Claude Opus 5 to 5.5: What Changed and What Breaks

Claude Opus 5.5, released on 22 September 2026, costs $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5. It also changes behavior: thinking can no longer be switched off, forced tool calls return an error, and the default effort drops from high to medium. After the 9 November Fable 5.2 launch, Anthropic's docs still point most teams to Opus 5.5 first.

Checked against Anthropic's pages on 13 November 2026. Benchmark and cost-per-task claims are Anthropic's own.

Key Takeaways

• Opus 5.5 is $4 input and $20 output per million tokens, against $5 and $25 for Opus 5 and Opus 4.8. Cache reads fall from $0.50 to $0.20 per million, according to Anthropic's launch post.
• Anthropic says typical workloads cost about 40% less than on Opus 5 at default settings. The 20% price cut explains half of that; the rest presumably comes from fewer tokens and the lower default effort, which is my inference, not Anthropic's statement.
• Requests that disable thinking, use manual thinking budgets or force a tool with tool_choice return a 400 error on Opus 5.5, per the migration guide.
• The default effort is medium, where Opus 5's was high. A request that omits effort now runs one level lower.
• Responses can start with thinking blocks, so code that reads content[0].text breaks.
• Anthropic's models overview still tells most teams to start with Opus 5.5, and its earliest retirement date is 22 September 2027.

The progression: Opus 4.8, Opus 5, Opus 5.5

Three consecutive Opus releases look similar on a price list and differ a lot in behavior. All figures are per million tokens, from Anthropic's pricing page and launch posts, as of 13 November 2026.

Anthropic's Opus 5 announcement of 24 July described stronger self-verification and iteration at the same price as Opus 4.8. Opus 5.5 then kept the model family but changed the price and the request contract. It is available on all platforms, including AWS, Google Cloud and Microsoft Azure, with zero data retention and watermarking to meet the EU AI Act, per the launch post.

What Anthropic says improved

Anthropic's launch table compares Opus 5.5, Opus 5 and Fable 5.1. Effort was set to Max with adaptive thinking unless a benchmark is noted, and all numbers are Anthropic-reported.

Anthropic also says Opus 5.5 produces output more than 30% faster, uses fewer tokens per task, writes more clearly with key information first, and attempts boundary circumvention roughly 85% less often than Opus 5 or Mythos 5.1. At default medium effort it reports 54.6% on FrontierCode and 52.5% on CursorBench, so most of the headline performance survives the lower setting. Anthropic adds that the gap to Fable 5.1 is narrower in real use than the scores suggest.

An outside view is thinner. Arena's Alignment Index, which I covered in GPT-6.1 Sol vs Claude Opus 5.5, measures agent behavior rather than quality, and it does not settle which model writes better for your brand.

What breaks when you move from Opus 5

Anthropic's migration guide lists changes that apply to every starting model. These are the ones most likely to hit a marketing or SEO pipeline built on the API, including no-code tools that call it.

1. Thinking cannot be disabled. thinking: {"type": "disabled"} and manual budget_tokens return a 400 error. Remove the field and use effort instead. Thinking tokens are billed as output even when the thinking text is hidden, so a job that ran with thinking off can use more output tokens per request.
2. Responses start with thinking blocks. Code that reads the first content block as text fails. Select blocks by their type, and pass thinking blocks back unmodified in tool loops, or the API rejects the request.
3. Forced tool calls are rejected. tool_choice types any and tool return a 400 error. Anthropic suggests auto with strict tool use or structured outputs, and saying in the prompt when the tool applies. Many extraction workflows use forced tools, so this is the likeliest silent failure.
4. Sampling parameters must be omitted. temperature, top_p and top_k should be omitted. Under API version 2026-10-21 they return a 400 on every model; I cover that in my post on future-proofing Claude prompts.
5. Do not end a request with a prefilled assistant turn. The guide lists it as rejected. Use structured outputs or system-prompt instructions.
6. Text between tool calls moves into thinking blocks. An app that streams progress text to users goes quiet unless you request a thinking display setting.
7. Refusals can appear. Responses can return stop_reason: "refusal", with categories such as bio and cyber. Configure fallback or a retry.
8. Priority Tier is not supported on Opus 5.5 even though Opus 4.8 keeps it. If you have a capacity commitment, plan separately.

One change is a benefit. The minimum cacheable prompt on Opus 5.5 is 512 tokens, down from 1,024 on Opus 4.8, so shorter style guides can now be cached.

Cost math: the price cut is smaller than it looks

Take 1,000 content pieces, each with 8,000 input tokens and 4,000 output tokens, excluding thinking tokens. This is my illustrative workload.

The cache-read cut from $0.50 to $0.20 sounds large, but on this workload output tokens dominate the bill, so the total moves only about one point more than the base price cut. Cache writes are excluded and would add a small one-time cost. The 40% figure Anthropic cites depends on token savings and lower default effort that I cannot reproduce from a price list. To see your own number, log total tokens per accepted piece on both models. Mixed teams should route simple jobs to a cheaper tier anyway, as in my model routing strategy.

What this means for marketers and founders

• If you only use claude.ai, Cowork or Claude Code, none of the API errors apply to you. The changes you will notice are speed and effort defaults.
• If you run Claude through scripts, Make, Zapier or n8n, search your workflows for thinking, tool_choice, temperature and prefilled assistant messages before changing the model ID. Fix errors in staging, not production.
• Set effort explicitly. Run one effort sweep: the same 20 real tasks at low, medium and high, scored blind. Use the lowest level that holds quality.
• Do not rush to Fable 5.2. Opus 5.5 is half its price, and the Fable 5.2 post shows the knowledge-work gap is narrow.
• Keep Opus 5 as a fallback for a few weeks. It remains active, with an earliest retirement date of 24 July 2027, so you have time.

What we do not know yet

• How many total tokens a content task uses on Opus 5.5 versus Opus 5 once thinking is always on. Anthropic says fewer; I have not measured it.
• Whether independent writing-quality tests favor Opus 5.5 over Sonnet 5.5 for your voice. The Sonnet vs Opus vs Fable comparison frames how to test it.
• How often safety classifiers reroute ordinary marketing requests. Anthropic reports lower false positives on Fable 5.2, not on Opus 5.5 specifically.

FAQ

How much does Claude Opus 5.5 cost?

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, as of 13 November 2026. Cache reads are $0.20 per million and cache writes are $5. Fast mode, up to 2.5 times faster, is $8 and $40.

Is Opus 5.5 cheaper than Opus 5?

Yes, by 20% on base prices: Opus 5 is $5 and $25. Cache reads fall 60%, from $0.50 to $0.20. Anthropic claims typical workloads cost about 40% less at default settings, which includes effects beyond the list price.

Can I turn off thinking in Claude Opus 5.5?

No. Adaptive thinking is always on, and requests with thinking set to disabled or with manual budgets return a 400 error. You control depth with the effort parameter, which has five levels from low to max.

What is the default effort for Opus 5.5?

The default on the Claude API is medium. Opus 5's default was high, so a request that omits effort runs one level lower than before. Set effort explicitly if you want consistent results across both models.

Is Opus 5.5 better than Fable 5.2?

For most knowledge work Anthropic recommends starting with Opus 5.5. Fable 5.2 leads on agentic coding benchmarks, but on Anthropic's knowledge-work Elo measure Opus 5.5 scores 1800 against 1785. Opus 5.5 also costs half as much.

Will my existing Opus 5 integration stop working?

Not unless you change the model ID. Opus 5 remains active, and Anthropic lists its retirement as no sooner than 24 July 2027. Anthropic gives at least 60 days' notice before retiring a publicly released model.

Do I need to migrate if I only use Claude in the browser?

No. The migration items apply to API and agent integrations. In the apps you simply get the new model, with different speed and usage-limit behavior covered in my post on Claude plan changes.

Want help with this?

I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help testing model upgrades against real content and SEO workflows, see my work and reach me through the contact form at younusfardeen.com.