SAN FRANCISCO, July 25 — Artificial-intelligence lab Anthropic on Thursday released Claude Opus 5, its most capable reasoning model to date, claiming performance that approaches industry leader Fable 5 while undercutting rival pricing by roughly 50 percent.
The announcement intensifies an arms race among frontier AI developers that has accelerated dramatically in recent weeks, with Chinese labs unveiling their own large-scale models and Silicon Valley giants racing to integrate reasoning capabilities into consumer products.
In benchmark tests, Anthropic said Opus 5 set a new record on ARC-AGI-3, a rigorous evaluation designed to measure abstract reasoning and skill acquisition — a domain where earlier Claude generations had lagged behind competitors. The model also scored within striking distance of Fable 5 on graduate-level science and coding challenges, while using “significantly fewer reasoning tokens” than its predecessor to reach similar conclusions, according to internal evaluations.
Perhaps more significant for enterprise customers, Anthropic said Opus 5 demonstrates markedly improved self-correction abilities. “It is better at checking its own work, spotting logical errors and iterating,” the company wrote in a technical blog post. “For complex software engineering and research workflows, this translates into fewer hallucinations and less time spent on prompt engineering.”
The pricing strategy signals Anthropic\’s intent to compete aggressively on value. At roughly half the cost of comparable-tier offerings from OpenAI and Fable, Opus 5 could attract price-sensitive developers and midsize companies that have been priced out of frontier-model access. Industry observers noted the timing is particularly pointed: just days earlier, more than 20 technology firms including Nvidia and Meta signed an open letter endorsing open-weight AI models, a movement that threatens the business model of closed API providers.
Elon Musk, whose xAI has championed open-source releases, responded to the announcement by declaring that “all code running on X systems will be open sourced” — a pledge that, if fulfilled, would add another volatile element to an already tumultuous competitive landscape.
Security researchers offered a more cautious assessment. Anthropic acknowledged that Opus 5\’s safety classifiers are “noticeably more permissive” than those of Fable 5, potentially increasing the risk of jailbreaks or misuse in sensitive applications. Cybersecurity benchmarks placed the model\’s defensive capabilities near those of Mythos 5, a specialized security-oriented system, but independent auditors said they had not yet completed red-team evaluations.
“The performance jump is real,” said Dr. Yuki Tanaka, an AI researcher at Stanford\’s HAI institute. “But the safety trade-offs deserve scrutiny. We\’re seeing a pattern where every major lab is pushing capability boundaries faster than evaluation frameworks can keep up.”
Anthropic said Opus 5 is available immediately through its API and will roll out to Claude Pro subscribers within two weeks. The company projected the model would become the default choice for “enterprise daily work, software development, scientific research and agentic workflows” — a direct challenge to incumbents who have dominated those use cases.