Claude Opus 5: the flagship experience, at half the price.
On July 24, 2026, Anthropic released Claude Opus 5 — and quietly rearranged what you get on every paid Claude plan. It replaces Opus 4.8, becomes the default model on Claude Max, and is now the strongest model on Claude Pro. Anthropic's pitch: intelligence close to its Fable 5 flagship, at half the API price. Here’s what actually changed, and what to do about it.
Opus 5 is available now on every Claude surface — apps, Claude Code, Cowork, and the API (claude-opus-5, $5/$25 per million tokens, unchanged from Opus 4.8). Anthropic says it's the new state of the art on coding and knowledge-work benchmarks like Frontier-Bench and GDPval-AA, an effort dial lets you trade cost against depth, and a Fast mode runs ~2.5× quicker at 2× the base price.
01 Where it lands in the lineup
Anthropic's model family now reads like this: Fable 5 stays the most intelligent generally available model. Opus 5 sits just below it — Anthropic's own charts show it within a whisker of Fable 5 on several coding benchmarks — at half the price and built for everyday use. Sonnet 5 remains the fast free-tier default, and Haiku handles the lightweight jobs.
02 The effort dial
Opus 5 exposes an effort setting: how hard the model works before it answers. Crank it up for a gnarly analysis; keep it low for routine tasks and it answers faster while spending fewer tokens. Anthropic claims that even at the lowest effort, Opus 5 passes more tasks than any other model on Zapier's business-automation benchmark — which is the real story here: the floor got high.
The practical habit: stop treating "the best model" as one fixed thing. On hard problems, high-effort Opus 5 is a near-Fable experience. On everything else, low effort gets you a fast, cheap answer from the same brain.
03 It checks its own work
The through-line in early reports is verification behavior: Opus 5 opens its own web pages to catch layout bugs, builds test harnesses when none exist, and pushes back (politely, with reasons) when your design idea has a flaw. In one of Anthropic's examples, the model was asked to rebuild a machine part from a drawing it couldn't view — so it wrote its own computer-vision pipeline to read the pixels, then built the part. That thoroughness is what "agentic" is supposed to mean, and it's why this release matters beyond benchmark charts.
04 The safety plumbing moved too
Two routing changes worth knowing if you've followed our Fable 5 lessons:
- Biology questions: requests blocked by Fable 5's safety classifiers now route to Opus 5 (previously Opus 4.8) — so the fallback answers got smarter.
- Cybersecurity: Opus 5's own cyber classifiers are lighter-touch than Fable's (Anthropic expects ~85% fewer interventions), but flagged requests in the Claude apps still fall back to Opus 4.8.
05 Opus 5 or Fable 5?
- Default to Opus 5 for daily work: drafting, analysis, coding, agents, long documents. Near-flagship quality, faster and cheaper.
- Reach for Fable 5 when the last few percent of reasoning matters — the hardest strategy, research, or debugging problems.
- Let Sonnet 5 keep the small stuff. Quick questions don't need a heavyweight.
One honest caveat: the benchmark numbers above are Anthropic's own, published on launch day. Independent results will trickle in over the coming weeks — the pattern with Sonnet 5 and Fable 5 was that the headline claims held up directionally, with some gaps narrower than advertised.
Take a task you'd normally hand to Fable 5 — a real one from this week — and run it through Opus 5. Compare the answers side by side. If you can't tell the difference, you've just found out where your personal “good enough” line is, and it's cheaper than you thought.