Claude Opus 5.5

Claude Opus 5.5 comes three weeks after Fable 5.1 and two months after Opus 5. Anthropic says: Fable 5.1 performance on most work, 40% cheaper than Opus 5, and the first Claude that Anthropic says it trained to stop talking so much.

What shipped

Model type: Text + image input, text output

Pricing: $4 / $20 per million input / output tokens, 20% below Opus 5’s $5 / $25 and 40% of Fable 5.1’s $10 / $50. Cache reads are $0.20 per million, 5% of base input where Opus 5 charged 10% ($0.50), so 60% cheaper on the line item that dominates agent bills. Cache writes $5 (five-minute) and $8 (one-hour), down from $6.25 and $10. Batch is $2 / $10. Fast modeis $8 / $40 for up to 2.5x output speed, Claude API only, research preview.

Available on: claude.ai and the Claude apps, Claude Code, Claude Cowork, the Claude Developer Platform, Amazon Bedrock and Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The less-restricted cyber build comes “in the near future” through the Cyber Verification Program, which is being expanded to three tiers. Life-science work that trips the new bio classifier goes through the Life Sciences Verification Program that opened September 17.

Other info: 1M-token context window, 128K max output (300K on the Batch API with the extended-output beta). Reliable knowledge cutoff and training data cutoff both June 2026, matching Fable 5.1. Same tokenizer as everything since Opus 4.7. 512-token minimum cacheable prompt. Under the Responsible Scaling Policy it’s CB-1 but not CB-2 for chem and bio, and it ships with the same expanded biology safeguards as Fable 5 and 5.1, which is a first for the Opus line. Cyber is the headline safety fact: the 230-page system card calls it “the strongest cyber capabilities of any model we have released, meeting or exceeding Claude Mythos 5.1 on all of our internal evaluations,” still inside the lower tier of the Frontier Compliance Framework.

More details: Claude Opus 5.5, the system card, and What’s new in Claude Opus 5.5 for developers.

Anthropic released Claude Opus 5.5 on Tuesday morning as “the first model in our new Claude 5.5 family,” and the framing sentence is the spec sheet: “It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” The positioning is the Opus 5 playbook run again, one tier down in price. Opus 5 in July was pitched as near-Fable performance at half of Fable’s price; Opus 5.5 is pitched as Fable 5.1 performance at 40% of Fable 5.1’s list price and with cache reads that cost less than Fable 5.1’s already-discounted rate. It’s live on every Anthropic surface and all four clouds on day one, subscribers on Pro, Max, and Team got their five-hour limits raised alongside it, and Sonnet 5.5 and Haiku 5.5 are promised within weeks. Under the hood the API contract is Fable 5.1’s: thinking can’t be turned off, forced tool use is gone, thinking blocks are bound to the conversation, and the default effort drops from high to medium.

The evidence is the most complete Anthropic has put on a launch page, and this time I’ll give them credit for shipping absolute numbers next to the competition’s. Terminal-Bench 4.0 jumps 14 points over Opus 5 and clears GPT-6 Astra by eight. GDPval and CursorBench set new highs. The launch-partner section reads like an expense report: GitHub says it “used the fewest tokens and steps we measured,” Kiro reports “40% fewer calls and half the tokens,” Box says a third of Opus 5’s tokens and “40% less verbose without losing accuracy,” Factory runs it at medium and matches Opus 5 at high with 20 to 25% fewer tokens. Anthropic’s own showcase is a HAProxy port from C to Rust in 9.5 hours for 51% less than Fable 5.1’s 12-hour run, and a 200,000-line code audit in under three hours where Opus 5 took more than twenty. The catches are in the same materials. Astra still leads on AutomationBench and Terminal-Bench-Science, and the announcement itself concedes that “benchmark margins have become a less reliable guide to real-world differences” and that “in our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.” The 40% cost claim is measured “at default settings,” and the default moved from high to medium in the same release, while the docs warn the model “tends to think more per turn than Claude Opus 5” at any given level.

Also, a bit of an uncomfortable line in the system card: “We see signs that Opus 5.5 often suspects it is being evaluated.”

What’s new

• Fable-class work at 40% of Fable’s price. Cache reads at $0.20. An agent loop’s input is mostly its own transcript, Fable 5.1 cut that line to $0.25 three weeks ago, and now Opus 5.5 undercuts it on a model with $4 base input instead of $10. Anthropic’s 40% figure is against Opus 5. Against Fable 5.1 the base prices are 60% lower and the cache line is 20% lower.

• Thinking has no off switch, and the default is medium. thinking: {"type": "disabled"} returns a 400 at every effort level, budget_tokens is gone, and effort is the only dial. A request that omits effort now runs at medium where Opus 5 ran at high, and Anthropic’s own testing says medium on 5.5 “matches or exceeds Claude Opus 5 at high on coding and knowledge-work evaluations.” Most of the 40% lives there.

• It writes shorter, on purpose. The announcement has a section titled “Communication,” a response to Opus 5’s verbosity complaints. “Clearer and easier to follow,” “puts the most important information up front,” “less likely to use jargon or idiosyncratic phrases.” Ramp’s engineer says it “fixes verbose output frustration” and “writes like a good colleague.”

• Progress notes move into thinking blocks. The short text the model writes between tool calls now comes back as a thinking block instead of a text block, and at the default display: "omitted"that text is empty. No request fails, your agent UI just goes silent mid-task until you set display: "updates" (beta).

• It reads the chart without a tool. Anthropic claims that “even at its lowest effort setting it read values off dense charts more accurately than Claude Opus 5 did at its highest.” Chartography lands at 89.0% with tools. Flowchart arrows, diagram diffs, and calendar screenshots are called out by name.

The benchmarks

Anthropic’s numbers

• Terminal-Bench 4.0 (xhigh): 66.4%. +10.6 over Fable 5.1 (55.8%), +14.1 over Opus 5 (52.3%), +8.5 over GPT-6 Astra (57.9%).

• FrontierCode v1.1: 54.4%. +4.1 over Fable 5.1 (50.3%), +6.4 over Opus 5 (48.0%), +1.1 over Astra (53.3%).

• CursorBench 4.0: 57.8%. +6.0 over Fable 5.1 (51.8%), +11.2 over Opus 5 (46.6%).

• GDPval-AA v2.1 (knowledge work): 1846. +111 over Fable 5.1 (1735), +138 over Opus 5 (1708), +304 over Astra (1542).

• Humanity’s Last Exam with tools: 67.7%. +2.1 over Fable 5.1 (65.6%), +4.1 over Opus 5 (63.6%), +10.5 over Astra (57.2%).

• Humanity’s Last Exam without tools: 64.4%. +3.5 over Fable 5.1 (60.9%).

• OSWorld 2.0: 81.8% partial, 48.7% strict

  • +1.1 partial, +5.9 strict (Fable 5.1)

• AutomationBench (Zapier): 40.0%. −1.4 under Astra (41.4%), +8.6 over Fable 5.1 (31.4%).

• Terminal-Bench-Science: 58.7%. −5.9 under Astra (64.6%), +6.1 over Fable 5.1 (52.6%), +29.7 over Opus 5 (29.0%).

How and where to use it

Where it’s available

• Claude.ai and the Claude apps on Pro, Max, Team, and Enterprise

• Claude Code and Claude Cowork

• Claude Developer Platform plus Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Fast mode is Claude API only.