Anthropic Slashes AI Costs With Claude Opus 5, Pricing It at Half of Fable 5 Ahead of IPO

Anthropic is betting that cheaper artificial intelligence will win over cost-conscious enterprises and Wall Street skeptics. The San Francisco-based startup on Thursday unveiled Claude Opus 5, a new model that it says approaches the capabilities of its most powerful publicly available system, Fable 5, while costing customers exactly half as much.

The launch marks a strategic pivot in the AI arms race. Instead of chasing raw performance at any price, Anthropic is packaging near-flagship intelligence into a model designed for high-volume, everyday business tasks. The move comes as the company prepares for a closely watched initial public offering later this year and as competition from both U.S. rivals and aggressive Chinese labs intensifies.

Claude Opus 5 is the fourth member of the Claude 5 family, sitting just below the company’s most advanced—and heavily restricted—Mythos 5 series. Anthropic positions Opus 5 as the default model for knowledge work, software development, scientific research, and autonomous AI agent workflows. The pricing is stark: Opus 5 costs $5 per million input tokens and $25 per million output tokens. Fable 5, which Anthropic released last month and describes as suitable for “days-long autonomous projects,” costs $10 per million input tokens and $50 per million output tokens.

“Cost has become a major issue,” the company noted, as business customers increasingly push back on the ballooning expense of running cutting-edge AI systems. The scrutiny is particularly acute because both Anthropic and its archrival OpenAI, the maker of ChatGPT, confidentially filed paperwork in June to go public. Investors are now examining their unit economics more closely than ever.

Anthropic says Opus 5 tops industry benchmarks for coding and knowledge work while consuming fewer computing resources than earlier versions. The company deliberately avoided training Opus 5 on offensive cybersecurity tasks, a sensitive area that has embroiled its Mythos line in controversy. Mythos models are so adept at finding and weaponizing software vulnerabilities that Anthropic restricts access to a handful of trusted partners. Opus 5 can find software flaws roughly as well as Mythos 5, according to the company, but it lags far behind in turning those flaws into working attacks.

Citing that reduced risk, Anthropic said its automated safety filters will intervene roughly 85% less often on Opus 5 than on Fable 5. Opus 5 also does not retain user data for 30 days, a practice Anthropic maintains for Fable 5 as part of its safety monitoring and misuse prevention efforts.

The safety architecture around Anthropic’s models has been shaped by direct government intervention. In June, the U.S. Commerce Department imposed export controls that briefly forced Anthropic to pull Mythos 5 and Fable 5 from public access. The restrictions were lifted at the end of that month. OpenAI, meanwhile, voluntarily submitted its GPT-5.6 Sol model to the administration for evaluation before a wider release.

These episodes illustrate how national security concerns are now a permanent feature of the frontier model business. By designing Opus 5 to be less capable in domains that regulators fear most—offensive cyber operations and sensitive biological research—Anthropic can offer a powerful model with fewer compliance headaches.