Opus 5.5's lower pricing comes with four breaking changes for agents

Anthropic has released Claude Opus 5.5 with meaningful price cuts—input tokens drop from $5 to $4 per million, and output tokens fall from $25 to $20 per million. The 1 million-token context window remains the same. While the cost savings look straightforward on paper, the upgrade path is complicated: four breaking changes will cause requests built for Opus 5 to fail with 400 errors once you switch the model ID.

Reasoning is mandatory now#

In Opus 5.5, thinking (the model's reasoning process) runs by default and cannot be disabled. Requests that previously set thinking to off or specified budget_tokens will now return errors, forcing a rethink of how to manage reasoning overhead.

The new control mechanism is effort levels—low, medium, or high—replacing the old budget approach. Since the default effort dropped from high to medium, agents that don't explicitly set this parameter will quietly run at a lower reasoning level, potentially changing latency and quality. Developers should re-evaluate effort settings for each agent step, as the cost-to-benefit tradeoff may have shifted.

Another consequence: thinking blocks now appear first in responses. Any code that assumes the first content block is text will break and needs updating.

Forced tool calls no longer work#

Opus 5.5 rejects tool_choice set to "any" or to a specific tool, returning errors on both regular requests and token-counting calls. This breaks agents that force tool invocation at specific decision points (e.g., when a database lookup or external API call is mandatory).

The recommended replacement combines auto tool selection with either strict tool use or structured outputs, along with clearer prompting about when a tool should be invoked. This approach is more permissive but requires adjusting agent logic to handle cases where the model might skip a tool.

Thinking blocks have model affinity#

Thinking blocks are now tied to the specific model that generated them. Only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks—if a conversation is handed to any other model via routing or fallback logic, earlier reasoning will be silently discarded rather than producing an error.

Conversations must also remain append-only for thinking blocks to stay valid. Trimming old messages, rewriting the system prompt mid-conversation, or editing tool definitions invalidates stored blocks. , though existing accounts won't throw errors (yet—Anthropic plans to enforce this universally in future models).