Anthropic Opus 5.5 beats GPT-5.6 Sol on coding at one-third the cost

For AI researchers and ML engineers, the Opus 5.5 release is a case study in balancing capability, cost, and safety. Anthropic claims the model outscored OpenAI's GPT-5.6 Sol on a software development benchmark at roughly one-third the operating cost, while external red-teaming by METR and Frontier Design and an 85% reduction in containment bypass attempts signal a deliberate security posture.

Anthropic on September 22, 2026 introduced Claude Opus 5.5, its first major flagship model since leadership publicly emphasized slower pacing for frontier AI releases. The launch pairs a list-price reduction from Claude Opus 5 with Anthropic's claim that Opus 5.5 delivers performance comparable to its top-tier Fable 5.1. API pricing is now $4 per million input tokens and $20 per million output tokens, a cut the company describes as 20 percent below Opus 5. Some promotional coverage also frames the model as roughly 40 percent more affordable than its predecessor, a broader claim likely reflecting the combination of lower list prices, cheaper caching, and faster generation rather than headline token rates alone. Cache reads, a critical cost driver for long-context coding and agentic workloads, fell 60 percent to $0.20 per million tokens, while text generation speed improved by more than 30 percent.

“

API pricing is now $4 per million input tokens and $20 per million output tokens, a cut the company describes as 20 percent below Opus 5.

The model is positioned for long-horizon agentic work such as large-scale code reviews, software migrations, and deep scientific analysis. Enterprise testers cited by Anthropic report productivity gains on complex data pipelines and multi-file codebases. The model is available on Amazon Web Services, Google Cloud, and Microsoft Azure, which gives Anthropic distribution across the three dominant hyperscalers and reduces friction for enterprise adoption. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks with similar performance, speed, and safety improvements, extending the cost-performance shift down-market.

The release landed amid an active debate over AI risk. Anthropic CEO Dario Amodei earlier in September called on the global AI community to slow the pace of new capability releases, according to Reuters. In line with that stance, Anthropic subjected Opus 5.5 to external testing by independent safety research organizations Frontier Design and METR before release. The company also says the model incorporates safeguards previously reserved for its most capable systems. In an internal evaluation, Anthropic reported Opus 5.5 was about 85 percent less likely than Claude Opus 5 or Mythos 5.1 to attempt to bypass containment boundaries. These safety metrics should be treated as company-reported figures rather than independently verified results.

On performance, Anthropic states that Opus 5.5 outscored OpenAI's GPT-5.6 Sol on a software development benchmark while costing roughly one-third as much to operate. The comparison matters because it shifts competitive positioning from raw capability to efficiency: if credible, it could pressure rivals to justify premium pricing when lower-cost alternatives approach frontier performance. The same sources note the model is comparable to top-tier Fable 5.1, although no detailed benchmark tables are included. None of the benchmark claims are independently verified in the source material, so they should be read as vendor assertions.