Claude Opus 5.5

Anthropic's Opus-tier model after Opus 5, announced alongside OpenAI's GPT-6 Sol and Luna on 22 September 2026 (opus-and-sol-and-luna-willison). It is the first model of a Claude 5.5 family; Sonnet 5.5 and Haiku 5.5 are due "in the coming weeks" (claude-opus-5-5-announcement).

Price (Anthropic's pricing docs, read 2026-09-23 — anthropic-model-docs)

• $4 input / $20 output per million tokens, against Opus 5's $5/$25: a 20% cut.
• Cache hits cost $0.20, 0.05× input. Other Opus models pay 0.1× and Fable 5.1 pays 0.025×, so Anthropic now sets the cache discount per model.
• Batch $2/$10. Fast mode $8/$40, still double the standard rate, as it is on Opus 5.
• Output limit 128,000 tokens (Willison).

What Anthropic claims (launch post, 2026-09-22 — claude-opus-5-5-announcement)

All first-party. Read the source page for the footnotes.

• Close to Fable 5.1 on most work, and 40% cheaper than Opus 5 on typical workloads: the 20% per-token cut plus fewer tokens per task. Output is more than 30% faster.
• Leads its own table on Terminal-Bench 4.0 (66.4%), FrontierCode (54.4%), CursorBench (57.8%), GDPval-AA (1846 Elo), Humanity's Last Exam and OSWorld 2.0. GPT-6 Astra leads on AutomationBench (41.4% vs 40.0%) and Terminal-Bench-Science (64.6% vs 58.7%).
• Anthropic says its own margins overstate the gap to Fable 5.1 in real use.
• Medium effort is the default, and the post's main argument is cost per task at medium: on FrontierCode, medium (54.6%) scores slightly above max (54.4%).
• Safeguards that fall back to other models, a first for an Opus: most cyber tasks go to Opus 4.8; biology gets Fable 5.1's safeguards. Preserved thinking blocks reasoning extraction for API accounts created on or after 2026-08-31. Thinking cannot be turned off.
• On AWS, Google Cloud and Azure from launch. Cache writes $5 (Opus 5: $6.25).

One hands-on report

Willison's: at maximum thinking it spent its whole output budget reasoning on a small drawing task, failed, and cost about $2.56 per try (opus-and-sol-and-luna-willison). That is one task, and he says so. It is the only test here not run by Anthropic or its early-access customers. The launch post says nothing about the context window or the output limit.

Where it sits

At $4/$20 it matches GPT-5.6 Sol's promotional price exactly and costs twice GPT-6 Sol ($2/$10). See llm-api-pricing for the September repricing.

Related

claude-opus-5 · claude-fable-5-1 · claude-sonnet-5 · anthropic · gpt-6 · opus-and-sol-and-luna-willison · claude-opus-5-5-announcement