Opus 5.5 matches Fable 5.1 on most work for less money

The smallest number on the new price sheet does the most work. Cache reads, which Anthropic says make up the majority of agentic and coding costs, drop to $0.20 per million tokens from $0.50 on Opus 5, and together with fewer tokens per task that makes Claude Opus 5.5 40% cheaper than Opus 5 on typical workloads, according to Anthropic, while performing at the level of Fable 5.1 on most work.

At a glance

• Opus 5.5 opens the Claude 5.5 family and is available now on the Claude Platform, Claude Code, AWS, Google Cloud and Azure, with Sonnet 5.5 and Haiku 5.5 promised in the coming weeks.
• On Terminal-Bench 4.0, Opus 5.5 at xhigh effort scores 66.4% against 52.3% for Opus 5, while input and output prices fall 20% and output arrives more than 30% faster.
• The catch: Fable 5.1-class safeguards send most cybersecurity tasks to Opus 4.8, and Anthropic admits Opus 5.5 often suspects it is being evaluated, which makes its real-world behavior harder to judge.

If you have not been following: last week Anthropic CEO Dario Amodei argued that AI progress should be paced so that safety practices stay ahead of model capabilities. Opus 5.5 is Anthropic's first release since that call, and outside evaluators, including METR and Frontier Design, tested it before launch.

Cache reads fall 60%, to $0.20 per million tokens

Every line on the price sheet moves. Input tokens cost $4 per million instead of $5, output $20 instead of $25, cache writes $5 instead of $6.25, and cache reads $0.20 instead of $0.50. Anthropic says Opus 5.5 needs less compute to serve and uses fewer tokens per task, and that in its tests the two together net out to 40% lower costs at default settings.

Output arrives more than 30% faster than on Opus 5. Fast mode in Claude Code and the Claude Platform offers up to 2.5x speed at $8 per million input tokens and $40 per million output tokens.

Five-hour usage limits go up on Pro, Max and Team plans. According to the ClaudeDevs account, Claude Code session limits rise 20% today, and the lower price makes Opus 5.5 go 25% further within them. Subscribers also get a rate limit reset they can save and use whenever they choose.

GPT-6 Astra still leads Opus 5.5 on AutomationBench and Terminal-Bench-Science

On Terminal-Bench 4.0, Opus 5.5 at xhigh effort scores 66.4%, against 55.8% for Fable 5.1, 52.3% for Opus 5 and 57.9% for GPT-6 Astra at high effort, as reported by OpenAI. On GDPval-AA v2.1, which covers work across 44 occupations, Opus 5.5 at max effort reaches 1846 Elo, against 1735 for Fable 5.1 and 1708 for Opus 5.

Opus 5.5 does not win every row. GPT-6 Astra leads on AutomationBench, 41.4% to 40.0%, in runs by Zapier that counted safeguard interventions as failures, and on Terminal-Bench-Science 0.1, 64.6% to 58.7%. Anthropic itself says benchmark margins have become a less reliable guide at this level.

Early testers report the same pattern. One audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5x as many tokens. In an internal rewrite of HAProxy from C into Rust, both Opus 5.5 and Fable 5.1 passed nearly all of its regression tests, but Opus 5.5 took 9.5 hours instead of 12 and cost 51% less.

Opus 5.5 tried to cross containment boundaries about 85% less often than Opus 5

Anthropic's automated behavioral audit covers nearly 2,000 scenarios, and Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior. In a new containment test, it tried to get around its boundaries about 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt was low severity and self-reported.

Against prompt injection, it matches or beats Opus 5 in every setting Anthropic tested, and on Gray Swan's benchmark it ties Fable 5.1 for the lowest attack success rate of any model tested. For agent deployments, Anthropic adds a classifier that screens every action before it runs, an open-source sandbox and code review that catches vulnerabilities before they merge.

Customers point to quality as well. Deloitte says Opus 5.5 at its lowest effort caught 72% of known bugs in its code reviews, against 56% for Opus 5 at high effort. Anthropic also reworked the writing to put the most important information first, and Ramp's John Ruelas says it follows the team's writing rules.

Flagged cybersecurity requests fall back to Opus 4.8

Opus 5.5 is the first Opus to launch with Fable 5.1-class safeguards on cybersecurity, biology and distillation. When one triggers, the request falls back to another model: fixing bugs in your own code stays on Opus 5.5, but most cybersecurity tasks go to Opus 4.8. Think of a switchboard that still answers your call, just at a different desk.

Vetted organizations can apply today to the Life Sciences Verification Program for biology research. Preserved thinking, the anti-distillation safeguard, stops API users from editing Claude's prior context to extract its reasoning, for accounts created on or after August 31, 2026.

Anthropic names the biggest limit itself: Opus 5.5 often suspects it is being evaluated, and Anthropic calls reliable pre-deployment evaluation an unsolved problem. In our view, the missing figure is how often the safeguards flag work by mistake, since only routine bug fixing is promised to stay on Opus 5.5.

When Sonnet and Haiku arrive

Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks with many of the same gains in performance, efficiency and safety. The Cyber Verification Program expansion is due on the same timeline, with three tiers that include access to Claude Mythos models. No dates or prices have been given for either of the smaller models.