The number worth dwelling on in Anthropic’s Opus 5 announcement isn’t a benchmark result. It’s $5 per million input tokens and $25 per million output tokens — exactly what the model it replaces already cost.
That pricing is basically the entire story. Anthropic released Opus 5 today, the newest iteration 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 gets close to Fable for roughly half the money.
The benchmark chart reads as iteration, not revolution
Anthropic’s own chart, which spans benchmarks including Frontier-Bench and DeepSWE, places Opus 5 roughly level with Fable on coding tasks, or a little in front. Set against Opus 4.8 and OpenAI’s GPT-5.6-Sol, it ostensibly comes out ahead in nearly every category.
Pay attention to the shape of those gains, though, not just their direction. It’s a step up — not the sort of leap Opus 4.5 delivered in agentic coding. Anyone waiting for the ground to shift again will be waiting a while longer.
Zoom out across a full year of releases and the progress looks dramatic. Line up any two consecutive releases and it looks modest. Both readings are accurate, and which one registers with you depends on how closely you’ve been following along.
One thing Anthropic made worse on purpose
This is the piece that sits awkwardly next to the usual launch-day story. Anthropic deliberately withheld cutting-edge training on cybersecurity tasks from Opus 5, and the results reflect it: the model falls well short of both Fable and Mythos in that area.
By the company’s account, Opus 5 is relatively good at locating cybersecurity vulnerabilities. Acting on them is a different question. Owing to choices made during training, Anthropic says the model is “substantially behind Mythos 5 on the exploitation of those vulnerabilities.”
There’s a practical knock-on effect: Opus 5 doesn’t come with all of the same controversial protections Fable shipped with, among them the policy of keeping data for review for 30 days should an incident occur. Whether you read that as a feature or a hole comes down to what you’re building.
Kimi K3 is the figure Anthropic has to respond to
Opus 5 holds the line on price against its predecessor and comes in under Fable. Fair enough. But then there’s Kimi K3, the Chinese open-weight model announced recently, offering comparable performance at $15 per million output tokens.
Call it a $10 gap per million output tokens — against something you can host yourself. For teams burning tokens on routine work, that arithmetic starts to sting quickly.
The real threat is model routing
Both Cursor and Meta have been working on model routers — systems that automatically select from a spread of models of differing size and capability based on what a given prompt actually demands. The reasoning is straightforward: don’t burn Fable-tier tokens on work that never needed Fable.
Each router that ships makes paying frontier prices a little more discretionary. And among software developers and engineering managers, the discussion at the moment centers on cost rather than capability, with genuine momentum gathering behind open-weight models, local models and other alternatives to frontier systems.
What this means in practice
Already running Opus 4.8? The upgrade is effectively free, since your costs don’t move — so take it. Paying Fable rates for general coding work? That’s the part of this release worth acting on.
And if your work is about exploiting vulnerabilities rather than surfacing them, Anthropic has said outright that this isn’t your model. Reread that line about Mythos 5.
From here, the squeeze on Anthropic is arithmetic. Keep driving token costs down, or keep delivering more performance at a flat price as it did today — or watch usage slide toward smaller and open models the instant those become good enough to handle the easy half of the job.
