Claude Opus 5, explained: what Anthropic shipped, and why it's half the price of Fable 5
On July 24, 2026, Anthropic shipped Claude Opus 5 (model ID claude-opus-5) at $5/$25 per million input/output tokens — half of Fable 5's price. It matches much of Fable 5's capability (CursorBench within 0.5%, OSWorld above it), more than doubles Opus 4.8 on Frontier-Bench (43.3%), and is the new default in Claude Code.
Shipped July 24, 2026: Claude Opus 5 (
claude-opus-5) is generally available on Claude.ai, Claude Code, Claude Cowork, and the Claude API, plus AWS, Google Cloud Vertex AI, and Microsoft Foundry. It’s the new default model in Claude Code and is included in paid plans (the new default on Claude Max, the strongest model on Claude Pro). The one-line pitch from Anthropic: much of Fable 5’s capability at half the price. Below is my builder’s read of what that actually means.
What did Anthropic actually ship on July 24?
On July 24, 2026, Anthropic released Claude Opus 5 — model ID claude-opus-5. If you build on Claude in production like I do, that’s the string you’ll be putting in config for the next while, and it’s the model Claude Code now reaches for by default.
The framing that matters most is the price-to-capability move. Opus 5 lands at $5 per million input tokens and $25 per million output tokens — the same list price the Opus tier has held — while, in Anthropic’s words, it “matches many of Fable 5’s capabilities.” Fable 5 costs $10/$50. So the headline isn’t “a bit better than Opus 4.8.” The headline is: near-Fable-5 quality at half of Fable 5’s price. That reframes a lot of model-selection decisions I’d already made this year.
I’ll walk through the benchmarks, the two behavioral changes that actually change how it codes, the new /effort control, and the caveats — then give you my honest bottom line.
The headline: Fable 5 capability at half the price
Two proof points make the “half the price” claim concrete rather than marketing:
- On CursorBench 3.2, Opus 5 “performs within 0.5% of Fable 5’s peak score, but at half the cost.”
- On OSWorld 2.0 (computer-use / agentic desktop tasks), it’s “surpassing Fable 5’s best result at just over a third of the cost.”
Read those again. On real agentic-coding and computer-use benchmarks, Opus 5 is essentially matching — and in one case beating — the model that until now was Anthropic’s ceiling, while costing a half to a third as much. For anyone who looked at Fable 5’s $10/$50 and quietly decided the top tier wasn’t worth it for everyday work, that math just changed.
Claude Opus 5 at a glance
How good is it really? The benchmarks
Here are the numbers that actually move my decisions, not the whole scorecard.
On Frontier-Bench v0.1, Opus 5 hits 43.3% — up from 18.7% on Opus 4.8. Anthropic frames that as surpassing all other models and “more than doubling” the prior Opus. On DeepSWE v1.1, it reaches 68.8%, up from 59.0% on Opus 4.8. Those are the two agentic-coding numbers I watch, because they track long, messy, multi-step work rather than tidy one-shot puzzles.
Opus 5 vs Opus 4.8: agentic benchmarks
Beyond coding, a few results stood out to me:
- ARC-AGI 3: score is “three times as high as the next-best model” — a real jump on the abstraction-and-reasoning benchmark that models have historically struggled with.
- Zapier AutomationBench: pass rate “around 1.5× the next-best model” — that’s the practical “can it wire up real automations” signal.
- GDPval-AA: state-of-the-art on this economically-grounded task set.
- Life sciences: 10.2 points higher than Opus 4.8 on organic chemistry, 7.7 points higher on protein prediction.
My caveat, same as always: these are Anthropic’s own reported numbers on Anthropic’s chosen benchmarks. They’re directionally strong and consistent with the pricing story, but the benchmark that matters is your workload. I’d still run Opus 5 against your own eval set before you flip a production default — which is exactly what I’d do on a client build, and why I keep an eval harness in the loop.
What changed under the hood: judgment and self-checking
The benchmark deltas are nice. The two behavioral changes are what I think you’ll actually feel day to day.
It exercises judgment before it acts. Anthropic says Opus 5 “more regularly asks clarifying questions before guessing, pushes back on flawed instructions, and considers the implications of its work before jumping ahead to implementation.” In practice that means fewer confidently-wrong tangents. When I hand an earlier model an ambiguous ticket, it picks an interpretation and runs — and sometimes that’s 20 minutes of work aimed at the wrong target. A model that stops to ask “did you mean A or B?” or says “this instruction will break X, are you sure?” produces cleaner code because it isn’t guessing.
It checks its own work as it goes. Opus 5 “checks its own work as it goes and catches issues that earlier models missed.” That’s the difference between an agent that writes a change and declares victory, and one that re-reads what it wrote, notices the edge case it broke, and fixes it before handing back. For long autonomous runs, self-checking is what keeps a small early mistake from compounding across an hour of work.
What actually changed vs Opus 4.8
Better judgment
- Asks clarifying questions before guessing
- Pushes back on flawed instructions
- Weighs implications before implementing
- Result: fewer confidently-wrong tangents
Self-correction
- Checks its own work as it goes
- Catches issues earlier models missed
- Recovers from errors mid-task
- Routes around blockers instead of stopping
Long-running work: hours, not minutes
This is the capability I care about most. Anthropic says Opus 5 “can work for hours on a task, recovering from errors and routing around blockers instead of stopping.”
“Instead of stopping” is the operative phrase. The failure mode of earlier agents wasn’t usually a wrong answer — it was giving up: hitting a failing test, a missing dependency, an unexpected API shape, and halting with a shrug. A model that routes around the blocker (try another approach, stub the dependency, read the error and adjust) is the one you can actually leave running on a real migration or refactor.
Combine that with the self-checking above and you get a different shape of system. You stop building elaborate scaffolding to chop every job into tiny supervised steps, and you start handing over larger tasks whole. That’s the same architectural shift I wrote about when Fable 5 landed — Opus 5 brings a big chunk of it down to the $5/$25 tier.
Setting effort with /effort
New control worth knowing: in Claude Code you set Opus 5’s effort level with /effort. It lets you trade intelligence against speed and token spend on a per-need basis.
The levels top out at Extra and Max, which “push hardest on difficult problems and use more of your usage limit.” Anthropic’s own guidance: use High or below for more routine work — it’s faster and lighter on your usage. So the mental model is a dial, not an always-on setting.
How to think about /effort
- High or below — routine workEveryday edits, refactors, glue code. Faster, lighter on your usage limit. This is the default posture.
- Extra — genuinely hard problemsGnarly bugs, tricky design work, long multi-file changes where a wrong turn is expensive.
- Max — push hardestThe difficult tail. Uses the most of your usage limit — reach for it when correctness clearly beats cost.
Practically: leave it at High for the bulk of your day and reach for Extra/Max on the hard tail. Spending Max effort on boilerplate is just burning your usage limit for no gain — the whole point of the dial is to spend more only where it pays off. If you’re new to the tool, my Claude Code getting-started guide covers the basics first.
What it costs, and where it runs
Opus 5 is $5 / $25 per million tokens (input / output) on the API. There’s a Fast mode — faster output at the same Opus intelligence, no downgrade to a smaller model — priced at roughly twice the base ($10/$50). On subscription plans it’s included: the new default on Claude Max, and the strongest model available on Claude Pro. It runs on the Claude API, Claude.ai, Claude Code, Claude Cowork, plus AWS, Google Cloud Vertex AI, and Microsoft Foundry.
For how per-token pricing turns into a real monthly bill — and how prompt caching changes the picture — I broke that down in how much it costs to add AI to your app and cutting token costs with prompt caching.
The quiet win: no mandatory data retention
Here’s a detail that didn’t make the headline but matters if you carry compliance obligations. When Fable 5 launched, it came bundled with a mandatory 30-day data-retention policy that applied even to customers with zero-retention agreements — a real sticking point I flagged at the time.
Opus 5 ships with no data-retention requirement for general access. If a retention clause was the reason you couldn’t route production traffic through the top tier, that specific blocker is gone here. For regulated workloads, that can matter more than a couple of benchmark points.
Safety: where Opus 5 still sits below Mythos 5
Anthropic is upfront that Opus 5 “remains behind Mythos 5 on cybersecurity tasks” and is “substantially behind” on exploit development — deliberately. At the same time, its cyber classifiers are “proportionally less restrictive than those on Fable 5.” The Opus 5 announcement doesn’t describe the Fable-style automatic reroute to a lower-tier model on flagged topics, so for everyday coding and security-adjacent debugging I’d expect fewer surprise handoffs than Fable 5 gave us — but treat that as my read, and test your own domain if you live in security tooling.
FAQ: quick answers for builders
What’s the model ID? claude-opus-5.
What does it cost? $5 / $25 per million input / output tokens. Fast mode is about 2x that.
Is it really as good as Fable 5? On many tasks, close. CursorBench 3.2 is within 0.5% of Fable 5’s peak at half the cost; OSWorld 2.0 surpasses Fable 5’s best at about a third of the cost. On the very hardest frontier tasks, Fable 5 / Mythos 5 can still lead — see my Opus 5 vs Fable 5 breakdown.
Is it the default in Claude Code? Yes — Opus 5 is the new default model there.
How do I control cost vs quality? Use /effort. High or below for routine work; Extra and Max for hard problems (they use more of your usage limit).
Does it force 30-day retention like Fable 5 did? No — general access has no data-retention requirement.
My bottom line
The interesting thing about Opus 5 isn’t a single benchmark — it’s the price line moving. Near-Fable-5 capability (CursorBench within 0.5%, OSWorld above it), Frontier-Bench more than doubled over Opus 4.8, genuine judgment and self-checking, hours-long autonomy that routes around blockers — all at half of Fable 5’s price, with no mandatory retention. That combination makes it my new default for most agentic work, and Anthropic clearly agrees, since it’s now the default in Claude Code.
Where I’d still think twice: the very hardest frontier tail, and security/bio-adjacent work where Mythos 5 is deliberately ahead. For everything else, the honest verdict is that the top-tier-quality tax just got cut in half. If you’re weighing whether to move a production build onto Opus 5 — or how to wire /effort so you’re not overpaying — that’s the kind of work I do.