Claude Opus 5.5 Launches With Major AI Upgrade for Coding, Research, and Agentic Work

Anthropic released Claude Opus 5.5 on September 22, 2026, describing it as the first model in a new Claude 5.5 generation and its new leading model for coding, research and long-running AI agent work. The company says the model performs at roughly the level of Claude Fable 5.1 on most tasks while costing 40 percent less to run than its predecessor, Opus 5, on typical workloads.

Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens, a 20 percent reduction from Opus 5's rates. Cache reads, which Anthropic says make up the majority of agentic and coding costs, fall 60 percent to $0.20 per million tokens. A faster mode is also available in Claude Code and the Claude Platform at up to 2.5 times the speed, priced at $8 and $40 per million input and output tokens. The model is available immediately through the Claude API, Claude Platform, and supported cloud services including Amazon Web Services, Google Cloud and Microsoft Azure.

Coding and Long-Running Agentic Work

Anthropic points to several internal examples to illustrate the coding gains. One early tester reportedly completed a 680,000-line code migration in under a day. In another internal test, Opus 5.5 audited and fixed a 200,000-line codebase in under three hours, compared with more than 20 hours for Opus 5 using two and a half times as many tokens. Asked to port HAProxy, widely used load-balancing software, from C to Rust, Opus 5.5 finished in 9.5 hours versus 12 hours for Fable 5.1, at 51 percent lower cost, with both versions passing nearly all of HAProxy's own regression tests.

Enterprise customers cited by Anthropic described similar results on extended, multi-step work. Stripe said a single Opus 5.5 session directed a dozen additional sessions overnight across a multi-day rebase of 40 stacked pull requests, with all 40 passing continuous-integration tests the next afternoon. Chicago Trading Company said the model autonomously diagnosed and fixed a bug in its systems overnight, passing its test suite by morning. Anthropic also says Opus 5.5 includes a classifier that screens agent actions before they run, an auditable open-source sandbox, and code review meant to catch vulnerabilities before code merges.

Benchmark Claims, With Caveats

On benchmarks Anthropic has published, Opus 5.5 leads Fable 5.1, Opus 5, and OpenAI's GPT-6 Astra and GPT-5.6 Sol on most measures shown, including Terminal-Bench 4.0, a test of complex command-line tasks, where it scored 66.4 percent against 55.8 percent for Fable 5.1 and 52.3 percent for Opus 5. On GDPval-AA v2.1, a benchmark from Artificial Analysis covering real-world professional work across 44 occupations, Opus 5.5 scored 1846 Elo versus 1735 for Fable 5.1.

Anthropic itself cautions against reading too much into these margins. The company states that at current capability levels, benchmark scores have become a less reliable guide to real-world differences, and that in its own use, the practical gap between Opus 5.5 and Fable 5.1 is narrower than the published scores suggest. Anthropic also notes that its safeguards, described below, intervened during some benchmark runs and likely reduced Opus 5.5's measured scores in cybersecurity and biology-related tests, since those tasks were rerouted to less capable models. None of the benchmark results have been independently reproduced by a third party as of this writing.

Research and Knowledge Work

In an internal evaluation, Anthropic asked Opus 5.5, Fable 5.1 and Opus 5 to write a company earnings report using only information available on a deliberately obscured copy of the web, with an automated grader checking every figure and quote against sources. Sixteen of eighteen Opus 5.5 attempts passed, where any fabricated figure or quote counted as a failure; neither Fable 5.1 nor Opus 5 passed on any attempt. Investment firm Walleye Capital, an early tester, said the model largely solved its evaluation suite at its lowest effort setting and, at higher settings, identified and corrected an error in the firm's own evaluation instructions that no prior model had caught.

Speed, Cost and Communication Style

Anthropic says Opus 5.5 generates output more than 30 percent faster than Opus 5 and requires less compute to serve, which the company says accounts for the bulk of the 40 percent typical cost reduction, alongside using fewer tokens per completed task. The company also highlights changes to how the model communicates, saying it now places the most important information earlier in a response and uses less jargon. "It writes the way I do," one early tester told Anthropic. Several enterprise customers, including Box and Ramp, said Opus 5.5's answers were substantially less verbose than Opus 5's without a loss in accuracy.

Safety, Safeguards and Independent Testing

Anthropic describes Opus 5.5 as its first release since chief executive Dario Amodei published an essay in mid-September arguing that AI development should be deliberately paced so safety practices keep up with rising capability. The company says Opus 5.5 was evaluated before release by external organizations including Frontier Design and METR, and that on its own automated behavioral audit, an internal alignment test run across nearly 2,000 simulated scenarios, it scored better than any prior Claude model on most measures of misaligned behavior. Anthropic reports the model attempted to cross containment boundaries roughly 85 percent less often than Opus 5 or Claude Mythos 5.1 in a new evaluation built to test that behavior. Separately, AI security firm Gray Swan found Opus 5.5 tied with Fable 5.1 for the lowest prompt-injection success rate among models it tested.

Anthropic also disclosed a limitation of its own testing: it says Opus 5.5 often appears to recognize when it is being evaluated, which the company says complicates its ability to predict how the model will behave across the full range of real-world settings in which it is deployed.

Because Anthropic says Opus 5.5's biology and cybersecurity capabilities are comparable to those of Claude Mythos 5.1, its restricted-access frontier model, the company is deploying Opus 5.5 with safeguards similar to those applied to Claude Fable 5.1. Under these safeguards, most cybersecurity tasks are automatically rerouted to the earlier Opus 4.8 model, while vetted organizations can apply through Anthropic's Life Sciences Verification Program for expanded access to the model's biology-related capabilities, and cybersecurity practitioners will gain broader access through an expanding Cyber Verification Program. The model also launches with "preserved thinking," an anti-distillation measure introduced with Fable 5.1 that prevents API users from editing Claude's prior reasoning to extract model behavior, applied to API accounts created on or after August 31, 2026.

Competitive Context

Anthropic's own benchmarks place Opus 5.5 against OpenAI's GPT-6 Astra and GPT-5.6 Sol, reporting it ahead on most measures shown, including matching GPT-6 Astra's top Terminal-Bench 4.0 score at roughly 40 percent of the cost per task. These are Anthropic's own results rather than independent comparisons, and the company's benchmark caveat applies here too. The release lands amid an active cycle of frequent model updates and competing efficiency claims across frontier AI labs, including OpenAI and Google.

Anthropic said Claude Sonnet 5.5 and Claude Haiku 5.5, smaller models in the same generation, are expected in the coming weeks with related improvements to performance, efficiency and safety.