claude/effort); see [Choosing an effort level](https://platform.claude.com/docs/en/models/opus-5/migration-guide#choosing-an-effort-level).
 
-2. **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](https://platform.claude.com/docs/en/build-with-claude/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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
+2. **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](https://platform.claude.com/docs/en/build-with-claude/thinking). `max_tokens` remains a hard limit on total output, thinking plus response text, so revisit it for workloads that ran without thinking. Thinking tokens are billed as output tokens even when the thinking text is not returned to you, so although per-token pricing is unchanged, a workload that ran without thinking can produce more output tokens per request on Claude Opus 5; see [Cost control](https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost#cost-control). 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](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5#running-with-thinking-disabled) for mitigations where you can't.
 
+   The response shape changes with it. With thinking on, a response can begin with one or more `thinking` blocks before the first `text` block, and because thinking content is omitted by default on Claude Opus 5 (item 5 in this list), those blocks arrive with an empty `thinking` field alongside their `signature`. Code that reads the reply by position, such as `content[0].text` or a stream handler that treats the first `content_block_start` event as text, breaks on these responses. Select content blocks by their `type` field instead: read `text` from the blocks whose `type` is `"text"`, and branch on the block type when handling stream events.
+
+   If you run a tool-use loop, pass the `thinking` blocks from each assistant response back to the API complete and unmodified when you return tool results, including blocks whose `thinking` field is empty. Echo the assistant message as received rather than filtering its content blocks by type or rebuilding it: the API rejects edited, reordered, or partially dropped thinking blocks with a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
+
 3. **Disabling thinking is capped at `high` effort:** You can turn thinking off with `thinking: {type: "disabled"}`, but only at an [effort](https://platform.claude.com/docs/en/build-with-claude/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.
 
 4. **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 Python SDK (v1.0 and later) does not define them, and passing them raises a `TypeError`. 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.

 
 ### Behavior changes
 
-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.
+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 in this list.
 
 1. **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.
 

 * Remove `temperature`, `top_p`, and `top_k` from request payloads.
 * Replace `thinking: {type: "enabled", budget_tokens: N}` with `thinking: {type: "adaptive"}` plus the [effort parameter](https://platform.claude.com/docs/en/build-with-claude/effort), or remove the `thinking` field entirely; adaptive thinking is on by default on Claude Opus 5.
 * Review workloads that ran without a `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.
+* Update response parsing that reads content by position, such as `content[0].text` or a stream handler that assumes the first content block is text: with thinking on, `thinking` blocks arrive before `text` blocks. Select content blocks by `type` instead.
+* If you run a tool-use loop, pass `thinking` blocks back complete and unmodified when you return tool results; modified blocks return a 400 error. See [Preserving thinking blocks](https://platform.claude.com/docs/en/build-with-claude/thinking#preserving-thinking-blocks).
 * Audit requests that disable thinking: `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.
 * Remove any assistant-message prefills.
 * If your UI displays thinking content, explicitly opt in to thinking summarization.
-* Re-benchmark end-to-end cost and latency under the updated tokenization.
+* Re-benchmark end-to-end cost and latency under the updated tokenization; thinking tokens are billed as output tokens, so workloads that ran without thinking can also produce more output tokens per request.
 * Re-tune `max_tokens` to account for the updated tokenization.
 * Re-test any client-side token-count estimations.
 * If your application sends images, re-budget for [high-resolution image support](https://platform.claude.com/docs/en/build-with-claude/vision#high-resolution-image-support-on-claude-opus-4-7) (up to approximately 3x more image tokens per full-resolution image). Downsample before sending if you do not need the additional fidelity.

 
 ### What changed
 
-1. **Pricing:** Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. Claude Sonnet 5 is priced at $2/$10 per million input/output tokens. See [Claude pricing](https://platform.claude.com/docs/en/about-claude/pricing) for complete pricing.
+1. **Pricing:** Claude Opus 5 is priced at $5 USD per million input tokens and $25 USD per million output tokens. Claude Sonnet 5 is priced at $2/$10 USD per million input/output tokens. See [Claude pricing](https://platform.claude.com/docs/en/about-claude/pricing) for complete pricing.
 
 2. **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](https://platform.claude.com/docs/en/build-with-claude/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.