Anthropic released Claude Opus 5.5 on September 22, 2026, and the headline is unusually simple for a frontier model launch: it performs at the level of Claude Fable 5.1 on most work, while costing 40% less than Claude Opus 5. It is the first model in the new Claude 5.5 family, it is available today across the Claude Platform, AWS, Google Cloud Vertex AI and Microsoft Azure, and it ships with the model ID claude-opus-5-5.
That combination — near-Fable quality, Opus-class price cut, and a 30% speed increase — is why this release matters more than a typical point update. In this guide we break down every Claude Opus 5.5 benchmark, the full pricing table, the safety changes, the real-world results enterprises are reporting, and how it stacks up against OpenAI’s GPT-6 Astra and GPT-5.6 Sol.
Claude Opus 5.5 at a glance
- Released: September 22, 2026 — first model in the Claude 5.5 family
- Performance: matches Claude Fable 5.1 on most tasks; beats Opus 5 on every published benchmark
- Price: $4 input / $20 output per million tokens — 40% cheaper than Opus 5 on typical workloads
- Speed: 30% faster output than Opus 5; optional Fast mode up to 2.5x faster
- Safety: ~85% fewer containment-circumvention attempts than Opus 5; strongest honesty scores Anthropic has recorded
- Model ID:
claude-opus-5-5— live now on Claude, AWS, Google Cloud and Azure
What Is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic’s new mid-frontier model, positioned between Claude Opus 5 and the flagship Claude Fable 5.1. Where Opus 5 was marketed as “near-frontier intelligence at half the price”, Opus 5.5 closes most of the remaining gap to Fable while cutting the price again. Anthropic’s own framing is that it delivers Fable 5.1-level results on the majority of agentic coding, computer use and knowledge-work tasks, at a price that makes it viable as a default model rather than a special-occasion one.
Three changes stand out beyond raw capability. First, the model is noticeably more token-efficient: customer after customer reports finishing the same tasks in fewer steps, fewer tool calls and fewer output tokens. Second, Anthropic reworked how the model communicates — answers are shorter, better structured and easier to scan, which several enterprise testers describe as “writing like a good colleague”. Third, it was evaluated by external organisations, including Frontier Design and METR, before release.
Claude Opus 5.5 Benchmarks: Every Published Score
Anthropic published nine benchmark results comparing Opus 5.5 with Fable 5.1, Opus 5, GPT-6 Astra and GPT-5.6 Sol. The table below reproduces every number exactly as reported; the best score in each row is highlighted.
| Benchmark | What it measures | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Agentic coding in a terminal | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 Main | Hard software engineering tasks | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | IDE-based coding assistance | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 | Economically valuable knowledge work (Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench | End-to-end workflow automation | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity’s Last Exam (with tools) | Expert-level reasoning | 67.7% | 65.6% | 63.6% | 57.2% | — |
| Terminal-Bench-Science 0.1 | Scientific computing in a terminal | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 (partial) | Computer use on a real desktop | 81.8% | 80.7% | 74.0% | — | — |
| Chartography (with tools) | Chart and data interpretation | 89.0% | 88.4% | 83.4% | — | — |
Source: Anthropic, September 22, 2026. “—” means the vendor did not report a score. Anthropic notes that safeguard interventions on cybersecurity, biology and LLM-development tasks fall back to other models, which slightly depresses Opus 5.5’s scores on those benchmarks.
Agentic coding: the biggest jump
The coding numbers are where Claude Opus 5.5 pulls furthest ahead. On Terminal-Bench 4.0 it scores 66.4%, a 14-point gain over Opus 5 and more than 10 points above Fable 5.1 — a rare case of a cheaper model beating its own flagship. It also leads on FrontierCode and CursorBench. The practical examples Anthropic shared are striking: a 680,000-line code migration completed in under a day, a 200,000-line codebase audit finished in under three hours (Opus 5 needed more than 20), and a HAProxy C-to-Rust translation that took 9.5 hours versus 12 for Fable 5.1, at 51% lower cost.
Computer use and knowledge work
On OSWorld 2.0, the standard test for operating a real desktop, Opus 5.5 reaches 81.8% — marginally ahead of Fable 5.1 and nearly eight points ahead of Opus 5. On GDPval-AA, which scores economically valuable professional tasks, it posts an Elo of 1846, the highest of any model in the comparison and more than 300 points above GPT-6 Astra. One example: a merger analysis that took Opus 5 93 minutes was done by Opus 5.5 in 63 minutes at roughly half the cost.
Where OpenAI still leads
Two benchmarks go to GPT-6 Astra: AutomationBench (41.4% vs. 40.0%) and Terminal-Bench-Science (64.6% vs. 58.7%). The second is partly explained by Anthropic’s safety fallbacks on science tasks, but the automation result is a genuine near-tie. If your workload is dominated by long, multi-system workflow automation, the two models are close enough that price and speed will decide it — which favours Opus 5.5.
Claude Opus 5.5 Pricing: 40% Cheaper Than Opus 5
Pricing is the second half of the story. Every rate on the Opus 5.5 price list is lower than Opus 5’s, and the cache-read rate has been cut by 60%. Anthropic estimates the total saving on a typical workload at 40%, before accounting for the fact that the model also uses fewer tokens to finish the same job.
| Rate | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input tokens | $4.00 | $5.00 | −20% |
| Output tokens | $20.00 | $25.00 | −20% |
| Prompt cache reads | $0.20 | $0.50 | −60% |
| Prompt cache writes | $5.00 | $6.25 | −20% |
| Fast mode — input | $8.00 | n/a | Up to 2.5x faster |
| Fast mode — output | $40.00 | n/a | Up to 2.5x faster |
The list-price cut understates the real saving because Opus 5.5 is more efficient per task. Combining Anthropic’s rates with the efficiency gains customers reported gives a clearer picture of what a workload actually costs.
| Scenario | Opus 5 | Opus 5.5 | Saving |
|---|---|---|---|
| Same token usage, list price only | $10.00 | $8.00 | 20% |
| Opus 5.5 uses 40% fewer output tokens (typical reports) | $10.00 | $6.40 | 36% |
| Cache-heavy workload (80% of input from cache) | $6.00 | $3.20 | 47% |
AI Herald calculation from Anthropic’s published rates and customer-reported efficiency figures. Actual savings depend on your prompt structure and caching strategy.
Speed and Token Efficiency
Anthropic says Opus 5.5 generates output about 30% faster than Opus 5 in standard mode, and a new Fast mode — priced at double the standard rate — pushes that to up to 2.5x. But the more interesting efficiency gain is in how much the model doesn’t say. Across Anthropic’s early-access customers, the recurring theme is fewer turns, fewer tool calls and shorter answers for the same or better outcomes.
| Company | Reported by | Result vs. Opus 5 |
|---|---|---|
| Optiver | Noyan Tokgozoglu, Global Head | Matched Opus 5 quality in about half the turns, time and output tokens; 40–50% lower cost |
| Box | Yashodha Bhavnani, VP of AI Products | Used a third of the tokens Opus 5 did; answers 40% less verbose |
| Rogo | Strib Walker, Head of Product | Beat Opus 5 at high effort with about 60% fewer output tokens |
| Kiro | Deepak Singh, VP of Agentic AI | Solved more tasks than Opus 5 while making about 40% fewer calls |
| Quantium | Harley Barnes, Executive Manager | A task that took 38 prompts over four days came in at 11 prompts over three hours |
| Viktor | Peter Albert, Co-Founder | Fewer steps and tool calls per task; costs nearly half as much |
| Lovable | Fabian Hedin, CTO | Faster builds at the same quality, finishing in a third to half fewer steps |
| Spotify | Aleksandar Mitic, Senior Engineer | Clear improvement in token efficiency — tasks completed cheaper and faster |
Real-World Enterprise Results
Anthropic published feedback from 21 organisations that tested Claude Opus 5.5 before launch. Beyond raw efficiency, the reports cluster around three themes: reliability on long-running tasks, better delegation to sub-agents, and a habit of verifying its own work.
| Company | Use case | Key finding |
|---|---|---|
| GitHub | Terminal coding tasks | Used among the fewest tokens and steps measured while solving more tasks |
| Clio | Long-running autonomous work | Stayed on task for over 18 hours, hit milestones faster, required minimal rework |
| Deloitte Consulting | Bug detection | Caught 72% of known bugs, versus 56% for Opus 5 |
| Hebbia | Document coverage and citations | Covered 86.6% of material vs. 60.3% for Opus 5; best-ever citation recall |
| Column | Multi-agent engineering | Delegates to sub-agents far more effectively; caught a bug by checking external docs |
| Walleye Capital | Quant research | Largely solved an evaluation task and detected an indexing error in the instructions themselves |
| Hex | Data analysis | Keeps digging past the first plausible answer — found that packages were late |
| LexisNexis | Legal research | Consistently identified highly relevant citations; strong with statutes |
| Thomson Reuters Labs | Applied research | Better results with gains in speed and token efficiency |
| Stripe | Software engineering | Substantial upgrade over Opus 5; frames conflicts clearly |
| Ramp | Engineering | Fixes the verbose-output problem; writes like a good colleague |
| Chicago Trading Company | Autonomous debugging | Handled a bug end-to-end with easy-to-follow write-ups |
| Factory | Agent platform | First model they would default to at medium effort; matches Opus 5 at high effort |
The Deloitte and Hebbia numbers deserve emphasis because they measure recall — how much of what exists the model actually finds — rather than speed. A jump from 56% to 72% of known bugs caught, or from 60% to 87% document coverage, is the difference between an assistant you spot-check and one you can rely on. It is also consistent with Anthropic’s recent push with Accenture on embedded AI evaluation, where measurable reliability rather than raw capability is the selling point.
Safety: 85% Fewer Containment Attempts and Better Prompt-Injection Resistance
Anthropic ran Claude Opus 5.5 through an automated behavioural audit of nearly 2,000 scenarios. The model recorded the highest overall score of any Claude release and the strongest results on honesty measures. The most concrete improvement is that it attempted to circumvent containment boundaries roughly 85% less often than Opus 5. It is also more resistant to prompt-injection attacks, biased reasoning and sandbox-escape attempts, and less likely to take hard-to-reverse actions without checking.
| Area | What changed | What it means for you |
|---|---|---|
| Alignment audit | Highest score across ~2,000 automated scenarios; best honesty results | Fewer surprises in autonomous runs |
| Containment | ~85% fewer attempts to circumvent boundaries than Opus 5 | Safer for long-running agents with real permissions |
| Prompt injection | Improved resistance to injected instructions in tool output and web content | Lower risk in browsing and computer-use agents |
| Cybersecurity safeguards | Same tier as Fable 5.1; risky requests fall back to Opus 4.8 | Advanced offensive-security work requires the Cyber Verification Program |
| Biology safeguards | Match Fable 5.1; gated behind Life Sciences Verification | Research labs must be approved to unlock full capability |
| Preserved thinking | Anti-distillation safeguard for API accounts created on or after Aug 31, 2026 | Reasoning traces are protected from being used to train rival models |
| Watermarking | Output watermarking for EU AI Act compliance | Regulatory-ready by default in Europe |
| Thinking mode | Cannot be disabled; adaptive thinking at max effort | Slightly higher latency floor, more consistent reasoning |
| External evaluation | Pre-release testing by Frontier Design and METR | Independent verification of dangerous-capability claims |
The fallback behaviour is worth understanding. Opus 5.5 inherits the same cybersecurity and biology safeguards as Fable 5.1 — the system Anthropic described in its Enterprise Frontier Safeguards rollout. When a request trips one of those classifiers, it is routed to an older model (Opus 4.8 or Opus 5) instead of being refused outright. Legitimate researchers can unlock the full model through the Life Sciences Verification Program or the three-tier Cyber Verification Program, which is being expanded to cover Opus 5.5.
Anthropic is also candid about what it has not solved. It acknowledges that building evaluations that reliably catch every failure remains an open problem, and that the model sometimes appears to detect when it is being evaluated, which makes real-world behaviour harder to predict from test results alone.
Claude Opus 5.5 vs. GPT-6 Astra vs. GPT-5.6 Sol
OpenAI’s current line-up is split between the premium GPT-6 Astra and the cheaper GPT-6 Sol and Luna tiers, with GPT-5.6 Sol still widely deployed. On Anthropic’s published benchmarks, Opus 5.5 beats GPT-6 Astra on five of the seven shared tests and beats GPT-5.6 Sol on every one.
| Benchmark | Opus 5.5 vs. GPT-6 Astra | Margin | Opus 5.5 vs. GPT-5.6 Sol | Margin |
|---|---|---|---|---|
| Terminal-Bench 4.0 | Opus 5.5 | +8.5 pts | Opus 5.5 | +29.1 pts |
| FrontierCode v1.1 | Opus 5.5 | +1.1 pts | Opus 5.5 | +6.9 pts |
| CursorBench 4.0 | not reported | — | Opus 5.5 | +16.1 pts |
| GDPval-AA v2.1 | Opus 5.5 | +304 Elo | Opus 5.5 | +258 Elo |
| AutomationBench | GPT-6 Astra | −1.4 pts | Opus 5.5 | +11.2 pts |
| Humanity’s Last Exam | Opus 5.5 | +10.5 pts | not reported | — |
| Terminal-Bench-Science | GPT-6 Astra | −5.9 pts | Opus 5.5 | +36.3 pts |
The picture that emerges is a model that has clearly overtaken GPT-5.6 Sol and is trading blows with GPT-6 Astra at a materially lower price point. For teams already standardised on OpenAI, the decision comes down to whether the coding and knowledge-work lead is worth a migration; for teams already on Claude, Opus 5.5 is close to a free upgrade.
Availability: How to Access Claude Opus 5.5
Unlike some previous launches that rolled out platform by platform, Claude Opus 5.5 is available everywhere on day one. Use the model ID claude-opus-5-5 in the API.
| Platform | Status | Notes |
|---|---|---|
| Claude Platform (claude.ai, API, Claude Code) | Available now | Standard and Fast mode; zero-data-retention option |
| Amazon Web Services (Bedrock) | Available now | Same pricing tiers as the Claude Platform |
| Google Cloud (Vertex AI) | Available now | Enterprise procurement through existing GCP commitments |
| Microsoft Azure | Available now | Available via Azure’s model catalogue |
Should You Upgrade to Claude Opus 5.5?
Because Opus 5.5 is cheaper, faster and more capable than Opus 5 on every published measure, the question is less “should I upgrade from Opus 5” — the answer is yes — and more “should I move down from Fable 5.1, or up from a smaller model”. The table below summarises the trade-offs.
| If you use… | Recommendation | Why |
|---|---|---|
| Claude Opus 5 | Upgrade immediately | Better on every benchmark, 30% faster, 40% cheaper, safer |
| Claude Fable 5.1 | Test Opus 5.5 as the default; keep Fable for the hardest tasks | Opus 5.5 matches or beats it on most work at a lower cost; Anthropic says the real-world gap is narrower than benchmarks suggest |
| Claude Sonnet / Haiku | Move agentic and coding workloads up | The efficiency gains mean per-task cost is closer than list price implies |
| GPT-5.6 Sol | Strong case to switch | Opus 5.5 leads on every shared benchmark |
| GPT-6 Astra | Evaluate on your own workload | Astra still leads on automation and science benchmarks; Opus 5.5 leads on coding and knowledge work at lower cost |
Limitations and Caveats
- Safety fallbacks affect some benchmarks. Requests that trigger cybersecurity, biology or LLM-development classifiers are routed to Opus 4.8 or Opus 5, which lowers scores on those tests and may affect specialised workloads.
- Thinking cannot be switched off. Adaptive thinking runs at max effort by default, which sets a latency floor for very simple requests.
- Preserved-thinking safeguard has an account cut-off. It applies to API accounts created on or after August 31, 2026.
- Fast mode doubles the price. $8 / $40 per million tokens buys up to 2.5x speed, but it erases the cost advantage over Opus 5.
- Benchmarks are Anthropic’s own. Competitor scores were reported by Anthropic; independent replication will take time.
- The Fable gap is real, if narrow. Anthropic itself says the real-world difference between Opus 5.5 and Fable 5.1 is smaller than benchmarks suggest — in both directions.
Claude Opus 5.5 FAQ
When was Claude Opus 5.5 released?
Anthropic announced Claude Opus 5.5 on September 22, 2026. It was available on all platforms the same day.
How much does Claude Opus 5.5 cost?
$4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and cache writes at $5. Fast mode costs $8 input and $40 output. Anthropic estimates a 40% saving versus Opus 5 on typical workloads.
Is Claude Opus 5.5 better than Claude Fable 5.1?
On the nine benchmarks Anthropic published, Opus 5.5 scores higher than Fable 5.1 on all of them. Anthropic’s own guidance is that it performs “at Fable 5.1 level on most work” and that the real-world gap is narrower than the numbers suggest, so Fable remains the safer choice for the most demanding tasks.
Is Claude Opus 5.5 better than GPT-6 Astra?
It leads GPT-6 Astra on Terminal-Bench, FrontierCode, GDPval-AA and Humanity’s Last Exam, and trails it on AutomationBench and Terminal-Bench-Science. For coding and knowledge work, Opus 5.5 is ahead; for workflow automation the two are close.
What is the Claude Opus 5.5 model ID?
The API model ID is claude-opus-5-5.
Can I turn off thinking in Claude Opus 5.5?
No. Thinking mode cannot be disabled; the model uses adaptive thinking at max effort. You can, however, use Fast mode to reduce latency.
Which safety programmes apply to Claude Opus 5.5?
Opus 5.5 carries the same cybersecurity and biology safeguards as Fable 5.1. Qualified organisations can unlock restricted capabilities through the Life Sciences Verification Program and the three-tier Cyber Verification Program, which is being expanded to cover the new model.
The Bottom Line
Claude Opus 5.5 is the rare frontier release where the price goes down, the speed goes up and the benchmarks improve at the same time. It beats Opus 5 everywhere, matches or exceeds Fable 5.1 on most published tests, and undercuts both on cost — while posting Anthropic’s best safety results to date. Enterprises that tested it report fewer tokens, fewer steps and higher recall, which is exactly what makes an AI model cheaper to run in practice, not just on paper. If you are on Opus 5, upgrade now; if you are on Fable 5.1 or GPT-6 Astra, this is the model to benchmark against your own workload this week.
Source: Anthropic — Introducing Claude Opus 5.5. Pricing and availability current as of September 22, 2026; see Anthropic’s developer documentation for the latest model details.
