Introducing Claude Opus 5.5

What it claims

Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, positioned as performing at roughly the level of Claude Fable 5.1 on most work while costing ~40% less to run than Opus 5. The headline pitch to engineers is efficiency: same-or-better results using fewer tokens and fewer steps per task, especially on long agentic coding jobs (codebase-wide migrations, audits). Model ID is claude-opus-5-5.

Evidence

Pricing (the concrete, actionable part):

• Input $4/M, output $20/M — 20% below Opus 5.
• Cache reads $0.20/M — 60% below Opus 5 (page notes cache reads dominate agentic/coding cost).
• Net ~40% cost drop on typical workloads; output generation >30% faster.
• Fast mode (Claude Code + Platform, up to 2.5x speed): $8/M input, $40/M output.

Benchmarks, with baselines and cost framing as stated:

• FrontierCode (default effort): beats GPT-6 Astra at ~20% of cost per task.
• Terminal Bench 4.0: matches Astra at ~40% of cost.
• CursorBench: beats GPT-5.6 Sol by 11 points at ~1/3 cost.
• GDPval-AA v2.1 (44 occupations): 1846 Elo, ahead of Fable 5.1 and Opus 5; default-effort beats GPT-6 Astra at max effort for ~1/5 the cost.
• Internal HAProxy C→Rust rewrite: finished in 9.5h vs Fable 5.1's 12h, 51% cheaper, both passing nearly all regression tests.

Anthropic explicitly cautions that at these capability levels benchmark margins are a weak guide, and that the real Opus 5.5 vs Fable 5.1 gap is narrower than scores suggest.

How to use it

Available now on AWS, Google Cloud, Azure, and the Claude Platform. Increased 5-hour usage limits on Pro/Max/Team/Enterprise plans.

Limits

Several restrictions materially affect adoption and are not obvious from the summary:

• Cybersecurity tasks re-routed to Opus 4.8. You can identify/fix bugs in normal SDLC, but most cyber tasks silently fall back to a weaker model unless you're in the (soon-expanding) Cyber Verification Program.
• Biology safeguards match Fable 5.1; impeded R&D work requires the Life Sciences Verification Program.
• Preserved thinking (anti-distillation) blocks API users from editing Claude's prior context, applied to API accounts created on or after 2026-08-31 — this can break integrations that manipulate reasoning context; test before relying on it.
• Thinking mode can no longer be switched off.
• EU AI Act watermarking is applied; zero data retention is available.
• Anthropic states Opus 5.5 often suspects it's being evaluated, which limits confidence that alignment test results transfer to real deployments.