Claude Opus 5.5: What Changed

The short version: Anthropic released Claude Opus 5.5 on September 22, 2026. It costs 40% less to run than Opus 5 on typical workloads, generates output about 30% faster, and posts new highs on agentic coding, computer use, and knowledge-work benchmarks. Pricing is $4 per million input tokens and $20 per million output tokens, both down 20% from Opus 5. It is live now in the Claude apps and the API as claude-opus-5-5.

I keep a running note of every Claude release because the pricing and the benchmark claims change what I recommend to people who ask “which model should I pay for.” Opus 5.5 is a bigger jump than the version number suggests, so here is what changed, what the numbers mean if you are not a benchmark nerd, and whether it is worth switching.

What is Claude Opus 5.5?

Opus 5.5 is Anthropic’s update to its flagship reasoning model, replacing Opus 5 as the top tier in the Claude lineup. It sits above Fable 5.1 and Sonnet 5 on Anthropic’s own model tier chart, aimed at the hardest coding, research, and multi-step agent work rather than everyday chat.

Anthropic’s release notes describe an early tester who used it to finish a 680,000-line code migration in under a day, work that would normally take an engineering team weeks. It’s a single anecdote rather than a benchmark, but it points at the intended use case: large, messy, multi-file work that used to need a team babysitting it.

The benchmark numbers, in plain English

Anthropic published a comparison table against its own prior models (Fable 5.1, Opus 5) and against OpenAI’s GPT-6 Astra. Here are the headline rows:

BenchmarkOpus 5.5Fable 5.1Opus 5GPT-6 Astra
Agentic coding (Terminal-Bench 4.0)66.4%55.8%52.3%57.9%
Knowledge work (GDPval-AA v2.1)1846173517081542
Multidisciplinary reasoning (Humanity’s Last Exam, with tools)67.7%65.6%63.6%57.2%
Computer use (OSWorld 2.0, partial)81.8%80.7%74.0%—
Business workflows (AutomationBench)40.0%31.4%26.9%41.4%

Two things stand out. First, Opus 5.5 beats its own larger sibling, Fable 5.1, on most of these. That’s unusual: Anthropic normally saves the biggest jumps for a size increase, not a point release. Second, GPT-6 Astra still wins on AutomationBench, the one benchmark built around Zapier-style business workflows, so this is not a clean sweep. If your work is mostly agent-driven business automation rather than coding or research, that gap is worth knowing about before you switch.

Anthropic also says Opus 5.5 posted the best scores of any Claude model to date on its own automated behavioral audit, the internal alignment test suite that runs the model through thousands of simulated scenarios. That is a self-reported number with no outside benchmark to check it against, so treat it as a company claim, not an independent result.

How much does Opus 5.5 cost?

ItemOpus 5.5Opus 5
Input tokens (per 1M)$4.00$5.00
Output tokens (per 1M)$20.00$25.00
Cache reads (per 1M)$0.20$0.50
Cache writes (per 1M)$5.00$6.25

That is a straight 20% cut across input and output pricing, and a much bigger cut on cached tokens, which matters if you run long agent sessions that re-read the same context repeatedly. Anthropic frames this as a 40% total-workload savings versus Opus 5, combining the lower per-token price with the model needing fewer tokens to finish the same job. See the full breakdown, including Pro and Max plan limits, on our Claude pricing page.

Should you switch from Opus 5?

If you are on a paid Claude plan already, yes, there is no reason not to. Opus 5.5 is available now inside the Claude apps on Pro, Max, Team and Enterprise, so switching just means picking it from the model dropdown, no new purchase involved. Anthropic also raised the five-hour usage limits on those paid plans with this launch (it did not publish the new numbers) and gave subscribers a one-time rate limit reset they can save and use any time before October 22. If you call the API directly, swap your model string to claude-opus-5-5 and rerun your eval suite before trusting it on production traffic; a faster, cheaper model is not automatically a drop-in replacement for your specific prompts.

Coders and heavy agent users get the biggest win here: the Terminal-Bench score jumped 14 points over Opus 5. Read our Claude Code guide if you have not tried the agentic coding workflow yet.

Where you can use it right now

Opus 5.5 is live in Claude.ai on the paid plans (Pro, Max, Team and Enterprise), in the Claude Platform API as claude-opus-5-5, and through AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure. Anthropic also added a “preserved thinking” safeguard aimed at making the model’s reasoning traces harder to strip out and reuse to train a rival model, and a zero data retention option for enterprise customers who cannot have their prompts logged at all. Neither changes how the model answers you day to day, but both are worth knowing if procurement or compliance ever asks.

Common questions about Opus 5.5

Is Opus 5.5 available for free Claude.ai accounts?

No. Claude’s pricing page lists Haiku and Sonnet for the free plan, and Opus models only on paid plans starting with Pro at $20 a month. Free users stay on Sonnet 5 and Haiku. See our Claude pricing guide for what each plan includes.

Does Opus 5.5 replace Fable 5.1?

No. Fable 5.1 is Anthropic’s larger, separately positioned model, released three weeks earlier on September 1. Opus 5.5 matches or beats it on most published benchmarks while costing less to run, but Anthropic has not announced retiring Fable 5.1.

What is Terminal-Bench and why does it matter?

Terminal-Bench 4.0 measures how well a model can complete real command-line and coding tasks end to end, not just write a snippet, so it is one of the better proxies for whether a model can finish a coding job unsupervised.

Did the context window change?

Anthropic’s own release page does not list a new context window figure for Opus 5.5. Several third-party blogs published specific numbers within hours of launch that we could not verify against Anthropic’s materials, so we are leaving that line out until Anthropic’s docs confirm it rather than repeat an unverified figure.

Get releases like this the day they happen

We track every Claude, Gemini, and ChatGPT release the same day it ships and break down what changes for regular users, not just developers. Join the daily newsletter to get it in your inbox.

Sources

You may also like

Two ways to go further

The AI Prompt Library

1,000+ ready-to-use prompts for Claude, ChatGPT, and Gemini. Stop staring at a blank box.

Get it for $39 →

2-Hour Live AI Crash Course

A private, beginner-friendly session across Claude, ChatGPT, Gemini, and the wider landscape.

Book for $125 →