Anthropic's Opus 5 is smaller and cheaper, yet it beats its bigger sibling

Anthropic launched a new model called Opus 5 on July 24, and it comes with a twist. Usually a fresh flagship is bigger, hungrier, and more expensive than what came before. Opus 5 goes the other way: it is smaller and cheaper than Anthropic’s pricier Fable 5, yet it matches or slightly beats it on several benchmarks. A benchmark, in plain terms, is a standardized test that lets you compare models on the same tasks.

The number developers noticed is the price. Opus 5 costs $5 per million input tokens and $25 per million output tokens. Tokens are the small chunks of text a model reads and writes (roughly three-quarters of a word each), and you pay per token. That price is the same as the older Opus 4.8 but well below Fable, so you get more capability for no extra spend. According to reporting by Ars Technica, Opus 5 also edges out OpenAI’s GPT-5.6-Sol across most task categories, while landing as an efficiency step rather than a giant leap.

Here is what is really going on, put simply. The loudest question in company engineering teams right now is not “how smart is the model” but “how much does it cost to run at scale.” Anthropic built Opus 5 for exactly that pressure. It also made a deliberate safety choice: the company did not give Opus 5 cutting-edge training on offensive cybersecurity, so the model is weaker at finding software exploits. Because its risk profile is lower, it needs fewer guardrails. Anthropic expects its safety classifiers (automatic filters that block risky requests) to trigger about 85% less often than on Fable 5. Fewer false alarms means fewer interrupted workflows. A new opt-in beta called Automatic Fallbacks goes further: when a filter does fire, it quietly reroutes the request to a smaller model and returns a working answer instead of an error.

A fair caveat: this is an iterative release, not a breakthrough, and competition is fierce. Ars Technica notes the Chinese open-weight model Kimi K3 offers similar performance at even lower output prices, so no single model holds a cost edge for long.

What this means for you: If you use Claude through an app, you may simply notice answers getting a little better and refusals getting rarer. If you build with the API, Opus 5 is the practical pick: strong results, predictable behavior, and lower bills. For most people, nothing changes today, but the trend is clear. The 2026 race is shifting from “biggest brain” to “best value,” and that competition tends to be good for whoever is paying.

Sources

• The Cryptonomist: Anthropic Opus 5 Launch
• Releasebot: Anthropic release notes, July 2026