Claude Opus 5 Guide: Refusals, Default Thinking and When to Stay

Most write-ups of Claude Opus 5 focus on its launch claim: close to Claude Fable 5's intelligence at half the price. That claim mattered in July. By October the more practical question is different, because Opus 5.5 has since replaced it at a lower price. If you are on Opus 5, should you move, and if you are choosing fresh, is there any reason to pick it?

The answer depends less on benchmarks than on two API behaviours Opus 5 introduced and one set of controls Opus 5.5 later took away.

What Opus 5 changed from Opus 4.8

Anthropic released Opus 5 on 24 July 2026 at the same price as Opus 4.8: $5 per million input tokens and $25 per million output, with a 1M-token context and 128K maximum output. It became the default model on Claude Max. Anthropic described it as thoughtful and proactive, aimed at software engineering, knowledge work, long-running agents and scientific research.

For developers, the important changes were in the request contract:

• Thinking is on by default. On Opus 4.8, omitting the thinking field meant no thinking. On Opus 5, omitting it runs adaptive thinking. A request that was cheap and fast on 4.8 may now spend output tokens reasoning before it answers.
• Disabling thinking is capped. {type: "disabled"} is accepted only at effort high or below; pairing it with xhigh or max returns a 400.
• Safety classifiers can decline a request. You receive a normal HTTP 200 with stop_reason: "refusal" rather than an error.
• A lower cache minimum. The smallest cacheable prompt dropped to 512 tokens from 1,024 on Opus 4.8, so some short prompts start caching with no code change.

Handling the refusal stop reason

This is the change most likely to break a working integration. Code that reads content[0] without checking stop_reason fails on a refusal, and it fails as a confusing downstream error rather than a clear one.

Anthropic says it deliberately avoided training Opus 5 on cyber tasks and that its cyber classifiers intervene about 85% less often than on Fable 5. Benign security and life-sciences work can still trip them. Three things to put in place:

1. Branch on stop_reason first, before touching content, everywhere you call the model.
2. Opt into fallbacks. Opus 5 launched with automatic fallbacks in beta: a flagged request can be re-run on another model inside the same call. Cyber-category refusals route to Opus 4.8 by default. The Opus 4.8 guide explains why that model is the landing spot.
3. Log the refusal category. If one route trips classifiers repeatedly, you want to know before your users do.

Model routing and fallbacks covers building a chain that degrades cleanly rather than failing.

What Anthropic claimed on capability

Anthropic's launch page leaned on relative claims rather than headline scores. It said Opus 5 more than doubles Opus 4.8 on Frontier-Bench v0.1, comes within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task, and surpasses Fable 5's best OSWorld 2.0 result at just over a third of the cost. It also called Opus 5 its most aligned model to date at launch.

The Opus 5.5 announcement later published Opus 5 figures alongside its successor: 52.3% on Terminal-Bench 4.0, 74.0% on OSWorld 2.1 and 1,708 on GDPval-AA v2.1. Those are useful mainly as a baseline for how far 5.5 moved. As with every vendor chart, the evaluations and scaffolds are the vendor's choice; vendor-claimed vs independent benchmarks covers how to read them.

Opus 5 or Opus 5.5?

Opus 5.5 costs $4 / $20 per million tokens against Opus 5's $5 / $25, uses the same tokenizer and context size, and scores higher on every row Anthropic published. For most new work it is the obvious choice. But it also introduced four breaking changes, and each one is a reason some integrations should stay on Opus 5 for now:

• Thinking cannot be disabled on 5.5 at any effort level. If a latency-critical route depends on thinking off, Opus 5 still allows it at high or below.
• Forced tool use returns a 400 on 5.5. tool_choice of any or a named tool still works on Opus 5. Extraction pipelines built on forced tool calls need rework before moving.
• Thinking blocks are bound to the model and conversation on 5.5. Harnesses that edit earlier turns may need restructuring to be append-only.
• Computer use needs a new toolset on 5.5 on the Claude API and Google Cloud.

The effort default also differs: Opus 5 defaults to high, Opus 5.5 to medium. Moving without setting effort explicitly changes behaviour silently. The Opus 5.5 guide covers the migration in detail. Anthropic lists Opus 5's retirement as not sooner than 24 July 2027, so staying is a schedulable decision rather than an emergency.

Prompting adjustments that still apply

Anthropic's notes describe two behavioural shifts on Opus 5 worth correcting for:

• Longer visible responses. Effort does not reliably change this; a short conciseness instruction in the system prompt does.
• More readiness to spawn subagents. Useful for parallel work, expensive if uncapped. Cost-sensitive workloads should limit delegation explicitly.

Anthropic also found low and medium effort unusually effective on this model. Sweep downward from the default before assuming you need high.

Variants and a short checklist

The catalogue lists anthropic/claude-opus-5 and a :batch variant, which is the same model on the asynchronous Batch API at half price ($2.50 / $12.50 per million tokens). Fast mode runs at about 2.5 times normal speed for twice the price, on the Claude API only.

If you are staying on Opus 5 for now:

1. Check stop_reason before reading content on every call.
2. Enable fallbacks and log refusal categories.
3. Set thinking and effort explicitly rather than relying on defaults.
4. List which of the four 5.5 breaking changes affect you, and schedule the work to remove them.