Anthropic has released Claude Opus 5.5, the first model in its new 5.5 family, positioning it as a performance-per-dollar play rather than a raw capability moonshot. It matches or approaches Claude Fable 5.1 on most benchmarks while costing 40% less to run than its predecessor Opus 5, with input tokens at $4/M (down from $5) and cache reads at $0.20/M (down from $0.50, a 60% cut). Output generation is 30% faster. The pricing structure aggressively targets the agentic coding workload where cache reads dominate costs. The benchmark numbers tell a nuanced story. Opus 5.5 leads on Terminal-Bench 4.0 (66.4% vs. GPT-6 Astra's 57.9%), FrontierCode (54.4% vs. Astra's 53.3%), and Humanity's Last Exam with tools (67.7%). But Anthropic itself notes that 'at these levels of capability, benchmark margins have become a less reliable guide to real-world differences' and that the gap between Opus 5.5 and Fable 5.1 is narrower than scores suggest. On AutomationBench, GPT-6 Astra actually edges Opus 5.5 (41.4% vs 40.0%), and on Terminal-Bench-Science, Astra leads substantially (64.6% vs 58.7%). The efficiency case is where Opus 5.5 genuinely separates. An early tester audited and fixed a 200,000-line codebase in under three hours; Opus 5 took over 20 hours and used 2.5x the tokens. In the HAProxy C-to-Rust translation, Opus 5.5 finished in 9.5 hours versus Fable 5.1's 12, at 51% lower cost. On Terminal-Bench 4.0, it matches GPT-6 Astra at roughly 40% of the cost per attempt. On FrontierCode, it beats Astra at about 20% of cost. This is the competitive moat Anthropic is building: not the smartest model, but the smartest model per dollar. Safety and alignment received unusual emphasis. Opus 5.5 posts the best scores on Anthropic's automated behavioral audit — thousands of simulated scenarios testing for hard-to-reverse actions, boundary violations, and prompt injection resistance. Gray Swan's independent benchmark ties it with Fable 5.1 for the lowest prompt injection success rate of any model tested. Because Opus 5.5 approaches Mythos 5.1 in biology and cybersecurity capability, it ships with comparable safeguards: vetted organizations must apply through Life Sciences and Cyber Verification Programs. Anthropic has broadened its alignment testing to cover longer tasks, impossible tasks, and scenarios modeled on real incidents. The deployment architecture includes a classifier that screens every action before execution, an open-source auditable sandbox, and code review that catches vulnerabilities pre-merge. On the knowledge work side, an internal test had Opus 5.5 produce reports from a simulated web where earnings data was hard to find — 16 of 18 reports cleared a quality bar where any invented figure would fail, versus zero passes for Fable 5.1 or Opus 5. The broader strategic signal: Anthropic is executing on its 'pacing the frontier' pledge by shipping a model that is materially better at the cost curve rather than at the capability ceiling. External evaluators (Frontier Design, METR) tested Opus 5.5 before release. Sonnet 5.5 and Haiku 5.5 will follow in weeks, presumably extending the efficiency-first approach down the product line. Pro, Max, Team, and Enterprise plans get higher usage limits and a saveable rate-limit reset. The honest read: this is a consolidation release dressed in launch-day fanfare. Opus 5.5 does not blow past Fable 5.1 or GPT-6 Astra on raw capability. What it does is make frontier-tier performance dramatically cheaper to deploy at scale, especially for the long-running agentic coding tasks that enterprises actually pay for. The safety infrastructure — external audits, verification programs, action-level classifiers — sets a procedural standard that competitors will be pressured to match. Whether that standard is sufficient is a separate question, but the receipts are more detailed than anything else on the market.
