Anthropic’s Opus 5 isn’t the leap Opus 4.5 was, and that’s fine

The most interesting number in Anthropic’s Opus 5 announcement isn’t a benchmark score. It’s $5 per million input tokens and $25 per million output tokens, which is the same as what you were paying for the model it replaces.

That’s the whole story here. Anthropic shipped Opus 5 today, the latest version of the model that’s become a default pick for coding and other software development work, and the pitch isn't that it's dramatically smarter. The pitch is that it’s close to Fable for roughly half the money.

The benchmark chart says iterative, not revolutionary

Anthropic’s own chart, covering benchmarks including Frontier-Bench and DeepSWE, puts Opus 5 at about the same level as Fable on coding tasks, or slightly ahead. Against Opus 4.8 and OpenAI’s GPT-5.6-Sol, it ostensibly wins across just about every category.

But look at the shape of the gains rather than the direction. This is a step up, not the kind of jump Opus 4.5 delivered in agentic coding. If you were expecting the ground to move again, it didn’t.

Stretch the timeline out to a full year of releases and the improvements look dramatic. Compare any two consecutive releases and they look modest. Both things are true, and which one you notice depends on how often you’re paying attention.

Anthropic deliberately made it worse at one thing

Here’s the part that doesn’t fit the usual launch-day narrative. Anthropic specifically avoided giving Opus 5 cutting-edge training on cybersecurity tasks, and it shows: the model trails Fable and Mythos there by a wide margin.

The company says Opus 5 is relatively good at finding cybersecurity vulnerabilities. Actually using them is another matter. Because of decisions made in training the model, Anthropic says it is “substantially behind Mythos 5 on the exploitation of those vulnerabilities.”

One practical consequence: Opus 5 doesn’t carry all of the same controversial protections Fable had, including the policy of keeping data for review for 30 days in case of an incident. Whether that reads as a feature or a gap depends entirely on what you’re building.

Kimi K3 is the number Anthropic has to answer

Opus 5 matches its predecessor on price and undercuts Fable. Fine. Then there’s Kimi K3, the recently announced Chinese open-weight model, at $15 per million output tokens with similar performance.

That’s a $10 gap per million output tokens against a model you can run yourself. For teams burning tokens on routine work, that math gets uncomfortable fast.

Model routers are the real threat

Cursor and Meta have both been building model routers, systems that automatically pick from a range of models of varying size and capability depending on what the prompt actually requires. The logic is simple enough: don’t spend Fable-tier tokens on work that doesn’t need Fable.

Every router that ships makes frontier pricing a little more optional. And the conversation among software developers and engineering managers right now is mostly about cost, not capability, with real momentum behind open-weight models, local models and other alternatives to frontier systems.

What to do with this

If you’re already on Opus 4.8, the upgrade is free in the sense that costs stay flat, so take it. If you’re paying Fable prices for general coding work, Opus 5 is the version of this release worth acting on.

If your work involves exploiting vulnerabilities rather than finding them, Anthropic has told you plainly that this isn’t the model. Go read that Mythos 5 line again.

The pressure on Anthropic from here is arithmetic. Keep pushing token costs down, or keep handing out more performance at the same price the way it did today, or watch usage drift toward smaller and open models the moment those get good enough for the easy half of the job.