Anthropic Releases Claude Opus 5.5, Beats GPT-6 Astra On Most Benchmarks

Anthropic released Claude Opus 5.5, which the company says outperforms OpenAI's GPT-6 Astra on most published benchmarks while costing 40% less to run than the outgoing Opus 5. Opus 5.5 scores 66.4% on Terminal-Bench 4.0 versus Astra's 57.9%, 54.4% on FrontierCode v1.1 versus 53.3%, and leads GDPval-AA v2.1 with an Elo of 1846 against 1542, though Astra wins AutomationBench and Terminal-Bench-Science (64.6% versus 58.7%). The model is priced at $4 per million input tokens and $20 per million output tokens, with cache reads down 60% to $0.20 per million tokens and output more than 30% faster than Opus 5.

Anthropic has released Claude Opus 5.5, the first model in what the company is calling its Claude 5.5 family. Anthropic says Opus 5.5 performs at roughly the level of its top-tier Claude Fable 5.1 on most tasks, while costing 40% less to run than the outgoing Opus 5. The release is Anthropic’s first since CEO Dario Amodei’s essay https://officechai.com/ai/anthropic-ceo-dario-amodei-says-ai-development-must-slow-down-to-pace-the-frontier/ calling on the industry to slow the pace of frontier AI development, and the company says Opus 5.5 was tested pre-launch by external evaluators including METR and Frontier Design. On Anthropic’s internal automated behavioural audit, its most extensive alignment test suite, the company says Opus 5.5 is the best-scoring model it has released. Claude Opus 5.5 Benchmarks On the benchmark charts Anthropic has published, Opus 5.5 comes out ahead of OpenAI’s GPT-6 Astra on most categories. It scores 66.4% on Terminal-Bench 4.0 against Astra’s 57.9%, 54.4% on FrontierCode v1.1 against Astra’s 53.3%, and leads on GDPval-AA v2.1 with an Elo of 1846 versus Astra’s 1542. It’s also ahead on Humanity’s Last Exam 67.7% with tools versus Astra’s 57.2% and on OSWorld 2.0 computer-use testing. Astra does come out on top on two of the categories Anthropic tested: AutomationBench, a Zapier-run business workflows benchmark, and Terminal-Bench-Science, a test of agentic scientific research where Astra scores 64.6% against Opus 5.5’s 58.7%. Anthropic has flagged that its own safeguards intervened during several of these evaluations, routing some cybersecurity and biology-adjacent tasks to older, weaker Claude models instead — which likely dragged its scores down on those specific tests. This continues a broader pattern of the gap between AI labs on the Artificial Analysis Intelligence Index https://officechai.com/ai/google-slips-to-10th-place-among-ai-labs-on-the-artificial-analysis-intelligence-index/ shifting from one release to the next. Claude Opus 5.5 Pricing and Speed Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, down from Opus 5’s $5 and $25. Cache reads — which Anthropic says account for most agentic and coding costs — drop 60% to $0.20 per million tokens, and cache writes fall to $5 from $6.25. Anthropic says the model also generates output more than 30% faster than Opus 5, and a “fast mode” running at up to 2.5x speed is available in Claude Code and the Claude Platform for $8 per million input tokens and $40 per million output tokens. Anthropic is also raising five-hour usage limits on its Pro, Max and Team subscription plans, and giving subscribers a saveable rate-limit reset they can use whenever they choose. Coding and knowledge work Anthropic is pitching Opus 5.5 heavily on long, sprawling engineering jobs. It cites one tester who used the model to audit and fix a 200,000-line codebase in under three hours, versus over 20 hours for Opus 5, and another who ran a 680,000-line code migration in under a day. In an internal test rewriting the load-balancing software HAProxy from C into Rust, Opus 5.5 finished in 9.5 hours against Fable 5.1’s 12, at just under half the cost. GitHub’s chief product officer Mario Rodriguez said the model used among the fewest tokens and steps of any model his team tested, solving more terminal tasks in VS Code in under half the steps Opus 5 needed. Deloitte Consulting’s CIO Carl Bennett said Opus 5.5 caught more known bugs in code review at its lowest effort setting than Opus 5 did at its highest. Safety and safeguards Because Opus 5.5’s cybersecurity and biology capabilities are close to those of Claude Mythos 5.1, Anthropic is deploying it with safeguards similar to those on Fable 5.1: most cybersecurity tasks get quietly rerouted to the older Opus 4.8, and deeper biology work requires vetted access through Anthropic’s Life Sciences Verification Program. This is part of a wider industry reckoning with what capable models can do unsupervised — Google recently disclosed that its own Gemini model hacked into companies during security testing https://officechai.com/ai/googles-gemini-too-hacked-into-companies-in-security-tests-google-says-it-stopped-when-it-realized-it-wasnt-in-a-simulation/ , a category of behaviour Anthropic says Opus 5.5 is now substantially less prone to than Opus 5 or Mythos 5.1. Anthropic also says Opus 5.5 often appears to suspect it’s being evaluated, which the company admits complicates its ability to predict how the model will behave once deployed in the wild. The model ships with “preserved thinking,” an anti-distillation measure introduced with Fable 5.1 that blocks API users from editing Claude’s prior reasoning to extract its capabilities — relevant given how much scrutiny frontier labs are under after a string of AI-assisted hacking incidents https://officechai.com/ai/hackers-were-able-to-use-claude-to-hack-openai-employee-accounts-and-create-a-pr-in-its-repo/ this year. It’s available with zero data retention and complies with EU AI Act watermarking rules. Claude Opus 5.5 Availability Claude Opus 5.5 is live today on the Claude Platform, AWS, Google Cloud and Microsoft Azure, with developers able to call it as claude-opus-5-5. Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to follow in the coming weeks.