Anthropic shipped Claude Opus 5 on July 24, 2026, and priced it exactly where Opus 4.8 sat: $5 per million input tokens, $25 per million output tokens. That single decision is the whole story. You now get a model Anthropic says closes most of the gap to Fable 5, its top-tier system, for half of what Fable costs to run. Opus 5 is the new default on Claude Max and the strongest model on Claude Pro, starting today.

Here's the part every other write-up skipped: Opus 5 lands 24 days after Fable 5 finished a 19-day blackout. On June 12, the Commerce Department ordered Anthropic to cut off Fable 5 and Mythos 5 for every foreign national worldwide, after Amazon researchers found a jailbreak that pushed Fable 5 into identifying exploitable vulnerabilities. The controls lifted June 30, and Fable came back July 1. If you build on Claude and your main coding model can vanish for three weeks over a government directive, price stops being the only thing you weigh. Opus 5 isn't a Mythos derivative. It never touched that shutdown. That's a real, checkable reason to make it your daily driver rather than a marketing line, and none of the five pieces I read on this launch connected it.

On Frontier-Bench v0.1, Anthropic's internal coding and knowledge-work benchmark, Opus 5 scored 43.3%. Opus 4.8 scored 18.7%. Fable 5 scored 33.7%. So Opus 5 didn't just beat its predecessor, it beat the model that costs twice as much. On CursorBench 3.2 at max effort, Opus 5 landed within half a point of Fable 5's peak score, at half the price per task. Sualeh Asif, Cursor's co-founder, put it plainly: Opus 5 delivers near Fable-level intelligence at Opus speed and cost.

The jump on ARC-AGI 3, a benchmark built around genuinely novel problems rather than pattern-matched ones, is the biggest single number in the release: 30.2%, roughly three times the next closest model. On Zapier's AutomationBench, which scores whether a model can carry a business task start to finish rather than just begin it, Opus 5's pass rate runs about 1.5 times the next-best model at matching cost, and it beats every rival's best result even on its cheapest effort setting. Wade Foster, Zapier's CEO, said Opus 5 took a raw account-health workbook and ran a full churn-prevention sequence end to end, something previous models simply didn't pass.

On OSWorld 2.0, a computer-use benchmark, Opus 5 beats Fable 5's top score while spending roughly a third of the compute Fable needed to get there. That's the pattern across this release: Anthropic isn't claiming Opus 5 is smarter than Fable in an absolute sense. It's claiming Opus 5 gets close enough, cheaply enough, that the smarter model stops being the obvious choice for most day-to-day work.

Science tasks moved too. Opus 5 beats Opus 4.8 on every life-sciences evaluation Anthropic tracks, with the largest jump on organic chemistry work like reading molecular structure out of spectroscopy data, up 10.2 percentage points, and on predicting how sequence variants change a protein's function, up 7.7 points.

Anthropic did not train Opus 5 on cybersecurity tasks, the same call it made with Sonnet 5. The model still improved on that front just from getting more generally capable, and it now finds vulnerabilities at close to Mythos 5's rate. Writing working exploits from those vulnerabilities is a different skill, and that's where Opus 5 falls well behind. Anthropic's own OSS-Fuzz results show the gap clearly: similar vulnerability-discovery scores, a much wider spread on turning that discovery into an actual exploit. If your work touches offensive security research, Fable 5 (built on the Mythos line) stays the stronger pick, with the tradeoff being that it also carries the 30-day data retention Fable requires and the export-control exposure that just took it offline for three weeks.

On alignment, Anthropic's automated behavioral audit scores Opus 5 at 2.3 for misaligned behavior, the lowest of any current Claude model, ahead of Opus 4.8, Sonnet 5, and Fable 5 itself. Safety classifiers on Opus 5 are expected to fire around 85% less often than they do on Fable 5. When one does trip inside Claude.ai, Claude Code, or Claude Cowork, the request falls back to Opus 4.8 by default rather than getting refused outright.

No single competitor writeup put every current Claude model side by side. Anthropic's own model page doesn't either, since Fable and Mythos share a page and Opus and Sonnet each get their own. Here's the complete grid as of this launch:

Fable 5 launched June 9, 2026 at exactly double Opus 4.8's price, as the first publicly available Mythos-class model. Mythos 5 and Fable 5 are the same underlying model; the difference is which safety classifiers sit in front of it. That's why Fable, not Opus, was the one caught in the export-control order: the directive targeted Mythos-class weights, and Opus 5 was never in scope.

Every piece I read on this launch was written for an enterprise buyer comparing benchmark charts. None of them addressed the reader who's on Claude Pro, building solo, and watching the token bill. Here's the gap answer.

If you're on Claude Pro and currently defaulting to Sonnet, Opus 5 is now the strongest model in your plan at no extra cost, so there's no reason not to switch it on for anything that isn't pure high-volume drafting. If you were paying for Max specifically to get Fable-level output, run your own workload through Opus 5 first. Anthropic's numbers say it lands within half a percentage point of Fable on coding-agent benchmarks at half the cost per task; if your usage looks anything like Cursor's or Zapier's reported patterns, the savings show up fast without a real quality hit. The one case where you still want Fable or Mythos over Opus 5 is offensive security research or exploit-adjacent work, where Opus 5's weaker exploit-generation score is a real ceiling, not a marketing footnote.

The other thing worth saying plainly: Opus 5 not being Mythos-derived is now a legitimate reliability argument, not just a pricing one. A model that can be pulled by government order with no advance notice is a bad thing to build a production pipeline around as your only option. Opus 5 gives teams a fallback that was never exposed to that risk in the first place.

What does Claude Opus 5 cost? $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. A Fast mode is available at 2.5 times the speed for double the base price.

Is Opus 5 better than Fable 5? Not across the board. Opus 5 closes most of the gap on coding and knowledge-work benchmarks at half the price, and beats Fable outright on some, like Frontier-Bench v0.1. Fable 5 stays ahead on the longest autonomous tasks and on cybersecurity, particularly exploit generation.

Is Opus 5 the same model family as Fable or Mythos? No. Fable 5 and Mythos 5 share the same underlying model with different safety classifiers. Opus 5 is a separate line, which is also why it was unaffected by the June export-control suspension that took Fable and Mythos offline.

Where can I use Opus 5? It's live today on Claude.ai, the Claude API as claude-opus-5, Claude Code, and Claude Cowork. It's the default model on Claude Max and the strongest model available on Claude Pro.

Does Opus 5 have a data retention requirement? No, consistent with prior Opus models, general access to Opus 5 carries no data retention requirement. Fable 5 requires 30-day retention with no zero-retention option, which matters if you're evaluating the two for a compliance-sensitive workload.