Claude Opus 5.5: more capability, lower cost, new rules

What Claude Opus 5.5 is and what it promises

Claude Opus 5.5 is the new model in Anthropic's Opus line, released on Tuesday, September 22, 2026, two months after Opus 5 (July 24). Anthropic describes it as a model for long-running agentic coding and knowledge work, and its documentation now recommends starting with Opus 5.5 for most workloads.

The headline promises, all from Anthropic: performance at the level of Claude Fable 5.1 (its most capable widely released model) on most work, a 40% lower running cost than Opus 5 on typical workloads and output generation more than 30% faster. It is available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry and Claude Platform on AWS, and in the Claude Pro, Max, Team and Enterprise plans.

It is also the first model Anthropic has shipped since its CEO, Dario Amodei, published the essay We Must Pace the Frontier in September 2026, calling for a slower pace of capability gains to leave time for alignment and external evaluation. The same day, OpenAI launched GPT-6 Sol and GPT-6 Luna, and Anthropic said Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.

Where it fits in the Claude family

Four current models, four price points.

The practical reading: Opus 5.5 takes the slot that used to require paying for the Fable tier. Fable 5.1 remains for the most demanding reasoning and very long-horizon agents, or for when your evaluations of Opus 5.5 at high effort still fall short. Sonnet 5 and Haiku 4.5 remain the options for volume and latency.

Two pricing details matter in production: an Opus 5.5 prompt-cache read costs $0.20 per million tokens (5% of the input price, versus $0.50 on Opus 5), and the Batch API applies a 50% discount. On batches, Opus 5.5 supports up to 300,000 output tokens with a beta header.

Reading the benchmarks carefully

What Anthropic publishes and what independent leaderboards say.

Three cautions before deciding on this table. First, these tests are run or chosen by the vendor itself, several in new versions (Terminal-Bench 4.0, OSWorld 2.0) with little comparison history. Second, it doesn't win everywhere: in the same table, GPT-6 Astra comes out ahead on AutomationBench and Terminal-Bench-Science. Third, none of these tests is in business Spanish or French, which is where we most often see models break.

Independent leaderboards point the same way, with nuances. On September 25 and 26, Opus 5.5 topped the Arena text leaderboard (1,509 points) and the Artificial Analysis Intelligence Index (58 points, version 4.3.2). It was not yet on the official Terminal-Bench 4.0 board, which GPT-6 Astra led at 58.2%, and on ARC-AGI-2 GPT-6 Astra scored 95.0% against 93.3% for Opus 5.5, although at almost three times the cost per task.

Anthropic also cites customer anecdotes: a 680,000-line code migration in under a day, or an audit of a 200,000-line codebase in under three hours. They come from its launch communication: a useful signal of where the model points, not a promise for your project.

What changes for engineering teams

Four changes break code written for Opus 5.

Thinking is always on. Opus 5.5 uses adaptive thinking that cannot be disabled; the only dial is the effort parameter (low, medium, high, xhigh, max). The default drops to medium, one level below Opus 5. Anthropic says Opus 5.5 at medium beats Opus 5 at high on its coding and knowledge-work evaluations, and that on several coding evaluations low comes close at much lower cost. We cover this in depth in when paying for thinking pays off.

• Sending thinking: disabled or a fixed thinking-token budget returns a 400 error.
• Forced tool use (tool_choice set to any or a specific tool) is no longer supported: use auto with strict tools, or structured outputs.
• Thinking blocks are tied to the model and the conversation: if your code edits earlier turns, accounts created on or after August 31, 2026 get an error. History must be append-only.
• Computer use only works through the new computer_toolset_20260801; the earlier tool is rejected.

One more change breaks nothing but surprises teams: the short notes the model writes between tool calls now arrive as thinking blocks, empty by default. If your interface shows that progress to users, it will go quiet until you enable the updates display mode. Finally, fast mode (research preview, Claude API only) promises up to 2.5x speed at $8 / $40 per million tokens.

What it really costs: a worked example

Price per token drops 20%; cost per task can drop much more.

Take a typical agent step: 12,000 input tokens, 10,000 of which are a stable prefix (instructions, tools, reference documents) that can be cached, and 2,000 output tokens including thinking, which is billed as output. These are illustrative assumptions, not measurements from a project.

The big saving is not the list price but the combination: cheap cache reads, a lower default effort and, according to Anthropic, fewer tokens per solved task. The number to watch is cost per completed task: a cheaper step that needs two retries is not cheaper. If the work isn't interactive, the Batch API halves the bill. More levers in the real cost of AI.

Safety and refusals: what your team should know

Opus 5.5 runs cybersecurity and biology safety classifiers similar to Fable 5.1's; the biology one is new compared with Opus 5. According to the documentation, finding vulnerabilities in source code is allowed, while dual-use biology research requests are declined except for participants in Anthropic's verification program. A new category, reasoning_extraction, declines attempts to make the model copy its internal reasoning into the answer.

For a product team this has a concrete consequence: a refusal arrives as a normal response with stop_reason: refusal, not as an error. Your code has to handle it, log it and, where appropriate, retry on another model; the API can configure that fallback server-side. Pentest and offensive-security teams should test their real workflows before migrating.

On robustness, Anthropic says Opus 5.5 matches or beats Opus 5 against prompt injection in every setting it tested, and attempts to circumvent containment boundaries 85% less often than Opus 5. Good news, but no model solves prompt injection on its own: agent design remains the main defense, as we explain in prompt injection and agent security.

Which model for which job

Our starting point before evaluating; it doesn't replace the evaluation.

This table is a starting point. The method we use to decide, with 30 real prompts per client in three languages, is in how we choose a model, and the full vendor landscape is in the September 2026 model map.

Checklist for migrating from Opus 5

1. Set effort per route and run your evaluation at low, medium and high before choosing.
2. Remove any thinking: disabled or fixed thinking budget.
3. Replace forced tool use with auto plus strict tools or structured outputs.
4. Make your conversation history append-only: no editing of earlier turns.
5. If you use computer use, move to computer_toolset_20260801 and test it on Opus 5 first, which accepts both versions.
6. Display between-tool progress updates if your interface relies on them.
7. Handle stop_reason: refusal with logging and a fallback model.
8. Check your tier's Opus 5.5 rate limits before moving volume.
9. Compare cost per completed task, not price per token, over at least a week of real traffic.

What it means for companies in Colombia and Europe

Procurement and contracts. Opus 5.5 is on the three big clouds (AWS, Google Cloud and Microsoft). For many Colombian and European companies that means using it under existing contracts and spend commitments, without onboarding a new vendor. Mind the cost of residency: US-only inference costs 1.1x the base price, and regional endpoints on Bedrock and Google Cloud carry a 10% premium.

Access risk. On June 12, 2026, a US export-control directive forced Anthropic to withdraw Fable 5 and Mythos 5 from all users; Fable 5 came back on July 1. Any company outside the US should keep a fallback model from another vendor, tested and ready to switch on.

Personal data. In Colombia, sending personal data to a model hosted abroad is an international transfer under Law 1581 of 2012: check the data subjects' authorization and the provider's safeguards. In the European Union the GDPR applies and, from August 2, 2026, so do the AI Act transparency obligations for anyone deploying a chatbot. General-purpose model obligations fall on Anthropic; yours concern how you use it. Details in our 2026 regulatory map.

Languages. No published benchmark measures the Spanish of a Colombian insurer or the French of a Paris law firm. That test is yours to run. At Slash AI Lab we run Opus 5.5 through the same 30-prompt suite per client that we use for any new model; we won't publish numbers until we have reproducible results.