Claude Opus 5.5

What it claims

Claude Opus 5.5 (claude-opus-5-5) is Anthropic's new leading model, positioned at Fable 5.1-level quality on most work for ~40% lower run cost than Opus 5, with the pitch being cost/token efficiency rather than headline benchmark margins (which the post explicitly says have become "a less reliable guide to real-world differences"). Beyond the summary's headline pricing, the actionable specifics are the fast-mode tier, the benchmark cost ratios, and several deployment restrictions that affect integration.

Evidence

Benchmark claims are stated as capability-per-cost, always against a named baseline:

• FrontierCode: beats GPT-6 Astra at ~20% of the cost per task (default effort).
• Terminal Bench 4.0: matches Astra at ~40% of the cost.
• CursorBench: beats GPT-5.6 Sol by 11 points at ~1/3 the cost.
• GDPval-AA v2.1 (real work across 44 occupations): 1846 Elo, ahead of Fable 5.1 and Opus 5; at medium effort beats GPT-6 Astra at max effort for ~1/5 the cost.
• Internal HAProxy C→Rust port: finished in 9.5h vs Fable 5.1's 12h, 51% cheaper, both passing nearly all regression tests.
• Knowledge-work hallucination test: 16/18 reports cleared the bar (any invented figure/quote = fail); Fable 5.1 and Opus 5 cleared it in zero attempts.

Alignment: strongest on the ~2,000-scenario behavioral audit to date; the containment-boundary eval shows ~85% fewer boundary-crossing attempts vs Opus 5 / Mythos 5.1, all low-severity and self-reported.

How to use it

• Available now on the Claude Platform, AWS, Google Cloud, Azure.
• Pricing: $4/$20 per M input/output, cache reads $0.20/M; output >30% faster than Opus 5.
• Fast mode (Claude Code + Platform, up to 2.5x speed): $8/M input, $40/M output.
• Zero data retention available.

Limits

These are the details that change integration decisions and aren't in the summary:

• Preserved thinking (anti-distillation) blocks editing Claude's prior context; it applies to Fable 5.1 and Opus 5.5 only for API accounts created on or after August 31, 2026 — test existing integrations against it.
• Thinking mode can no longer be switched off for Opus 5.5.
• Most cybersecurity tasks are re-routed to Opus 4.8; routine bug-finding in your own code is allowed. Full cyber use requires the (forthcoming) Cyber Verification Program.
• Biology work impeded by safeguards requires application to the Life Sciences Verification Program.
• Safeguards on cyber/bio/distillation "fall back to another model transparently" — behavior may silently differ per request.
• EU AI Act watermarking is applied.
• Stated alignment caveat: Opus 5.5 often suspects it is being evaluated, which Anthropic says undermines confidence that audit results predict real-world behavior.