Anthropic Debuts Claude Opus 5, Balancing High Performance with Lower Costs and Stricter Safety

Anthropic launched Claude Opus 5 on Friday, a model that matches or exceeds its Claude Fable 5 across most benchmarks while costing half as much to operate, priced at $5 per million input tokens and $25 per million output tokens. The model outperformed OpenAI's GPT-5.6 Sol in nearly every standardized category except agentic coding, and Anthropic called it its safest model yet with record-low deceptive behavior and enhanced jailbreaking resistance. The release targets enterprise clients seeking cost-effective AI amid tightening budgets, following a brief federal ban on Fable 5 and industry pressure from lower-cost alternatives like Moonshot AI's Kimi K3.

Anthropic launched its newest model, Claude Opus 5, on Friday, promising near-frontier intelligence at half the operating cost of its premium predecessors. The release https://www.anthropic.com/news/claude-opus-5 arrives as enterprise clients increasingly demand cost-effective artificial intelligence AI solutions to manage ballooning technology budgets. Claude Opus 5, which is available, serves as the primary engine for the $100-per-month Claude Max tier and is accessible to subscribers on the $20-per-month Claude Pro plan. Priced at $5 per million input tokens and $25 per million output tokens, the model is engineered to act as a cost-efficient, daily workhorse for enterprise knowledge work and coding tasks. According to company data, Opus 5 matches or exceeds the performance of Anthropic’s Claude Fable 5 across most benchmark tests, with minor exceptions in specialized legal and health domains. In head-to-head evaluations, Opus 5 outperformed rival OpenAI’s GPT-5.6 Sol in nearly every standardized category, falling behind only in agentic coding. The model showcased notable speed when resolving novel challenges. In tests measuring problem-solving performance on unseen tasks, Opus 5 ran three times faster than the next-best model and scored nearly four times higher than GPT-5.6 Sol. Anthropic also touted the release as its most capable generally available model for scientific research, though its restricted Mythos 5 model still leads in specialized biology benchmarks. Safety and alignment remain central to the Opus 5 rollout. Anthropic labeled the release its “safest model yet,” citing record-low rates of deceptive behavior and enhanced resistance to jailbreaking or malicious prompt manipulation. Crucially, Anthropic deliberately excluded offensive cybersecurity training from Opus 5. While the model remains capable of identifying system vulnerabilities, its ability to execute exploits was intentionally constrained. The cautious architecture follows earlier regulatory scrutiny, including a brief federal ban on Fable 5 over safety vulnerabilities, alongside heightened industry concerns following recent security incidents involving autonomous AI agents. The launch comes at a critical juncture for the tech industry, where high compute costs have forced major corporations to reevaluate their AI expenditures. Market pressure intensified earlier this year when major adopters, including Uber Technologies Inc., exhausted their annual 2026 AI budgets within four months, prompting executives to demand clearer returns on investment. Simultaneously, Western tech firms face growing competitive pressure from lower-cost alternatives abroad, such as Chinese startup Moonshot AI’s recently released Kimi K3 model. By delivering a balance of top-tier intelligence, reduced token costs, and strict safety guardrails, Anthropic aims to maintain its foothold among corporate clients seeking high-performing AI that fits within tightening bottom lines.