Anthropic drops Claude Opus 5 — near-Fable 5 at half the price

Right, so Anthropic's come out swinging — two big moves in one week. On 20 July they made Fable 5, their pricey flagship, a permanent part of every Max and Team Premium plan (at 50% of the limits). Then four days later, 24 July, they dropped a whole new model — Claude Opus 5 — that they reckon is nearly as good as Fable at half the price. Bang, bang. Nobody's even caught their breath. Here's everything that actually landed.

What shipped

What's actually new

The headline's the price — half of Fable — but a couple of things under it matter more to me. One, there's an effort dial now — low, medium, high — so you're not paying top whack to think hard about an easy job; you tell it how much to bother. Crank it to full and Anthropic reckon it's within half a percent of Fable, at half the cost. Two, and this is the real one: it checks its own work. It verifies what it did and keeps going until it's actually right, instead of confidently handing you rubbish. For anyone running agents, that's the whole game — fewer cock-ups, less time you spend cleaning up after it.

There's more they've crammed in. A Fast mode that runs about two and a half times the speed, and it thinks with roughly a seventh of the tokens the old Opus burned — so it's leaner, not just quicker. The visuals took a proper jump — charts, interactive illustrations, presentation stuff; early users said it's the clearest step up they've seen, and it can knock out things like wind-tunnel simulations. It got genuinely better at science too — a solid bump on organic chemistry and protein work, and one customer said it "behaves more like a careful scientist." And it's more independent: it'll write its own test harnesses and code pipelines and just get on with the job.

On safety they're calling it their "most aligned" model yet — the lowest rates of the sneaky deceptive behaviour they test for. And here's one that matters if you build agents like I do: its cyber safety filters trip about 85% less often than Fable's, so it stops shutting itself down mid-task on legit security work. Less getting flagged for doing nothing wrong.

Does it actually beat Fable?

For the money, it's no contest. It leads Fable on most of the published benchmarks — the reasoning tests, the knowledge work, three times the next-best model on the hard ARC-AGI puzzle — all at half the price, and it does one of the big computer-use tests at a third of the cost. Where Fable still wins: the very hardest pure coding, by a hair, and legal. And Mythos — their locked-down security model — still owns cybersecurity. So it's not a clean sweep. But as the new everyday default at half the price? Fair enough.

And it's not just their own numbers

The early customers backed it up, which counts for more than a benchmark chart. The Devin lot said it nailed the gnarly debugging and root-cause stuff. Zapier got a 100% pass rate on business-automation tasks — and about 1.5× the next-best model. A genomics team said it works "like a careful scientist," and a few said it was the clearest jump they'd seen in the quality of visual work for presentations. Real teams, real jobs — not just Anthropic's own slides.

Why now

Two things pushed it. Their last model, Opus 4.8, had fallen behind — getting beaten by OpenAI's GPT-5.6 Sol and even a free open-source model, Kimi K3. And having just made the pricey Fable a permanent plan feature, a near-as-good model at half the API price is the obvious next punch — keep your crowd, undercut the rivals. Word is there's a Fable successor lined up for August to take on GPT-6. Same move Google pulled last week: refresh the mid-tier while the next flagship cooks.