Claude Opus 5.5 Launches: Stronger, 40% Cheaper

Anthropic shipped Claude Opus 5.5, the first model in its new Claude 5.5 family. The headline isn't a benchmark score — it's a 40% drop in running cost compared to Opus 5, while it's simultaneously the strongest model Anthropic has released.

What happened

Anthropic's official blog, posted September 22, frames Opus 5.5 as its first release since CEO Dario Amodei called for "pacing the frontier" — slowing capability growth so safety practices can keep up. Before release, the model went through evaluation by outside organizations including METR and Frontier Design.

On pricing, input tokens drop to $4 per million (down 20% from Opus 5's $5), output tokens to $20 per million (from $25). The bigger number is cache reads — which make up most of the cost in agentic and coding workloads — down 60% to $0.20 per million tokens, from $0.50. Anthropic is also raising five-hour usage limits across Pro, Max, Team, and seat-based Enterprise plans.

Performance: a real jump over Opus 5

On Terminal-Bench 4.0, which measures multi-step professional task completion in a command-line environment, Opus 5.5 scores 66.4%, against 52.3% for Opus 5 and 57.9% for GPT-6 Astra (as reported by OpenAI). On GDPval-AA v2.1, which evaluates real-world work across 44 occupations, Opus 5.5 hits 1846 Elo, ahead of Opus 5 (1708) and Claude Fable 5.1 (1735).

One internal tester used Opus 5.5 to migrate a 680,000-line codebase in under a day. That's work that would normally take an engineering team weeks. In another test, asked to cut page load times across a web app, Opus 5.5 succeeded on 39 of 40 attempts; Opus 5 made smaller gains and sometimes altered app behavior unintentionally.

Worth noting: these are Anthropic's own figures, reported with specific test conditions (effort level, harness). Some benchmarks, like AutomationBench, were run independently by Zapier and reported with their own methodology.

Safety: the highest behavioral audit score yet

On its automated behavioral audit — the most comprehensive alignment suite Anthropic runs, testing Claude across nearly 2,000 simulated scenarios — Opus 5.5 scores better than any prior Claude model on almost every measure. It's less likely to take hard-to-reverse actions or act outside assigned boundaries, and it's more resistant to prompt injection than Opus 5.

Because its biology and cybersecurity capabilities are comparable to Claude Mythos 5.1, Anthropic is deploying Opus 5.5 with safeguards similar to Claude Fable 5.1's — including access through its Life Sciences Verification Program for biology research, with expanded Cyber Verification Program access for verified security practitioners coming soon.

Coding agents: open sandbox, action-level classifier

For enterprises running agents autonomously for hours at a time, Anthropic added three layers: a classifier that screens every action before execution, an open-source sandbox security teams can audit, and automated code review that catches vulnerabilities before merge.

On a prompt injection benchmark run independently by AI security firm Gray Swan, Opus 5.5 ties Claude Fable 5.1 for the lowest attack success rate of any model tested.

So what — what this means for developers

• Agentic coding gets meaningfully cheaper: a 60% drop in cache read pricing matters most for workflows that repeatedly reuse large context — code review, large refactors — cutting cost, not just latency
• Long, complex tasks become economically viable: migrating a codebase with hundreds of thousands of lines in hours instead of weeks changes the math on refactor projects many teams previously shelved as too expensive
• Prompt injection resistance matters more as agents run unattended: for agents with real system access — reading email, calling APIs, running commands — resistance to being steered by malicious content in their input isn't a nice-to-have anymore
• Sonnet 5.5 and Haiku 5.5 are coming: if you're on Sonnet for cost reasons, it's worth waiting a few weeks rather than migrating straight to Opus 5.5

What makes this release different isn't the benchmark numbers — every AI lab claims "strongest model yet" every few months. The notable part is Anthropic cutting price by 40% instead of holding the line and selling capability as a premium tier. Inference cost is still the biggest blocker to running agents at production scale. That's a signal the competition is shifting from "who's smartest" to "who's cheapest per unit of completed work."

Key takeaways

• Claude Opus 5.5 launched September 22, priced 40% below Opus 5 at default settings; cache reads down 60% to $0.20/million tokens
• Scores 66.4% on Terminal-Bench 4.0 and 1846 Elo on GDPval-AA v2.1, the highest Anthropic has published
• Highest score yet on its automated behavioral audit, a nearly 2,000-scenario alignment suite
• Ties Claude Fable 5.1 on prompt injection resistance in Gray Swan's independent benchmark
• Open-source sandbox, action-screening classifier, and automated code review for long-running agents
• Claude Sonnet 5.5 and Claude Haiku 5.5 are expected in the coming weeks

---

Content assisted by AI (Amy 🌸). Reviewed by the author.