Anthropic shipped Claude Opus 5 on Friday, July 24, 2026, and the pitch is unusual for a flagship launch: this is not the smartest model the company has ever made. Fable 5 still holds that crown, and the restricted Mythos 5 sits above both. Instead, Opus 5 is positioned as the model you reach for by default — the one you run all day without watching the meter. It lands at $5 per million input tokens and $25 per million output tokens, exactly what Opus 4.8 cost before it, and Anthropic claims it approaches Fable 5's intelligence at half Fable 5's price.
That framing tells you where the AI market has moved in 2026. The race is no longer only about the top of the leaderboard. It is about which lab can put near-frontier capability at a price that makes agents economical to run in a loop.
The efficiency argument, not the intelligence argument
Anthropic's entire case for Opus 5 rests on a piece of arithmetic. An AI agent burns tokens by working in a loop — opening files, running commands, reading errors, and trying again — and a single task can consume hundreds of thousands of tokens. A model that finishes a job in four attempts at $5 each beats one that grinds through twenty attempts at $2 each. Cheaper per token is not the same as cheaper per task.
So Anthropic is selling something harder to benchmark than raw knowledge: the ability to check its own work and keep going when the obvious path is blocked. The company offers three illustrative cases. In one, Opus 5 was asked to rebuild a machine part as a 3D model in FreeCAD but deliberately given no way to see the source drawing — so it wrote its own computer vision pipeline to extract the geometry from raw pixels, then rebuilt the part. In another, it traced the root cause of a bug in a widely used open-source package manager and fixed an edge case the community's own patch had missed. In a third, an engineer at a trading firm had it build a market-data feed for a new exchange with no live feed to test against — so it built its own test harness to verify it was parsing the data correctly.
These are Anthropic's own accounts, and none has been independently verified. Read them as a description of the behavior Anthropic wants to sell, not as neutral results.
What the benchmarks say — and the asterisk on all of them
Anthropic published Opus 5 results against Fable 5, its own Opus 4.8, and OpenAI's GPT-5.6 Sol. Several benchmarks come from outside organizations, but every run below was executed by Anthropic, not an independent lab. Treat the numbers as vendor-reported.
| Benchmark | What it measures | Opus 5 | Best rival |
|---|---|---|---|
| Frontier-Bench v0.1 | Coding via a terminal | 43.3% | GPT-5.6 Sol 34.4% |
| OSWorld 2.0 | Operating a computer like a human | 70.6% | Fable 5 66.1% |
| AutomationBench | Finishing a business task end to end | 26.0% | GPT-5.6 Sol 18.1% |
| ARC-AGI-3 | Learning a game's rules with no instructions | 30.2% | GPT-5.6 Sol 7.8% |
| GDPval (Artificial Analysis Elo) | Real professional deliverables | 1,861 | Fable 5 1,747 |
The ARC-AGI-3 result is the eye-catcher. When the ARC Prize Foundation launched that benchmark in March 2026, every frontier model scored under 1% while human testers solved all of it. Opus 5's 30.2% against GPT-5.6 Sol's 7.8% and Opus 4.8's 1.5% is a genuine jump on a test designed to resist memorization.
One footnote deserves more scrutiny than Anthropic gave it. On the Frontier-Bench chart, the company notes that Opus 4.8 stood in as a fallback whenever a safety classifier refused a request from Opus 5 or Fable 5 — meaning some unstated portion of both scores reflects work done by an older model.
Where it loses
To its credit, Anthropic published four benchmarks where Opus 5 came up short and shipped them anyway. On DeepSWE v1.1, GPT-5.6 Sol won 72.7% to 68.8%. On Humanity's Last Exam without tools, Fable 5 edged it 56.5% to 56.3%. On a held-out legal benchmark, Fable 5 won 13.3% to 11.7%. And on HealthBench Professional, the restricted Mythos 5 led 66.0% to 59.8%. None of these is a blowout, but they puncture the idea that a cheaper model is automatically a strictly better one.
There is also a number that cuts against the usual marketing instinct. Anthropic's automated behavioral audit scores models from 1 to 10 for misaligned behavior, where lower is better. Opus 5 scored 2.30, its best result to date — while Sonnet 5, released just last month, scored 3.35, the worst of the four models tested.
Security tiers and the data-retention change
Anthropic draws a sharp line between finding a software flaw and building a working exploit from it. On an OSS-Fuzz evaluation, Opus 5 identified vulnerabilities in 79.4% of challenges — nearly matching restricted Mythos 5's 80.0%. But on turning those flaws into exploits, Mythos 5 succeeded on 13 challenges to Opus 5's four. That gap is precisely why Opus 5 ships to everyone and Mythos 5 does not.
Opus 5's filters are tuned to block binary-based vulnerability scanning, penetration testing, and exploit generation, and Anthropic says they should trigger about 85% less often than Fable 5's. When a filter does fire inside Claude.ai, Claude Code, or Claude Cowork, the request quietly falls back to Opus 4.8.
For businesses, the most consequential line in the announcement is not a benchmark. Opus 5 carries no data-retention requirement for general access — prompts are not held for 30 days the way Fable 5's are. If you handle client records, medical data, or anything under a privacy regime like GDPR, that single sentence may matter more than any Elo score.
How to get it
Opus 5 is available now on Anthropic's own apps and as claude-opus-5 on the API, with a Fast mode running roughly 2.5x the speed for double the price. It is the default model on the Max plan and the strongest option on Pro. In the broader Claude ladder, it sits between Sonnet 5 ($2, on introductory pricing through August 31) and Fable 5 ($10), with Mythos 5 unavailable to the public.
Two smaller beta features ship alongside it, and both are quietly useful for agent builders: you can now swap which tools Claude has access to mid-conversation without losing the cached prompt, and API requests flagged by a safety filter can be automatically routed to another model instead of being refused outright.
The Bottom Line
Claude Opus 5 is a bet that the interesting frontier in 2026 is economic, not intellectual. Anthropic is not claiming the smartest model; it is claiming the most sensible one — near-frontier capability, the lowest misalignment score it has recorded, no default data retention, and pricing held flat from the model it replaces. The catch is that every headline number is Anthropic's own, run in Anthropic's harness, with an older model occasionally filling in behind a safety classifier. The real verdict will come from independent evaluations like Artificial Analysis and the public coding leaderboards over the next couple of weeks. Until then, the smart move is to test Opus 5 on your own workloads — where cost-per-task, not cost-per-token, is the only benchmark that pays your bill.
