Anthropic didn’t just release a new model on September 22, 2026. They released a model that completed a 680,000-line code migration in less than a day. A model that audited and fixed a 200,000-line codebase in under three hours, while its predecessor needed over 20. A model that caught a one-digit indexing error buried in a client’s own evaluation instructions… and then flagged it because it “noted that this would cost it points with the grader.”
That last one is the detail that sticks. It’s not just the performance. It’s the judgment.
Claude Opus 5.5 is Anthropic’s new frontier model and the first in the Claude 5.5 family. Here’s what it is, what it does, what it costs, and how it compares to everything else currently on the market.
Claude Opus 5.5 is Anthropic’s new leading model, succeeding Claude Opus 5 and positioned as the primary choice for complex agentic coding, knowledge work, and multi-step autonomous workflows. Per Anthropic’s official announcement, it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.”
That sentence is doing a lot of work. Fable 5.1 is Anthropic’s previous ceiling model, the one with safeguards so tight that certain biology and cybersecurity tasks require a verification program to access. Opus 5.5 matching it on most tasks at a lower price point than Opus 5 is, by any reasonable measure, a significant shift in the value equation.
The model ID is claude-opus-5-5. It’s available now on the Claude Platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry.
The improvements are substantial enough that they’re worth going through systematically, because “it’s better” covers a lot of ground here.
Opus 5.5 is a major step up from Opus 5 on real-world tasks, not just benchmark scores. The key distinction Anthropic draws: it’s not just more capable, it’s more efficient per task, using fewer tokens to complete the same work.
Per Anthropic’s official data:
This is the headline improvement for teams running production workloads. Opus 5.5 costs 40% less than Opus 5 on typical workloads. That number comes from two things happening at once: cheaper per-token pricing, and the model using fewer tokens per task to reach the same outcome.
Spotify’s engineering team confirmed this in early testing: “We’ve seen a clear improvement in token efficiency across our internal evaluations, as we’ve been able to complete the same tasks both cheaper and faster.”
Optiver ran Opus 5.5 on agentic coding tasks: “It matched Opus 5’s quality in about half the turns, time and output tokens, cutting the cost of that workload by 40 to 50%.”
One of the most consistent pieces of feedback about Opus 5 was that its outputs, particularly on coding and analysis tasks, were verbose and difficult to scan. Anthropic addressed this directly in 5.5.
Opus 5.5 puts the most important information first. It uses less jargon. It follows writing rules you give it. Multiple early testers commented that it writes the way a good senior colleague writes: clear, direct, no extra padding.
Ramp’s staff engineer: “Verbose, hard-to-follow output has been my biggest frustration with frontier models, and Claude Opus 5.5 fixes it. It writes like a good colleague, and follows our writing rules. When it rewrote one of our prompts I preferred its version to my own.”
Stripe’s engineering team ran a multi-day rebase of 40 stacked pull requests using Opus 5.5. After the session, the engineer could pick up context and answer clarifying questions in minutes, even after hours away from the work. All 40 passed CI the next afternoon.
Opus 5.5 generates output more than 30% faster than Opus 5 at standard settings. Fast mode, available in Claude Code and the Claude Platform, offers up to 2.5x speed at higher per-token cost.
Per Anthropic’s official announcement, all Opus 5.5 results use adaptive thinking at max effort unless otherwise noted.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 Main | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | N/A | 41.7% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench (business workflows) | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity’s Last Exam (with tools) | 67.7% | 65.6% | 63.6% | 57.2% | N/A |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 (computer use, partial) | 81.8% | 80.7% | 74.0% | N/A | N/A |
| Chartography (visual chart recognition) | 89.0% | 88.4% | 83.4% | N/A | N/A |
A few important context notes from Anthropic’s own footnotes:
Terminal-Bench 4.0 numbers for Opus 5.5 are reported at “xhigh” effort and GPT-6 Astra at “high” effort as reported by OpenAI, representing each model’s highest published score. Opus 5.5 was evaluated with its production safeguards enabled, meaning that on tasks where cybersecurity or biology safeguards intervened, those tasks were completed by a fallback model. This likely reduces Opus 5.5’s published scores on those specific benchmarks compared to what the base model would achieve without safeguards active.
The efficiency story is where the numbers get particularly interesting. Per Anthropic’s published data:
The consistent pattern: frontier benchmark results at fractions of competitor costs per task.
Per Anthropic’s official pricing page:
| Token type | Opus 5.5 | Opus 5 |
|---|---|---|
| Cache reads | $0.20/M | $0.50/M (60% less) |
| Input tokens | $4.00/M | $5.00/M (20% less) |
| Output tokens | $20.00/M | $25.00/M (20% less) |
| Cache writes | $5.00/M | $6.25/M (20% less) |
The cache read price is the most significant number for agentic and coding workloads. At $0.20 per million tokens, it’s 60% cheaper than Opus 5’s cache read rate. In agentic coding sessions, the majority of input costs are cache reads (repeated context, system prompts, large codebases kept in memory), so the effective cost reduction on real workloads typically exceeds the 20% headline input discount.
Fast mode is available in Claude Code and the Claude Platform at up to 2.5x speed, priced at $8 per million input tokens and $40 per million output tokens.
claude-opus-5-5This is the primary use case the model was built for, and where the performance gaps over prior models are most dramatic.
Opus 5.5 is designed for long, sprawling engineering work. Not one-shot code generation, but sustained multi-hour sessions across large codebases where maintaining context and judgment over time matters as much as raw capability.
Codebase migrations at scale. According to Anthropic’s announcement, an early tester completed a 680,000-line migration in under a day, and a separate tester audited and fixed a 200,000-line codebase in under three hours. These are examples from Anthropic’s early access program illustrating the class of work the model targets.
HAProxy translation (internal test). Anthropic asked both Opus 5.5 and Fable 5.1 to translate HAProxy, the widely-used load balancing software written in C, into Rust. Both rewrites passed nearly all of HAProxy’s own regression tests. Opus 5.5 finished in 9.5 hours compared to 12 for Fable 5.1, and cost 51% less.
Agentic multi-session orchestration. Stripe’s staff engineer ran a multi-day rebase of 40 stacked pull requests. A single Opus 5.5 session directed a dozen additional sessions, laid out every conflict clearly, and held calls in a way that was immediately followable after hours away. All 40 pull requests passed CI.
GitHub Copilot integration. Per GitHub’s own testing in Copilot CLI and VS Code, Opus 5.5 used among the fewest tokens and steps measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps.
Optiver’s Global Head of AI Engineering: “On our agentic coding tasks, Claude Opus 5.5 matched Opus 5’s quality in about half the turns, time and output tokens, cutting the cost of that workload by 40 to 50%. It posted the highest score we’ve recorded on one desk’s trading-support suite, passing tasks earlier Claude models had failed.”
Kiro’s VP of Agentic AI: “On a public benchmark of real command-line tasks, Claude Opus 5.5 solved more than Opus 5 while making about 40% fewer calls and using half the tokens.”
Lovable’s CTO: “It gathers context once, makes fewer and more complete edits, and doesn’t get stuck retrying, finishing in a third to half fewer steps and using significantly fewer tokens.”
For enterprises running agents unattended across systems, Opus 5.5 ships with:
The coding numbers get most of the attention. The knowledge work numbers deserve as much.
Financial analysis. Walleye Capital, an investment firm, ran Opus 5.5 on their evaluation suite. At its lowest effort setting, it largely solved the task. At higher settings, it went further: it detected that the minute indexing in their own instructions was off by one, corrected for it, and noted that this would cost it points with the grader. No model they’d tested had caught that before.
Research accuracy. An internal test asked Opus 5.5, Fable 5.1, and Opus 5 to write a report on a company’s quarterly performance using only information available on a web copy where the earnings release was hard to locate. An automated grader checked every figure and quote against sources. 16 out of 18 of Opus 5.5’s reports cleared the quality bar (any invented figure or quote would have failed). Neither Fable 5.1 nor Opus 5 cleared that bar in any attempt.
M&A analysis task. Anthropic asked Opus 5.5 and Opus 5 to analyze a proposed merger between two fictional HR software companies: build a financial model in Excel, then create an executive presentation. Both models reached the same conclusions. Opus 5.5’s model was more thorough, its presentation easier to read, and Opus 5 had minor errors. Opus 5.5 finished in 63 minutes versus 93 for Opus 5, at 50% less cost.
GDPval-AA v2.1. This is Artificial Analysis’s benchmark testing agents on real-world professional work across 44 occupations. Opus 5.5 scores 1846 Elo, ahead of Fable 5.1’s 1735 and Opus 5’s 1708. At default (medium) effort, Opus 5.5 beats GPT-6 Astra at max effort for about one-fifth of the cost per task.
What enterprise users reported:
Deloitte: “Even at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5’s 56% at high effort, with fewer false alarms and a fraction of the output.”
Rogo: “At its lowest effort setting, Claude Opus 5.5 beat Opus 5 at high effort on our BigFinance Bench with about 60% fewer output tokens.”
Hebbia: “On end-to-end finance workflows graded against expert rubrics, Claude Opus 5.5 covered 86.6% of what we look for versus 60.3% for Opus 5.”
Viktor: “At the same effort, Claude Opus 5.5 needs fewer steps and tool calls per task than Opus 5 and costs nearly half as much, while getting twice as many of our hardest tasks right.”
Opus 5.5 uses adaptive thinking. At launch, max effort is the only available setting via the API, with Anthropic noting that low- and high-effort modes will be introduced in subsequent updates. The benchmarks in Anthropic’s announcement use max effort unless otherwise noted, but the product’s default effort level for day-to-day use is separate from this benchmark setting and varies by context.
Thinking cannot be switched off on Opus 5.5. Anthropic has documented this explicitly. This is a design decision, not a limitation: always-on reasoning is tied to both the model’s capability ceiling and its alignment behavior. If you need a version of Opus with togglable thinking, Opus 5 remains available.
The model ID for API access: claude-opus-5-5. The context window is 1 million tokens, with a 128K output limit.
This section matters more than most AI model releases because Anthropic has been unusually transparent about it.
On Anthropic’s primary evaluation suite (an automated behavioral audit across nearly 2,000 scenarios), Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior. Specific improvements over prior models:
That last point is worth sitting with. A model that attempts a containment-boundary test and self-reports the attempt is not behaving the way a model trying to deceive its evaluators would behave. That’s the alignment property that matters for teams running Claude unattended overnight.
Anthropic also acknowledges the limits of their testing openly: “We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act in the vast variety of real-world settings it is deployed in.” That’s not marketing copy. That’s a company publishing its own evaluation gaps in the same document where it’s announcing its best model.
Because Opus 5.5’s capabilities in biology and cybersecurity are comparable to Claude Mythos 5.1, it ships with safeguards similar to Fable 5.1.
Cybersecurity. Most cybersecurity tasks re-route to Claude Opus 4.8. Standard bug identification and fixing as part of software development is unaffected. Verified cybersecurity practitioners can apply to the Cyber Verification Program for Opus 5.5 access. Claude Security (available to enterprise teams) already has access to Claude Mythos 5.1.
Biology. Opus 5.5 matches or beats Mythos 5.1 across many areas of biology work, per internal testing in collaboration with Dyno Therapeutics. The same biology safeguards as Fable 5.1 apply. Vetted organizations (academic labs, startups, pharmaceutical companies) can apply to Anthropic’s Life Sciences Verification Program for access to the full range of biology-related work.
Anti-distillation. Opus 5.5 launches with preserved thinking, the anti-distillation safeguard introduced with Fable 5.1. It stops API users from editing Claude’s prior context to extract reasoning at industrial scale. Applies to API accounts created on or after August 31, 2026.
Watermarking. As with Fable 5.1, Opus 5.5 includes watermarking measures to comply with the EU AI Act.
Claude app (claude.ai): Available now on Pro, Max, Team, and Enterprise plans. Anthropic is also increasing five-hour usage limits across all paid plans and providing a rate limit reset that subscribers can save and use whenever they choose.
Claude Code: Available now. Select claude-opus-5-5. Fast mode available in Claude Code with up to 2.5x speed.
API: claude-opus-5-5 at api.anthropic.com. Standard Anthropic API access. Zero data retention available as with previous Opus models.
Amazon Bedrock: Available now.
Google Cloud Vertex AI: Available now.
Microsoft Azure Foundry: Available now.
Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements to performance, efficiency, and safety.
Anthropic is direct about this: “In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.”
The benchmark scores show Opus 5.5 leading Fable 5.1 on several evaluations (Terminal-Bench 4.0: 66.4% vs 55.8%; GDPval-AA: 1846 vs 1735; FrontierCode: 54.4% vs 50.3%). But Anthropic explicitly notes that at these capability levels, benchmark margins have become a less reliable guide to real-world differences.
The practical question is not which model scores higher on a leaderboard. It’s cost efficiency at equivalent task quality. And there, Opus 5.5 makes a strong case: it beats Fable 5.1’s performance on most benchmarks at substantially lower per-token cost, with faster generation speed.
For teams currently on Fable 5.1: the performance delta between the two models on most practical tasks is narrow. The cost and speed advantage of Opus 5.5 is significant.
For teams on Opus 5: the performance upgrade is substantial, and the cost is lower. This is the clearest upgrade case.
Claude Opus 5.5 is Anthropic’s most capable model as of September 2026, the first in the Claude 5.5 family. Per Anthropic’s official announcement, it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. The model ID is claude-opus-5-5.
Per Anthropic’s published pricing: $4 per million input tokens, $20 per million output tokens, $0.20 per million cache read tokens (60% less than Opus 5), and $5 per million cache write tokens. Fast mode costs $8 input and $40 output per million tokens.
Opus 5.5 is a major step up. It completes the same work with fewer tokens (users report 40 to 50% cost reduction on equivalent tasks), generates output 30%+ faster, leads Opus 5 on all major benchmarks, and has substantially better alignment scores. It’s also cheaper per token across all categories.
It leads Fable 5.1 on several key benchmarks, including Terminal-Bench 4.0 (66.4% vs 55.8%), GDPval-AA (1846 vs 1735 Elo), and FrontierCode. Anthropic notes that in their own internal use, the performance gap is narrower than benchmark scores suggest. The practical advantages of Opus 5.5 over Fable 5.1 are cost efficiency and generation speed.
1 million tokens (1M context window) with a 128K maximum output limit.
Yes. It’s specifically designed and optimized for long-horizon agentic coding tasks, codebase migrations, audits, debugging, code review, and refactoring. It’s available in Claude Code and integrates with GitHub Copilot, Kiro, and other developer tools.
Yes to all three: Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure Foundry.
A speed-optimized inference option available in Claude Code and the Claude Platform, offering up to 2.5x faster output generation at $8 per million input tokens and $40 per million output tokens.
No. Thinking cannot be switched off on Opus 5.5, per Anthropic’s documentation. At launch, max effort is the only available API setting. Low- and high-effort modes will be added in future updates. If you need togglable thinking, Opus 5 remains available.
Opus 5.5 ships with Fable 5.1-level safeguards on cybersecurity, biology, and distillation. Cybersecurity tasks re-route to Opus 4.8; biology tasks are restricted with a verification program available for vetted organizations; preserved thinking (anti-distillation) applies to API accounts created after August 31, 2026.
There’s a version of this where Claude Opus 5.5’s release is a fairly routine model upgrade story: better benchmarks, new features, lower prices. And on paper, that’s what it is.
But the actual texture of the changes suggests something slightly more interesting is happening. The model that catches a one-digit indexing error in a client’s own evaluation instructions and self-reports its containment-boundary attempts… that’s not just a more capable model. That’s a model that’s starting to exhibit the kind of judgment and transparency that makes long autonomous sessions feel less like a liability management exercise and more like working with a competent colleague who happens to never sleep.
Whether that colleague is worth $4 per million input tokens is a math problem your specific workloads will answer. The efficiency numbers Anthropic and its early testers are reporting, 40 to 50% lower costs on equivalent tasks, half the steps, a third of the tokens, are consistent across a range of different companies and different use cases.
The HAProxy rewrite in 9.5 hours at 51% less cost than Fable 5.1. The 200,000-line codebase audit in under three hours. The M&A analysis in 63 minutes versus 93.
Opus 5.5 is the model you run when the task is genuinely hard and you need it done right… without burning your entire API budget to find out.
Most benchmark, pricing, performance, and customer-test data in this article is based on Anthropic’s September 22, 2026 announcement at anthropic.com/claude-opus-5-5 and related Anthropic documentation. Customer quotes are attributed to named individuals and organizations as published on the announcement page. Some contextual framing and structural organization are the author’s own. Model capabilities and safeguards are subject to change; verify current details at the Anthropic platform documentation.
Curated by Lorphic
Digital intelligence. Clarity. Truth.
We have sent a 6-digit verification code to your email. Please enter it below to continue.
We have sent a 6-digit verification code to your email. Please enter it below to continue.
