The first thing I did when Claude Opus 5 landed was not read the benchmark chart. I opened a terminal and ran claude --version. It printed 2.1.219 — and that number is the entire difference between reading about this model and using it. Opus 5 requires Claude Code v2.1.219 or later, and older builds don't perform worse with it; they don't show it at all. No row in the /model picker, no error explaining why. The opus alias quietly keeps handing you Opus 4.8.
I got lucky — my install had auto-updated the night before. If yours hasn't, run claude update before any number below means anything to you. That's a strange thing to lead with on a model launch, but it frames my whole read of this release: the benchmarks are real, and the setup determines whether you ever see them. Most people forming an opinion of Opus 5 this month are actually evaluating their own configuration.
Short version for anyone who wants the answer up front: Opus 5 beats Fable 5 on agentic, long-running work at half the per-token price — and loses to it on the evals almost nobody is citing. The losses are where the useful information is.
What Shipped on July 24
Claude Opus 5 went live on Friday, July 24, 2026, as claude-opus-5 on the Claude API, in Claude Code and Claude.ai, and on the major cloud platforms. The specs that matter:
| Spec | Claude Opus 5 | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|
| Input / 1M tokens | $5 | $5 | $10 |
| Output / 1M tokens | $25 | $25 | $50 |
| Context window | 1M | 1M | 1M |
| Fast mode | $10 / $50 | — | — |
The row worth staring at: the price didn't move. Opus 5 costs exactly what Opus 4.8 cost — the second consecutive Opus release with a flat sticker. If 4.8 is your daily driver, switching changes nothing on your bill except token volume, which makes this an unusually clean upgrade decision. A fast mode at double the price exists for latency-sensitive work.
The Benchmark Table, Trimmed to What I Could Verify
Anthropic's launch table has a dozen rows. I'm reprinting only the figures I could cross-check across multiple independent sources, because launch-day numbers travel badly — more on that in a moment.
| Benchmark | Opus 5 | Fable 5 | GPT-5.6 Sol | Opus 4.8 |
|---|---|---|---|---|
| Frontier-Bench v0.1 (agentic coding) | 43.3% | 33.7% | 34.4% | 18.7% / 21.1% * |
| ARC-AGI-3 (novel reasoning) | 30.2% | not published | 7.8% | — |
| DeepSWE v1.1 (agentic coding) | 68.8% | 69.7% | 72.7% | — |
| Legal agent benchmark | 11.7% | 13.3% | — | — |
* Both figures circulate. See the next section.
Frontier-Bench is the headline for a reason. It's agentic terminal coding — the closest published proxy for what Claude Code does all day — and Opus 5 scores 43.3% against 33.7% for Fable 5, a model that costs twice as much. Beating your own flagship on the benchmark nearest your primary workload is the strongest single argument in this launch, and the one I find most credible because the task shape matches the use case.
ARC-AGI-3 is the eye-catcher at 30.2%, roughly four times GPT-5.6 Sol's 7.8% on an eval built specifically to resist memorization. But attach the caveat coverage keeps dropping: I could not find a published Fable 5 score on ARC-AGI-3. The generational-leap chart compares Opus 5 to its predecessor and to OpenAI — not to the model it's priced beneath. If you're choosing between Opus 5 and Fable 5, that row tells you nothing.
DeepSWE is the loss that matters. Opus 5 finishes third on an agentic coding benchmark — behind GPT-5.6 Sol by about four points and behind Fable 5 by about one — in the exact territory this release is marketed for. That row appears in almost no coverage, and it's the most important one in the table. The legal agent benchmark goes to Fable 5 too, 13.3% to 11.7% — and notice the absolute values: the winner is failing 87% of tasks. Ranking language hides how much runway is left in these domains.
Two more losses circulated at launch in health and exam-style evals, but the published numbers differ between sources depending on tool access, so I'm not reprinting them.
The Opus 4.8 Score Nobody Agrees On
On Frontier-Bench, Opus 4.8's baseline is reported as 18.7% by some outlets and 21.1% by others — a 2.4-point spread on the same model and benchmark, in coverage published within hours of each other, and I've seen nobody flag it.
The most likely explanation is effort levels. Claude models run at settings from low to max, and a model evaluated at two different effort levels is two different data points wearing one name. If 18.7% is 4.8 at one setting and 21.1% at another, both are correct and neither is comparable without the setting attached. The practical consequence: every headline comparison in this launch is a single point plucked off a curve. "More than double Opus 4.8" is 43.3 vs 18.7. "Roughly double" is 43.3 vs 21.1. Same launch, different story, depending on which rung of the effort ladder got screenshotted. I saw the same effect first-hand when I mapped Opus 4.8's effort levels on real Laravel work — the model's identity changes more between effort settings than most people assume it changes between versions.
"Half the Price" Is a Claim About Effort, Not Tokens
Opus 5 at $5/$25 is literally half of Fable 5's $10/$50, so the headline is accurate on its face. But you don't pay per-token price; you pay price times tokens consumed, and consumption is governed by effort. Anthropic's framing is that Opus 5 reaches near-Fable results at a lower point on its own effort curve — the saving comes from needing less compute to get there, not from a discount at identical settings.
That distinction bites the moment you set effort yourself. Run Opus 5 at max on everything and the token multiplier can erase the sticker advantage; find the level where quality holds and the saving compounds. The number that decides your bill is token count, not price per token — the same lesson from cutting Fable 5 usage costs, and the same reason I keep a cheap-model daily driver in the rotation for work that doesn't need a flagship.
The wins and losses form a clean pattern once you stop reading rankings and start reading task shapes. Opus 5 dominates long-running, many-step work in an environment — agentic coding sessions, workflow execution, novel problem-solving. It ties or loses on single-shot, deep-domain judgment — law, health, one-pass evaluations. Opus 5 is not "Fable 5 for half price." It's a model tuned for duration and iteration, where Fable is tuned for depth on a single hard thing. Pick by task shape, not chart position — the same split logic I applied when Sonnet 5 launched against Opus 4.8.
Switching in Claude Code: The Three Steps That Actually Matter
Step 1 — clear the version gate. claude --version must show 2.1.219 or later; below that, Opus 5 is invisible and no amount of typing summons it. claude update fixes it. The VS Code extension rides the same binary, so updating the CLI updates both.
Step 2 — select the model, explicitly. /model opens the picker; /model claude-opus-5 pins the exact model. I'd pin rather than trust the opus alias — what the alias resolves to depends on your version and platform, and an unattended update can silently change which model your team is running.
Step 3 — set effort deliberately. Per the Claude Code model-config docs, Opus 5 doesn't reset effort on first use: whatever level you last set — for another model, for an experiment you forgot — carries over. If a stale low from three weeks ago is riding along, you'll benchmark a frontier model at its floor and conclude the leap isn't real. Run /effort, set it consciously, and sweep downward from high settings to find where quality breaks for your work.
Then do the counterintuitive part: delete your verification prompts. Anthropic's migration guidance is explicit that Opus 5 verifies its own work unprompted, and instructions like "double-check your answer" or "add a final verification step" now cause over-verification — paying tokens for work the model already did. Two years of accumulated self-check scaffolding in your CLAUDE.md files just became overhead. The same guidance flags that Opus 5 expands scope and delegates to subagents more readily than 4.8, so add a scope-discipline line and consider capping subagent counts.
One more behavior to know before you switch: Opus 5 runs safety classifiers, and flagged requests fall back to Opus 4.8. The classifiers evaluate everything in context — files, connectors, search results — not just your message, so a repo full of security tooling can trip a fallback before you've typed anything unusual. If you do offensive-security-adjacent work, the model answering you will sometimes not be the model you selected. That's documented routing, not an account problem, but it also muddies how to read published scores on security-adjacent evals.
What I've Verified, and What I Haven't
Verified first-hand: the version gate, on my own machine. The /model and /effort mechanics and the migration guidance are from Anthropic's current docs, read rather than recalled. Cross-checked against multiple independent sources: the pricing, the launch date, and every benchmark figure in the table above — which is why the table is four rows instead of twelve.
Not verified: any benchmark number, by me, on my own code. Nobody publishing within days of a launch has reproduced the vendor's evals, and a published number is a claim, not a result. The open questions I'd want answered before drawing hard conclusions: which effort level produced which Opus 4.8 baseline, how flagged-and-rerouted requests are counted in security-adjacent scores, and whether the DeepSWE third-place finish reproduces on real repository work. I'm running my standard Laravel service-extraction task across effort levels this week and will publish what I find — not before.
Until then, the practical read: if you're on Opus 4.8, switch — same price, better agentic scores, no downside I can find. If you're on Fable 5, split your workload by task shape instead of migrating wholesale. And whatever you run, check your version, set your effort, and delete the verification prompts. The benchmarks aren't the hard part. The setup is.
Model selection, agent pipelines, and cost tuning judged against a real codebase instead of a launch chart — that is the AI systems work I do for clients. The scope, and how the decision gets made with receipts, is on my services.
