Anthropic announced a new AI model called Claude Opus 5 on July 24, 2026. The company says it is a big step up from the previous Opus model, with the biggest improvements in complex reasoning, tasks that require multiple steps, and the ability to spend more processing time on a single question to give a better answer (Anthropic). The new model gets close to the performance of Anthropic's most powerful model, Claude Fable 5, but costs half as much. It also adds a faster mode and a feature that hands off a request to a simpler model if Opus 5 decides it cannot safely answer.

The price stays the same as the older Opus 4.8: $5 for every million words of input (what you send to the model) and $25 for every million words of output (what the model sends back). Anthropic says Opus 5 "comes close to the capabilities of Claude Fable 5 in many domains" and is stronger at complex coding tasks (The Verge). On CursorBench, a testing tool that measures how well AI models handle coding tasks, Opus 5 scores just under Fable 5 (Anthropic). Compared to outside competitors, Opus 5 is slightly cheaper than OpenAI's GPT-5.6 (The Verge).

Two Models, Two Purposes

Anthropic is being deliberate about how it positions its models. It calls Opus 5 the best choice for everyday knowledge work and biology, while Fable 5 is for "ambitious work and long-horizon AI agent projects" (The Verge). Fable 5 belongs to a category Anthropic calls "Mythos-class" — its most powerful but also most sensitive models. Fable 5 was taken offline for several weeks after government concerns, then brought back on July 1, 2026 with stronger cybersecurity protections (Anthropic; The Verge). Opus 5 arrives at a lower price without that history attached.

What Changed on Safety

Anthropic publishes a detailed document called a system card for each model, explaining how it behaves and what guardrails are in place. Opus 5's system card shows its safety protections match Fable 5's, with one difference: Opus 5 is allowed to help find security flaws in computer code at all access levels (Anthropic System Card). Anthropic calls Opus 5 "the most aligned Opus model and is the least susceptible to being tricked into misuse" (The Verge).

The new fallback feature matters for developers. When Opus 5 refuses a request for safety reasons, it automatically sends that request to a lower-tier model instead of just returning an error. That matters because some applications — like AI agents that carry out multi-step tasks on their own — break down if the model simply refuses to answer partway through.

A feature called Fast mode, still in research preview, offers faster responses at double the standard price. Teams that care more about speed than cost can switch it on without switching to a different model.

Anthropic spokesperson Danielle Ghiglieri confirmed the company continues to work with government partners on independent testing of its models, including Opus 5 (The Verge). Reuters called the release an "efficiency upgrade" in its July 24 coverage (Reuters.

A Summer Full of Announcements

Opus 5 arrives in the middle of a busy stretch for Anthropic. Claude Sonnet 5, a smaller and cheaper model, shipped on June 30 at $2/$10 per million words in and out. Anthropic described it as the most capable Sonnet yet for tasks involving tool use, such as controlling a web browser or a computer terminal (Anthropic). That same day, Anthropic joined Amazon, Microsoft, Google, and other partners in a group called Glasswing to propose a shared way of rating how hard it is to break through an AI model's safety rules (Anthropic). On July 9, former Federal Reserve Chair Ben Bernanke was appointed to Anthropic's Long-Term Benefit Trust, and the company launched a public initiative called "Inviting hard questions." On July 14, Anthropic pledged $10 million to Canadian AI research and released a product called Claude for Teachers. A $20 million donation to Public First Action followed on July 21 (Anthropic).

The pattern here is worth pausing on. Anthropic is releasing new models, setting up governance structures, and launching public trust initiatives all at the same time, rather than one after the other. Opus 5 is the product; the Glasswing safety framework, the Bernanke appointment, and the government testing partnerships are the scaffolding around it. Whether that scaffolding actually makes Anthropic stand out in a market where every major AI lab publishes safety commitments is something enterprise customers will decide with their purchases, not with press releases.

For engineering teams, the practical question is more straightforward. Opus 5 at $5/$25, with safety features close to Fable 5's, permission to find code vulnerabilities, and a fallback that prevents hard refusals, may be the easier model to put into production — even if Fable 5 is still the strongest on raw capability. For organizations that find Fable 5 too expensive or too restricted, Opus 5 closes most of the gap at half the cost.