The short version
- It replaces Opus 5; it is not a cut-down Fable 5.1. Anthropic says it matches Fable 5.1 on most work and is 40% cheaper than Opus 5 on typical workloads.
- The price cut is the headline. Input $5→$4, output $25→$20, cache reads $0.50→$0.20 (a 60% cut). Anthropic says cache reads make up the majority of agentic and coding work costs.
- Thinking can no longer be disabled. Depth is controlled with an
effortparameter (low / medium / high…). The default is medium — the announcement's chart notes say “at default effort (medium)”. - Sonnet 5.5 was released on 2026-09-28 — see our page for it. Haiku 5.5 is still described only as following "in the coming weeks"; as of 2026-10-01 we have not verified its availability.
Pricing (USD per 1M tokens)
Taken from the pricing table in Anthropic's announcement and the official pricing documentation — the same standard, pay-per-token basis we use on our pricing comparison.
| Item | Claude Opus 5.5 | Claude Opus 5 | Fable 5.1 (for reference) |
|---|---|---|---|
| Input | $4.00 | $5.00 | $10.00 |
| Output | $20.00 | $25.00 | $50.00 |
| Cache read (cache hit) | $0.20 | $0.50 | $0.25 |
| Cache write (5-minute TTL) | $5.00 | $6.25 | — |
Note the nuance: on Fable 5.1 a cache hit costs 2.5% of the standard input price ($0.25); on Opus 5.5 it costs 5% ($0.20). Opus 5.5 has the cheaper headline input rate, while Fable 5.1 keeps the deeper cache discount.
Fast mode (Claude Code and the Claude Platform)
| Mode | Input | Output | Speed |
|---|---|---|---|
| Standard | $4.00 | $20.00 | baseline |
| Fast mode | $8.00 | $40.00 | up to 2.5x |
Fast mode pricing “applies across the full context window, including requests over 200k input tokens” per Anthropic's pricing docs.
💡 The 40% claim is about tokens per task as well as price per token. Anthropic writes that Opus 5.5 “requires less compute to serve than Opus 5” and “uses fewer tokens per task”. The saving is the product of both. Don't read the table alone and conclude the older model is cheaper for your workload.
Primary sources: anthropic.com/claude-opus-5-5 (Pricing table) · platform.claude.com pricing docs (verified 2026-09-25)
Figure: how the rates line up
← scroll horizontally →
Specifications
Release date
2026-09-22
Context window
1M tokens
Max output
128K tokens
Knowledge cutoff
June 2026
Thinking mode
Always on (cannot be disabled)
Reasoning effort
low / medium / high… (default medium)
Model ID
claude-opus-5-5
Availability
Claude Platform / AWS / Google Cloud / Azure
Sources: Claude Platform Docs, Models overview (context window 1M, max output 128K, reliable knowledge cutoff Jun 2026, training data cutoff Jun 2026) · Anthropic announcement (availability, thinking behaviour).
Nine official benchmarks — and who actually ran each harness
From the table in Anthropic's announcement. Unless otherwise noted, every Opus 5.5 figure uses adaptive thinking at max effort (verbatim: “Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort”).
| 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% | — | 41.7% |
| GDPval-AA v2.1 (Elo) | 1,846 | 1,735 | 1,708 | 1,542 | 1,588 |
| 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% | — |
| 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% | — | — |
| Chartography (with tools) | 89.0% | 88.4% | 83.4% | — | — |
Who measured what (this is the important part)
| Benchmark | Run and reported by | Harness and conditions |
|---|---|---|
| Terminal-Bench 4.0 | Anthropic's own setup, plus the public leaderboard | Public leaderboard uses the Claude Code harness, 5 trials/task, and reports Opus 5 at 51.8%; Anthropic's setup reproduces it at 52.3% (within noise). Opus 5.5 was run at xhigh effort and GPT-6 Astra at high effort — each model's highest score. Standard error ±2.6 pts. GPT-6 Astra and GPT-5.6 Sol figures are as reported by OpenAI. |
| Terminal-Bench-Science 0.1 | Anthropic's setup, plus the public leaderboard | Public leaderboard uses 3 trials/task and the Claude Code harness, reporting Opus 5 at 30.0% versus 29.0% on Anthropic's setup. Standard error ±3.5–5 pts. GPT-6 Astra figure is as reported by OpenAI. |
| AutomationBench | Zapier ran and reported it | Runs were performed without fallback models, so safeguard interventions counted as failures — a condition that understates the real-world score. Opus 5.5 comes from Zapier's early-access evaluation; Opus 5, GPT-5.6 Sol and GPT-6 Astra come from Zapier's public leaderboard. |
| GDPval-AA v2.1 | Artificial Analysis' evaluation, quoted by Anthropic | Anthropic describes it as “Artificial Analysis's GDPval-AA v2.1 … real-world professional work across 44 occupations”, scored in Elo. |
| CursorBench 4.0 / FrontierCode v1.1 | Anthropic's own setup | Chart notes also give figures “at default effort (medium)” — FrontierCode 54.6%, CursorBench 52.5%. So the table shows max-effort scores while the charts show default-effort scores: different conditions. |
| WANDR (chart only, not in the table) | Perplexity's benchmark, run by Anthropic under modified conditions | Verbatim note: “Claude models were run with offline versions of the web search and web fetch tools, programmatic tool calling, code execution, and a 980k-token task budget. This differs from Perplexity's published setup, scores are not directly comparable”. Only models scored under identical conditions are shown. |
⚠️ How to read the table. ① The GPT-6 Astra column is OpenAI's self-reported figure — Anthropic did not re-run it (“as reported by OpenAI”). ② On Terminal-Bench-Science, Astra scores 64.6% against Opus 5.5's 58.7%. Anthropic does not claim a clean sweep; it says Opus 5.5 “leads in agentic coding, computer use, and knowledge work”. ③ On AutomationBench, Opus 5.5 is second to Astra (41.4%).
🔍 Anthropic itself warns that benchmark margins are getting less meaningful. Verbatim: “at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.” For who runs each of these benchmarks, see Glossary #6: the organisations behind the benchmarks.
Source: Anthropic, “Introducing Claude Opus 5.5” (benchmark table and footnotes 1–4, verified 2026-09-25)
What the efficiency claims mean
| Comparison | What the announcement says |
|---|---|
| Opus 5.5 vs Opus 5 | 40% lower cost on typical workloads; output generated more than 30% faster. |
| Terminal-Bench 4.0 | Opus 5.5 at default effort beats Opus 5 at max effort for about one fifth of the cost, and matches GPT-6 Astra at about 40% of the cost. |
| CursorBench 4.0 | Beats GPT-5.6 Sol's top score by 11 points for about a third of the cost. |
| Internal test: 200,000-line codebase audit | Opus 5.5 finished in under three hours; Opus 5 took over 20 hours and used 2.5x as many tokens. |
| Internal test: HAProxy C→Rust rewrite | Opus 5.5 finished in 9.5 hours versus 12 for Fable 5.1, and cost 51% less. |
All of these are vendor-run internal tests, not independent reproductions.
Safety: safeguards, and a stated limit of the evidence
| Item | What the announcement says |
|---|---|
| External evaluation | Tested before release by Frontier Design and METR. |
| Cybersecurity | Same class of safeguards as Fable 5.1; most cybersecurity tasks are re-routed to Opus 4.8, transparently. |
| Biology | Same biology safeguards as Fable 5.1. Vetted organisations can apply to the Life Sciences Verification Program. |
| Distillation | Launches with preserved thinking, the anti-distillation safeguard introduced with Fable 5.1. Applies to API accounts created on or after 2026-08-31. |
| New containment test | Attempted to circumvent boundaries about 85% less often than Opus 5 or Mythos 5.1; every attempt was low severity and self-reported. |
| Data handling | Available with zero data retention; ships with watermarking to comply with the EU AI Act. |
⚠️ The announcement states a limit plainly. Verbatim: “We see signs that Opus 5.5 often suspects it is being evaluated.” That challenges the ability to predict behaviour in the wide variety of real deployment settings — a vendor acknowledgement that pre-release evaluation has limits. If you run agents unattended, keep your own guardrails.
Source: Anthropic, “Introducing Claude Opus 5.5” (Safety section)
How this relates to our other Claude pages
| Model | On this site | Relationship |
|---|---|---|
| Claude Fable 5.1 | Dedicated page | A different model (higher tier, higher price) at $10 input / $50 output. Opus 5.5's pitch is Fable-5.1-level work at a lower cost. |
| Claude Mythos 5.1 | Dedicated page | A different model (trusted-access tier for vetted organisations). Opus 5.5's cyber and biology safeguards are the same class, and lifting them needs a verification program. |
| Claude Opus 5.5 | This page (new) | First model of the Claude 5.5 family; successor to Opus 5. Sonnet 5.5 was released 2026-09-28 (dedicated page). Haiku 5.5 availability is not verified. |
We have no page for /en/claude-opus-5/ as of 2026-10-01. Sonnet 5.5 is covered at /en/claude-sonnet-5-5/; we do not link to pages that do not exist.
Where it fits — and where to be careful
- Long-running agentic coding: cache reads at $0.20 and Anthropic's own note that cache reads dominate agentic costs make this a good fit for setups that resend repository context.
- Plan for thinking you cannot switch off: easy tasks still consume output tokens. Lower
effortor use a lighter model for simple work. - Cyber and biology work: requests transparently fall back to Opus 4.8 or Opus 5, so a task may be answered by a different model than the benchmark numbers describe.
- If raw peak accuracy is the goal: GPT-6 Astra leads Terminal-Bench-Science at 64.6% versus 58.7%. Check per domain rather than by overall reputation.
- The start of a price war: OpenAI announced GPT-6 Sol and Luna roughly a week later at about half the previous prices — see our GPT-6 Sol / Luna page.
Estimate your real cost
Compare token costs across models, including Claude Opus 5.5.
🧮 Open the token calculator →📊 Full pricing comparison · 🆚 GPT-6 Sol / Luna · 🤖 Fable 5.1 · 🗓️ Release timeline
⚠️ Disclaimer
- Pricing and specifications were verified on Anthropic's announcement and official documentation on 2026-09-25 and may change without notice. Always confirm on the official pages before you commit.
- Benchmark figures include vendor-reported numbers. GPT-6 Astra and GPT-5.6 Sol figures come from OpenAI's own reporting, not from Anthropic re-runs, and conditions may not be identical.
- This site is for information only and is not a recommendation or agent for any provider.