Claude Opus 5.5 is Anthropic’s September 22, 2026 flagship and the first model of the new Claude 5.5 family. Its pitch is blunt: roughly Claude Fable 5.1-level work at less than half Fable’s price, and 40% cheaper to run than Claude Opus 5 on typical workloads. Input and output tokens drop to $4 / $20 per million, cache reads to $0.20, and output is more than 30% faster.
It is also Anthropic’s first release since it called for “pacing the frontier”: external evaluators including Frontier Design and METR tested it before launch, and it posts the strongest score the company has recorded on its automated behavioral audit.
Benchmarks
Numbers below are vendor-reported (Anthropic, September 22, 2026). Opus 5.5 results use adaptive thinking at max effort unless noted; Terminal-Bench 4.0 is Opus 5.5 at xhigh effort and GPT-6 Astra at high effort as reported by OpenAI.
| Benchmark | What it measures | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Agentic terminal coding | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 (Main) | Mergeable agentic coding | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | Real Cursor coding sessions | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 | Real-world knowledge work (Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench | Cross-app business workflows | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity’s Last Exam | Multidisciplinary reasoning (tools) | 67.7% | 65.6% | 63.6% | 57.2% | — |
| Terminal-Bench-Science 0.1 | Agentic scientific research | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.1 | Computer use (partial credit) | 81.8% | 80.7% | 74.0% | — | — |
| Chartography | Visual chart recognition (tools) | 89.0% | 88.4% | 83.4% | — | — |
Anthropic itself warns that at these levels “benchmark margins have become a less reliable guide to real-world differences”, and that the gap between Opus 5.5 and Fable 5.1 is narrower in its own use than the scores suggest. The more durable story is efficiency:
- At default effort, Anthropic says Opus 5.5 beats GPT-6 Astra on FrontierCode at roughly 20% of the cost per task, and matches Astra on Terminal-Bench 4.0 for about 40% of the cost.
- On Terminal-Bench 4.0, Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost.
- In an internal C-to-Rust translation of HAProxy, both Opus 5.5 and Fable 5.1 passed nearly all of HAProxy’s regression tests, but Opus 5.5 finished in 9.5 hours versus 12 and cost 51% less.
- An early tester audited and fixed a 200,000-line codebase in under three hours; Opus 5 took over 20 hours and 2.5x the tokens.
Pricing
| Prices per 1M tokens | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes | $5 | $6.25 |
Fast mode (up to 2.5x speed) is available in Claude Code and the Claude Platform at $8 input / $40 output per million tokens. On the apps, Anthropic also raised five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and gave subscribers a saveable rate-limit reset.
Specs
| Spec | Value |
|---|---|
| API model ID | claude-opus-5-5 |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Thinking | Adaptive, always on (cannot be switched off) |
| Default effort | Medium |
| Reliable knowledge cutoff | June 2026 |
| Modalities | Text and image in, text out |
| Platforms | Claude Platform, Amazon Web Services, Google Cloud, Microsoft Azure |
Safety and safeguards
- Alignment. Best scores to date on Anthropic’s automated behavioral audit across thousands of simulated scenarios, with broader coverage of long tasks, impossible tasks and scenarios modeled on real incidents. It is less likely than recent models to take hard-to-reverse actions or act outside its boundaries.
- Prompt injection. Matches or beats Opus 5 in every setting tested (coding, tool use, computer use, browsing) and ties Fable 5.1 for the lowest injection success rate on Gray Swan’s benchmark.
- Agent security. A classifier screens every action before it runs, an open-source sandbox is auditable by security teams, and code review is meant to catch vulnerabilities before they merge.
- Dangerous capabilities. Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to Fable 5.1’s. Vetted organizations can apply to the Life Sciences Verification Program for biology work, and the Cyber Verification Program was expanded on October 6.
- Data and compliance. Zero data retention is available, outputs carry watermarking for EU AI Act compliance, and preserved thinking applies to API accounts created on or after August 31, 2026.
How Opus 5.5 compares
| Claude Opus 5.5 | Claude Fable 5.1 | Claude Opus 5 | GPT-6 Astra | |
|---|---|---|---|---|
| Price (in / out, per 1M) | $4 / $20 | $10 / $50 | $5 / $25 | $10 / $50 |
| Context | 1M | 1M | 1M | 1.05M |
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% |
| GDPval-AA v2.1 | 1846 | 1735 | 1708 | 1542 |
| Released | Sep 22, 2026 | Sep 1, 2026 | Jul 24, 2026 | Sep 3, 2026 |
- vs Claude Fable 5.1: Opus 5.5 is the default for most workloads; reach for Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short.
- vs Claude Opus 5: a straight upgrade — cheaper per token, cheaper per task, faster, and clearer writing, which was a common complaint about Opus 5.
- vs GPT-6: Astra leads on scientific-research and business-workflow benchmarks at 2.5x the price; GPT-6.1 Sol ($2 / $10) is the closer price competitor.
- Within the family: Claude Sonnet 5.5 ($2 / $10) and Claude Haiku 5.5 (from $0.10 / $0.50) take the well-scoped and high-volume work.
Limitations
- Safeguard fallbacks. Anthropic notes that when production safeguards intervened, cybersecurity tasks were completed by Opus 4.8 and biology and frontier-LLM-development tasks by Opus 5, which likely lowers its benchmark scores and can change behavior on security-adjacent work.
- Thinking is mandatory. There is no thinking-off mode, so very latency-sensitive paths should look at Haiku 5.5 instead.
- Vendor numbers. Every score above comes from Anthropic or from rivals’ own announcements; wait for independent replications before treating small gaps as real.
For the wider picture, see the 2026 LLM guide.
Best for
- Long, sprawling engineering jobs — codebase-wide migrations, audits and performance work
- Agentic computer use and browser workflows where per-task cost matters
- Knowledge work — financial analysis, spreadsheets, decks and research reports
- Default driver for Claude Code and Cowork, with Sonnet 5.5 / Haiku 5.5 as cheaper subagents
Pros & cons
Strengths
- Anthropic says it performs at Fable 5.1's level on most work while costing 40% less to run than Opus 5 on typical workloads
- Leads Anthropic's own table on Terminal-Bench 4.0 (66.4%), CursorBench 4.0 (57.8%), GDPval-AA v2.1 (1846) and OSWorld 2.1 (81.8%)
- Price cut across the board vs Opus 5 — $4 / $20 per 1M, cache reads $0.20 (60% lower) — and output is 30%+ faster
- Strong prompt-injection resistance (ties Fable 5.1 for the lowest rate on Gray Swan's benchmark) plus a screening classifier, an auditable open-source sandbox and code review for agent use
Limitations
- Tops the table on most rows but trails GPT-6 Astra on Terminal-Bench-Science 0.1 (58.7% vs 64.6%) and AutomationBench (40.0% vs 41.4%)
- Safeguards sit on top of the model — fallbacks to older models on cyber, biology and frontier-LLM-development tasks can lower scores and change behavior
- Thinking can no longer be switched off, and preserved thinking blocks API users from editing prior context (accounts created on or after Aug 31, 2026)
- Anthropic itself cautions that at this capability level benchmark margins are a less reliable guide to real-world differences
Frequently asked questions
When was Claude Opus 5.5 released?
September 22, 2026. It is the first model in Anthropic's Claude 5.5 family; Claude Sonnet 5.5 followed on September 28 and Claude Haiku 5.5 on October 7.
How much does Claude Opus 5.5 cost?
$4 per 1M input tokens and $20 per 1M output tokens — 20% below Opus 5's $5 / $25. Cache reads are $0.20 per 1M (Opus 5 was $0.50) and cache writes $5. Anthropic estimates a 40% lower cost than Opus 5 on typical workloads because it also uses fewer tokens per task. Fast mode costs $8 / $40 per 1M.
Is Opus 5.5 better than Claude Fable 5.1?
On Anthropic's published table Opus 5.5 scores higher than Fable 5.1 on every row shown (for example Terminal-Bench 4.0 66.4% vs 55.8%), at under half the price ($4 / $20 vs $10 / $50). Anthropic nevertheless says the real-world gap is narrower than those scores suggest, and Fable 5.1 remains the choice when evals on Opus 5.5 at higher effort still fall short.
What is the Claude API model ID for Opus 5.5?
`claude-opus-5-5`. It has a 1M-token context window and up to 128K output tokens, adaptive thinking that is always on, and a default effort of medium. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
Does Opus 5.5 beat GPT-6 Astra?
It depends on the task. In Anthropic's table Opus 5.5 leads GPT-6 Astra on Terminal-Bench 4.0 (66.4% at xhigh vs 57.9% at high), FrontierCode v1.1 (54.4% vs 53.3%) and GDPval-AA v2.1, while Astra leads on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science 0.1 (64.6% vs 58.7%). Astra costs $10 / $50, 2.5x Opus 5.5.
Why can't I turn thinking off on Opus 5.5?
Anthropic made adaptive thinking always-on for Fable 5.1 and Opus 5.5 and added preserved thinking, an anti-distillation safeguard that stops API users from editing Claude's prior context to extract its reasoning. It applies to API accounts created on or after August 31, 2026.