Claude Opus 5.5: Re-Test the Agent Contract Before Migration

What changed with Claude Opus 5.5

Reuters reported that Anthropic launched Claude Opus 5.5 on 22 September, saying it performs comparably to the higher-tier Claude Fable 5.1 while costing 40% less to run than Claude Opus 5. The company also said external groups including METR and Frontier Design tested the model before release. Those are vendor claims and launch evidence, not proof that a particular business workflow will improve.

Anthropic's developer documentation lists the Claude API model ID as claude-opus-5-5 and says the model is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. The published price is US$4 per million input tokens and US$20 per million output tokens, compared with US$5 and US$25 for Claude Opus 5.

Why this is more than a model-name change

Anthropic documents four breaking changes for existing Claude Opus 5 integrations. Thinking cannot be disabled; forced tool-choice modes any and tool return errors; thinking blocks are tied to the model and conversation; and the earlier computer_20251124 tool is not accepted on the Claude API or Google Cloud, where computer_toolset_20260801 is required. Amazon Bedrock retains different computer-use compatibility.

The response stream also changes. Text written between tool calls can arrive as progress-update thinking blocks, so an interface that only renders text blocks may appear silent during an agent run. Refusals can return HTTP 200 with stop_reason set to refusal and a policy category in stop_details. These details affect agent loops, telemetry, user feedback and fallbacks even when the final answer quality looks strong.

Re-baseline cost and behaviour together

Lower list prices do not establish lower cost per completed task. Claude Opus 5.5 defaults to medium effort, while Claude Opus 5 defaulted to high, and Anthropic says the newer model tends to think more at the same effort level. Teams should measure accepted outcome cost, latency, retries, refusal rate, tool-call accuracy and human correction rather than compare token prices alone.

Run the same representative workload set across the old and new models. Keep inputs, tool schemas, permissions and success criteria fixed. Segment results by workflow because a model that improves coding may still change communication, visual interpretation or refusal behaviour in ways that matter elsewhere.

Use an agent migration gate

Before changing production traffic, require one signed migration record that covers the complete runtime contract:

Promote through a small canary, compare live evidence with the staging baseline and keep the previous model configuration callable until the rollback exercise succeeds. ELYMENT AI helps teams turn model launches into controlled operating changes, with the evidence and approval needed to upgrade agents without treating a vendor announcement as production acceptance.

• pin the exact provider, platform and model ID, then record every changed parameter and beta header;
• replay task-level evaluations for output quality, tool selection, schema compliance, latency and accepted outcome cost;
• test thinking-block preservation, progress rendering, refusal parsing and fallback routes across multi-turn agent loops;
• exercise computer-use tools in the target cloud because compatibility differs by platform;
• verify logs capture model ID, effort, stop reason, tool calls, approvals, errors and final outcome without exposing hidden reasoning; and
• name the approver, canary threshold, stop condition and tested rollback configuration.

Sources

• Anthropic, What's new in Claude Opus 5.5 (22 September 2026) - Primary documentation for the model ID, breaking changes, feature behaviour, pricing and platform availability.
• Anthropic, Migrating to Claude Opus 5.5 (22 September 2026) - Primary migration guidance covering thinking, tool choice, computer use, refusals, progress updates, effort calibration and rollback preparation.
• Reuters, Anthropic unveils Claude Opus 5.5 (22 September 2026) - Independent reporting on the release, Anthropic's performance and safety claims, pricing and cloud availability.

Continue learning

• Treat model names as controlled production dependencies
• Design durable boundaries for long-running agents
• Design refusals and safety stops into the workflow

Frequently asked questions