Claude Opus 5.5: Price, Specs and When to Use It

Claude Opus 5.5 is Anthropic's model for long-running agentic coding and knowledge work. It costs $4 per million input tokens and $20 per million output tokens, takes text and images, and has a 1M-token context window with 128K max output. It is closed-weight. It launched on 22 September 2026. Checked 31 October 2026 against Anthropic's docs.

Specs and prices

All rows are from Anthropic's Opus 5.5 model page and pricing page, 31 October 2026. The 1M window is billed at the standard rate: a 900,000-token request costs the same per token as a 9,000-token one. The minimum cacheable prompt is 512 tokens.

Anthropic's launch post says Opus 5.5 costs 40 percent less than Opus 5 on typical workloads, which it attributes to a 20 percent cut in list price (Opus 5 was $5 and $25), a 60 percent cut in cache reads and lower token use. That is a vendor claim about typical workloads, not a measurement of yours.

Who it is for

Anthropic's one-line description is "for long-running agentic coding and knowledge work." Its overview tells readers to start with Opus 5.5 for most workloads, and to move to Fable 5.1 when evaluations on Opus at higher effort still fall short. In the Sonnet 5.5 launch post, Anthropic says Opus "remains clearly stronger at complex, open-ended work requiring sustained judgment."

In practice that means three kinds of work:

• Agent runs that last for hours across a large codebase.
• Hard debugging where the cost of a wrong answer exceeds the cost of tokens.
• Planning and review steps in a setup where a cheaper model does the typing.

The launch post gives examples, again vendor-reported: a tester completing a 680,000-line code migration in under a day, and a test where Opus 5.5 cut page load times across a web app on 39 of 40 attempts. Both are anecdotes from early testers, and we cannot judge how typical they are.

Where it sits in the family

The default effort row is easy to overlook. Opus defaults to medium and Sonnet to high, so out of the box Sonnet may spend more thinking on a prompt than Opus does. Whether Opus is the better model depends on the ceiling you can reach by raising effort, not on the defaults. See Claude models for the whole family and Claude Opus vs Sonnet for a direct comparison.

Published benchmarks (vendor-reported)

Anthropic's Opus 5.5 launch post includes this table. Every figure is vendor-reported by Anthropic. Fable 5.1 and Opus 5 are Anthropic's own models, and the GPT-6 Astra and GPT-5.6 Sol figures come from OpenAI's reports or public leaderboards, as Anthropic's footnotes explain.

Read the footnotes before you compare. Anthropic says Opus 5.5 results use adaptive thinking at max effort unless noted, with Terminal-Bench 4.0 reported at xhigh for Opus 5.5 and at high for GPT-6 Astra, "as reported by OpenAI." The standard error on Terminal-Bench is about 2.6 points for Opus 5.5. Opus 5.5 ran with production safeguards on, and where they stepped in, older models completed cybersecurity, biology and frontier LLM development tasks, which Anthropic says likely lowers its scores on those. AutomationBench results come from Zapier. Two numbers in the table are not Opus wins: Astra leads on AutomationBench and on Terminal-Bench-Science.

Anthropic also warns that "benchmark margins have become a less reliable guide to real-world differences," and says the gap between Opus 5.5 and Fable 5.1 is narrower in its own use than the scores suggest. We take that at face value and would run our own evaluation before choosing.

Behaviour changes to plan for

The Opus 5.5 "what's new" page lists four breaking changes for code moving from Opus 5.

1. Thinking cannot be disabled. A request with thinking: {"type": "disabled"} or a manual budget returns a 400 error. Lower the effort instead.
2. Forced tool use fails. tool_choice of any or a named tool returns a 400 error. Keep auto, and use strict tool use for schema-valid output.
3. Thinking blocks are bound to the model and conversation. Keep conversations append-only, and change instructions through mid-conversation system messages instead of editing earlier turns.
4. The earlier computer use tool is rejected on the Claude API and Google's managed platform. Use computer_toolset_20260801.

The page also lists supported features, among them per-message effort (beta), task budgets, prompt caching and fast mode. Fast mode is a research preview for Opus 5 and 4.8, and it is listed for Opus 5.5 at $8 and $40, with the caveat that it cannot be combined with the Batch API.

Effort, the main dial

Because thinking is always on, the effort parameter is how you steer Opus 5.5. The levels named in Anthropic's docs run from low and medium through high and xhigh to max, and the default on the API is medium. Higher effort means the model works for longer, which raises cost per task and generally raises the score. Anthropic's Sonnet 5.5 launch post plots score against cost per task at each level for that reason.

Two practical points follow. First, Opus 5.5 supports changing effort per message in beta, and Anthropic says doing so does not invalidate the prompt cache, so a long session can run at low effort for routine steps and spike to high for a hard one. Second, the benchmark table above is mostly max effort, so the scores describe the top of the cost curve. Your default medium runs sit lower.

Safeguards and access

Anthropic says Opus 5.5 is comparable to its limited-availability Mythos 5.1 model in biology and cybersecurity, so it ships with safeguards similar to those on Fable 5.1. In the launch post, vetted organisations can apply for wider biology research use, and verified cybersecurity practitioners can apply for cyber work. If your work is security testing, expect refusals or fallbacks that a smaller model may handle differently, and read the system card before you build on it.

Anthropic also says Opus 5.5 is its strongest model on its automated behavioural audit, with fewer hard-to-reverse actions and better resistance to prompt injection than Opus 5. Those are vendor-reported alignment claims, and the system card has the detail. On subscription plans, the post says five-hour usage limits rose on Pro, Max, Team and seat-based Enterprise.

Where it is served

Anthropic lists the Claude API, Amazon Bedrock, Claude Models on Google Cloud (Google managed and Anthropic managed), Microsoft Foundry and Claude Platform on AWS. Of the hosts in our directory, AWS Bedrock is the one to read. Partner platforms set their own lifecycle dates, and regional or multi-region endpoints carry a 10 percent premium over global ones. Using inference_geo: "us" on Anthropic's own API adds 1.1x.

One practical wrinkle: Claude Code's opus alias is Opus 5.5 on the Anthropic API and on Bedrock and Google Cloud's agent platform, but Opus 4.6 on Microsoft Foundry, per Claude Code's model docs. For a wider look at hosting choices, see AI hosting.

Which coding agents can use it

Claude Code uses Opus through the opus alias, plus a mode called opusplan that uses Opus during planning and then Sonnet for execution. GitHub Copilot lists Opus 4.8, 5 and 5.5. Cursor has a model page for Opus 5.5 and bills it at the list price from prepaid credit, on Premium plans and above. Cline and OpenCode can reach it through an Anthropic key. Kiro lists Claude models, and its pricing page names Opus 5 and Opus 4.8 but not Opus 5.5.

Cost in practice

Two scenarios from the list prices, our arithmetic and not a measurement. A call with 50,000 input tokens and 2,000 output tokens costs $0.20 for input and $0.04 for output, so $0.24, before caching and thinking tokens. If 45,000 of those input tokens are cache reads at $0.20 per million, input drops to $0.029 for the fresh tokens plus $0.009 for the cache hits, about $0.038, so the call costs about $0.078. On the Batch API, halve it.

Agent loops resend the same context on every turn, which is where cache reads at 5 percent of input matter. Anthropic's launch post says cache reads make up the majority of agentic and coding work costs. For subscription users, plan access to Opus is a separate topic: see Claude pricing. For API detail, see Claude API pricing.

Opus history, briefly

Opus is the top of Anthropic's mid-tier ladder, and the 5.5 release follows a busy year. The pricing page still lists Opus 5 at $5 and $25, Opus 4.8, 4.7, 4.6 and 4.5, and the retired Opus 4.1 at $15 and $75 and Opus 4. The legacy models stay available, and Anthropic's migration guide covers a move from Opus 5 or earlier. If you have older code, the trend is clear: each generation has cut the list price, and Opus 5.5 at $4 and $20 is a long way below the $15 and $75 of Opus 4.1.

For an agent that still pins Opus 4.8 or 4.7, the reasons to move are the lower price and the newer tokenizer behaviour, and the reason to wait is the breaking changes listed above. Test on a copy of your real traffic.

Local hardware needs

None. Opus 5.5 cannot be downloaded. If you need a local or self-hosted model, start from open source LLMs or the coding-focused best open source LLM for coding.

Who should pick it

Our take

Opus 5.5 is a price cut as much as a model. At $4 and $20, with cache reads at $0.20, it undercuts its own predecessor by a clear margin, and it narrows the distance to Sonnet. That makes the old default of "use Sonnet unless you must" less automatic. For agent loops that resend long context, the cheap cache reads can make Opus closer to Sonnet's bill than the list prices suggest.

We would still start most tasks on Sonnet, and move to Opus where a first attempt keeps failing or the task spans many files. The always-on thinking is a real change to plan for: latency has a floor, and you control cost through effort. All of this comes from Anthropic's documentation and launch post. We have not tested the model.

Checklist

• Set effort explicitly, because medium is the default.
• Remove thinking: disabled and forced tool_choice from existing code.
• Turn on prompt caching for every reused prefix.
• Keep conversations append-only if you replay thinking blocks.
• Use the Batch API for work that can wait.
• Pin claude-opus-5-5, not an alias, in production.
• Recheck pricing before committing; this page is dated 31 October 2026.

Questions we hear

Is Opus 5.5 better than GPT-6 Astra? Anthropic's table favours Opus on most rows and Astra on two, but it is Anthropic's table. See GPT models for the OpenAI side.

Does Opus work with images? Yes, text and images in, text out.

How long will it be available? Anthropic commits to not retiring it before 22 September 2027 on platforms it operates.