Anthropic's Opus 5 is about token efficiency, not a capability leap

Ronni Holmvig Strøm · 2026-07-26

Anthropic rolled out Opus 5 on Friday, the newest version of the model that has become a default for coding and agentic software work. On Anthropic's own benchmarks it edges out Opus 4.8 and

Anthropic rolled out Opus 5 on Friday, the newest version of the model that has become a default for coding and agentic software work. On Anthropic's own benchmarks it edges out Opus 4.8 and OpenAI's GPT-5.6-Sol, and it lands within reach of the much larger Fable at roughly half the price. The press read the release as a mild letdown: an iterative cost story, not a capability leap. That framing has the emphasis backward. At the point where these models actually get used, cost is the capability.

Cost Is a Capability

The reflex is to treat efficiency as the consolation prize you offer when the benchmark curve flattens. But a model that delivers something just shy of the frontier at half the token spend does not do less. It does the same work for people who could not previously afford to hand it that work. Opus 5 sits at $5 per million input tokens and $25 per million output, on par with its predecessor and well under Fable. The performance went up. The bill did not. That is a capability improvement measured in the only unit that survives contact with a production budget.

Developers already know this, which is why the discourse among engineering managers is about cost and not leaderboards. The interesting movement is downstream. Cursor and Meta have both built model routers, systems that read a prompt and pick the smallest model that can handle it, reserving the expensive frontier calls for the prompts that need them. Once selection becomes infrastructure, the question stops being "which model is best" and becomes "what is the cheapest model that clears the bar for this task." Opus 5 is built to be the answer for a wide band of those tasks.

The Frontier Is Splitting in Two

There are now two different things a model release can be. One is a reach for the ceiling: Fable and Mythos, tuned to do things nothing else can, priced accordingly. The other is a reach for the useful middle, where most real work lives, optimized so that the useful middle keeps getting cheaper. Opus 5 is unmistakably the second kind, and the split is deliberate.

The clearest tell is what Anthropic left out. The company chose not to give Opus 5 cutting-edge training on cybersecurity exploitation, so it can find vulnerabilities reasonably well but sits, in Anthropic's own words, "substantially behind Mythos 5 on the exploitation of those vulnerabilities." A lab that can build the sharper capability and declines to ship it here is deciding which capabilities belong in which product. A capability withheld on purpose is a design choice, and it signals a field that has stopped assuming every model should be maximal on every axis.

Competition Sets the Floor

The reporting notes the Chinese open-weight model Kimi K3 landing near Opus 5's performance at around $15 per million output tokens, and open and local models keep creeping up the difficulty curve. That pressure is exactly what forces a frontier lab to hand you more performance for the same price. When the open-weight floor rises, the paid tier has to justify itself on something other than being the only option, and the honest way to do that is efficiency.

This is the market working the way you would want it to. A year of updates still compounds into a dramatic jump, and the year-over-year curve stays steep even when any single step looks modest. What changed is that the steps now arrive with the price attached, and the price is falling.