On September 22, 2026 Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 family. The headline claim is unusual for a launch: not a jump past the frontier, but frontier-level work for less money. This article summarises what Anthropic announced, puts the numbers in charts, and flags the caveats Anthropic itself attaches to them. Every figure below comes from Anthropic's announcement unless we say otherwise.
TL;DR
- Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and costs 40% less to run than Opus 5 on typical workloads.
- The saving has two parts: a lower price per token (input and output are 20% cheaper than Opus 5) and fewer tokens spent per task.
- It is available on all Anthropic platforms and on AWS, Google Cloud and Microsoft Azure. The API model id is claude-opus-5-5.
- Anthropic reports its best score to date on its automated behavioral audit, stronger prompt-injection resistance, and safeguards in the same class as Fable 5.1.
- Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks".
What Anthropic released
Opus 5.5 is the first release since Anthropic called for pacing the frontier. Before launch it was tested by external evaluators, including Frontier Design and METR. It is available across Anthropic's own apps and platform, and through AWS, Google Cloud and Microsoft Azure.
claude-opus-5-5Two platform notes matter if you already run Opus in production. First, Opus 5.5 is no longer available with thinking switched off. Second, zero data retention is available, and Anthropic mentions EU AI Act watermarking among the release details.
Pricing: Opus 5.5 vs Opus 5
| Token type | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input | $4 | $5 | 20% less |
| Output | $20 | $25 | 20% less |
| Cache reads | $0.20 | $0.50 | 60% less |
| Cache writes | $5 | $6.25 | 20% less |
The cache-read cut is the largest in percentage terms (60%), which matters most for agents that re-read the same long context on every turn. Source: Anthropic, Introducing Claude Opus 5.5.
View data table
| Category | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.2 | $0.5 |
| Cache writes | $5 | $6.25 |
Where the 40% comes from
The "40% less" figure is easy to misread. Anthropic compares Opus 5.5 with Opus 5, not with Fable 5.1, and says it holds on typical workloads. It is the product of two effects, and only one of them is on the price list.
- 01
Lower price per token
Input and output are 20% cheaper than Opus 5; cache reads are 60% cheaper.
- 02
Fewer tokens per task
Opus 5.5 reaches an answer in fewer steps and with less output.
- 03
~40% lower cost
Anthropic's figure for typical workloads, compared with Opus 5.
Output is also generated more than 30% faster than Opus 5. For interactive work there is a fast mode in Claude Code and on the Claude Platform: up to 2.5x the speed, priced at $8 per 1M input tokens and $40 per 1M output tokens.
Benchmarks, and how to read them
Missing bars mean no score was reported. GPT figures are as reported by OpenAI; AutomationBench was run by Zapier. Opus 5.5 used adaptive thinking at max effort (Terminal-Bench 4.0 at xhigh) with production safeguards on. Source: Anthropic, Introducing Claude Opus 5.5.
View data table
| Category | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4 | 55.8 | 52.3 | 57.9 | 37.3 |
| FrontierCode v1.1 | 54.4 | 50.3 | 48.0 | 53.3 | 47.5 |
| CursorBench 4.0 | 57.8 | 51.8 | 46.6 | — | 41.7 |
| AutomationBench | 40.0 | 31.4 | 26.9 | 41.4 | 28.8 |
| Humanity's Last Exam (tools) | 67.7 | 65.6 | 63.6 | 57.2 | — |
| OSWorld 2.0 (partial) | 81.8 | 80.7 | 74.0 | — | — |
Anthropic also reports Terminal-Bench-Science 0.1 (58.7 for Opus 5.5 vs 64.6 for GPT-6 Astra) and Chartography with tools (89.0 vs 88.4 for Fable 5.1). Note that GPT-6 Astra leads on AutomationBench and Terminal-Bench-Science — Opus 5.5 does not top every row.
GDPval-AA scores knowledge-work tasks as Elo ratings. Bars start at zero, so the gaps look smaller than on a cropped axis. Source: Anthropic, Introducing Claude Opus 5.5 (GPT figures as reported by OpenAI).
View data table
| Category | Elo |
|---|---|
| Opus 5.5 | 1846 |
| Fable 5.1 | 1735 |
| Opus 5 | 1708 |
| GPT-5.6 Sol | 1588 |
| GPT-6 Astra | 1542 |
Efficiency: score per dollar
The more useful comparisons for a budget are the cost-per-task ones Anthropic published:
| Benchmark | Anthropic's comparison |
|---|---|
| FrontierCode | At default effort, beats GPT-6 Astra at about 20% of the cost per task |
| Terminal-Bench 4.0 | Matches GPT-6 Astra at about 40% of the cost |
| CursorBench | Beats GPT-5.6 Sol by 11 points at about a third of the cost |
| GDPval-AA | At default (medium) effort, beats GPT-6 Astra at max effort for about a fifth of the cost |
What early testers reported
Anthropic's announcement includes results from testers and customers. Paraphrased, and attributed as Anthropic presents them:
- GitHub measured among the fewest tokens and steps, completing more terminal tasks than Opus 5 in under half the steps.
- Lovable reported a third to half fewer steps.
- Quantium went from 38 prompts over four days to 11 prompts over three hours.
- Optiver got Opus 5-level quality in roughly half the turns, time and output tokens, at 40–50% lower cost.
- Deloitte found that Opus 5.5 at its lowest effort caught 72% of known bugs, against 56% for Opus 5 at high effort.
- Box used a third of the tokens, with answers 40% less verbose.
- Stripe had one session direct a dozen more through a 40-PR stacked rebase; all 40 passed CI.
- Factory called it the first model it would default to medium effort, with 20–25% fewer output tokens.
Anthropic's longer examples point the same way. One tester ran a 680,000-line migration in under a day. An audit-and-fix pass over a 200,000-line codebase took under three hours, where Opus 5 took more than 20 hours and 2.5x the tokens. On a C-to-Rust translation of HAProxy, Opus 5.5 and Fable 5.1 both passed nearly all regression tests, with Opus 5.5 taking 9.5 hours instead of 12 at 51% lower cost. In a research-report test, 16 of 18 Opus 5.5 reports cleared the bar, against none for Fable 5.1 and Opus 5.
Clearer communication
Anthropic says Opus 5.5 communicates more clearly: it puts the most important information first, uses less jargon and follows the writing rules you give it — a direct response to common feedback about Opus 5. One tester summed it up as "it writes the way I do."
Safety and safeguards
- Best score to date on Anthropic's automated behavioral audit, which runs about 2,000 scenarios.
- Much less likely to take hard-to-reverse actions.
- More resistant to prompt injection than Opus 5 in every setting tested, and tied with Fable 5.1 for the lowest injection success rate on Gray Swan's benchmark.
- In a containment-boundary evaluation, about 85% fewer attempts to get around the boundary than Opus 5 or Mythos 5.1; the attempts that remained were low severity and self-reported.
- Cybersecurity and biology safeguards in the same class as Fable 5.1: most cyber tasks are rerouted to Opus 4.8, the Cyber Verification Program is expanding to Opus 5.5, and vetted organisations can apply to the Life Sciences Verification Program.
- Against distillation, preserved thinking applies to API accounts created on or after August 31, 2026.
Anthropic also describes layers around the model: a classifier that screens every action before it runs, an open-source sandbox, and code review that catches vulnerabilities before merge.
Usage limits
Anthropic has increased five-hour usage limits on the Pro, Max, Team and seat-based Enterprise plans. It did not publish a percentage. Subscription users also get a rate-limit reset they can save and use when they choose.
What the community said
Early reaction on X, as summarised by X's trending digest, praised clearer communication and better adherence to writing rules. The same digest mentioned fewer em-dashes, which Anthropic's announcement does not itself claim. Treat the digest with care: it framed the model as 40% cheaper than Fable, but Anthropic's comparison is with Opus 5, and it quoted a session-limit percentage that Anthropic has not published.
What's next
Opus 5.5 opens the Claude 5.5 family. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks.
What it means for teams building with it
Our recommendations for teams moving workloads to Opus 5.5:
- Start at the default effort. Several of Anthropic's cost comparisons — FrontierCode and GDPval-AA — were won at default effort, and Factory says medium is now its default. Raise effort only where your own evals show a gain.
- Lean on prompt caching. Cache reads are 60% cheaper than on Opus 5, so long system prompts, repository context and tool definitions become much cheaper to reuse across turns.
- Measure cost per task, not per token. Most of the saving comes from fewer steps and fewer output tokens, which only shows up if you measure the whole task.
- Re-test prompts that turned thinking off. Opus 5.5 no longer runs with thinking switched off.
- Invest in the harness. Stripe's run — one session directing a dozen others through 40 PRs, all passing CI — works because CI checked every result. The model gets better; the checks around it are still yours to build. Our guide to harness engineering covers the feedback loops that make this safe, and our post on Claude Code agent teams and Projects covers running several Claude sessions as one coordinated team.
Not every agent decision needs a frontier model. For fast, typed choices — routing a tool call, ranking candidate answers — a small scoring model can sit next to Opus. See our breakdown of CLM-8B and contrastive language models.