Anthropic's Opus 5.5 Price Cut Comes With Four Breaking Changes for AI Agents

Anthropic unveiled Claude Opus 5.5 on Tuesday with a significant price reduction: input tokens dropped from $5 to $4 per million, while output tokens fell from $25 to $20 per million. The model retains the 1 million-token context window and 128,000-token maximum output of its predecessor. Yet the lower cost masks a steeper migration challenge than simply swapping a model identifier.

The company's official migration guide identifies four breaking changes capable of triggering 400 errors in agents built for Opus 5. Beyond these critical incompatibilities, several additional modifications may alter agent behavior without raising exceptions, complicating the upgrade path for production systems.

Thinking is always on

The first breaking change centers on how thinking operates. Opus 5.5 rejects requests that disable thinking or enable it with a budget_tokens parameter, making effort the sole mechanism for controlling reasoning depth. Agents that previously turned thinking off during straightforward tasks to conserve tokens must now assign those steps a lower effort level instead.

Since thinking now runs continuously, responses begin with thinking blocks. Any code expecting the first content block to be text will require modification. Additionally, the default effort setting has shifted from high on Opus 5 to medium on Opus 5.5, meaning requests omitting the parameter will silently execute at reduced intensity. Anthropic advises explicitly setting effort and re-evaluating appropriate levels for each step, as both cost and latency may have changed.

No more forced tool calls

Forced tool use has been eliminated. Setting tool_choice to any or tool now returns a 400 error, including on the token counting endpoint. This breaks cost estimation built on those settings and affects agent loops that force tool invocation when a step requires database queries, code execution, or external service calls. Anthropic's replacement combines auto-combined strict tool use with structured outputs, requiring the prompt to specify when a tool applies.

Routing and conversation history

Thinking blocks are now bound to the specific model and conversation that generated them. On the Claude API, only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks. A router or fallback that directs a conversation to any other model will process those turns without the earlier reasoning rather than returning an error. This adds complexity for teams already monitoring whether agent calls are being silently routed to older models.

Opus 5.5 itself can read thinking blocks from Opus 5 and earlier versions of Opus, Sonnet, and Haiku, but not from Fable or Mythos. Conversations must remain append-only for thinking blocks to stay valid. Trimming old messages, modifying tool definitions, summarizing earlier context on the client side, or changing the system prompt mid-conversation invalidates existing thinking blocks. For accounts created on or after August 31, 2026, at midnight UTC, replaying an invalid thinking block returns a 400 error by default. Older accounts receive no error, though invalid blocks still reach the model; Anthropic indicates future models will enforce this check universally. Integrations that never edit prior turns require no code changes. Anthropic notes that Claude Code, claude.ai, Claude Managed Agents, and the Claude Agent SDK already follow this pattern, while agents that compact their own context should consult the company's preserved thinking documentation.

The fourth breaking change affects computer-use agents on the Claude API and Google Cloud. Opus 5.5 rejects the computer_20251124 tool and accepts computer use only through the computer_toolset_20260801 toolset. While the request itself becomes simpler—the beta header disappears and the toolset entry requires no name or display dimensions—the agent loop demands substantial rework. Each action now arrives as its own tool_use block identified by the block's name rather than input.action, multiple actions can appear in a single turn, and every result must echo toolset_name. The older tool continues functioning on Amazon Bedrock; developers on other platforms should consult Anthropic's computer use tool compatibility documentation.

Changes that won't throw errors

One significant change produces no error at all. On Opus 5, text Claude generates between tool calls returns as text blocks, but on Opus 5.5 that narration arrives as progress-update thinking blocks. At the default thinking.display setting of omitted, these blocks are empty.

Any agent interface streaming that narration to users will fall silent between tool calls until developers set display to updates—a beta option returning progress updates while hiding reasoning—or to summarized, which returns both. Each non-empty thinking block must then render before the tool call it precedes.

Opus 5.5 also introduces broader safety classifiers. It can return a stop_reason of refusal with stop_details categories now including bio and reasoning_extraction alongside cyber. Anthropic's server-side fallback will not retry requests declined under reasoning_extraction, instead handing the refusal back to the application.

Agents lacking refusal handling will halt mid-task, a problem developers have already encountered with OpenAI's safety system interrupting API responses.

Upgrading from older models

Teams migrating from Opus 4.8 must first complete the Opus 5 migration, which addresses thinking being on by default and resulting response-shape changes, before tackling Opus 5.5 modifications. Teams on Opus 4.7 or earlier face additional work. Those on models predating Opus 4.7 encounter rejected sampling parameters, rejected manual extended thinking, removed prefill, and a newer tokenizer.

Claude Managed Agents users only need to change the model name. Developers in Claude Code can run /claude-api migrate to apply the model ID swap, parameter changes, prefill replacement, and effort calibration across a codebase before reviewing a manual verification checklist.

Anthropic recommends testing the migration in a development environment before switching production traffic. Developers maintaining custom integrations must also test the surrounding components. Tool calls, model handoffs, conversation history, and user-facing progress updates can all behave differently after the switch, since agent failures frequently originate outside the model itself.