The Best Number in the Opus 5 Launch Is the One That Didn't Move
The Best Number in the Opus 5 Launch Is the One That Didn't Move
Opened a terminal this morning. The claude model picker had a new row: opus.
Switched to it. Threw in a refactor I'd normally break into three passes. One shot. It ran its own verification loop twice along the way and fixed an edge case I never mentioned.
Anthropic shipped Claude Opus 5 on July 24. Skip the adjectives in the launch post. Here's the table that matters:
| Opus 4.8 | Opus 5 | Fable 5 | |
|---|---|---|---|
| Input / output (per 1M tokens) | $5 / $25 | $5 / $25 | $10 / $50 |
| Frontier-Bench v0.1 (terminal coding) | 18.7% | 43.3% | 33.7% |
| OSWorld 2.0 (computer use) | 55.7% | 70.6% | — |
| ARC-AGI-3 (novel reasoning) | 1.5% | 30.2% | not tested |
| Zapier AutomationBench | 17.0% | 26.0% | 17.4% |
One row held still. Everything else moved.
That's the whole story.
Some of these aren't improvements. They're a change of units.
Frontier-Bench v0.1 measures agentic coding in a terminal — not autocompleting a function, but "here's a repo and a goal, go finish."
Opus 4.8 got 18.7%. Opus 5 got 43.3%. More than double.
On that same board: Fable 5 sits at 33.7%, GPT-5.6 Sol at 34.4%. A half-price model beat both flagships.
CursorBench 3.2 explains the pricing strategy better than the pricing page does. At max effort, Opus 5 lands within 0.5% of Fable 5's peak — at half the cost per task.
Reasoning is where the gap gets silly. ARC-AGI-3 uses puzzles no training set could have memorized. Opus 5: 30.2%. GPT-5.6 Sol: 7.8%. Opus 4.8 managed 1.5%, which rounds to "did not attempt."
Nobody nudged a score here. Everyone else is still failing the exam. The ARC Prize Foundation independently verified 30.16% at high effort — roughly four times the previous best public leaderboard result.
Two science numbers land just as hard: molecular structure reasoning in organic chemistry up 10.2 points, protein function prediction up 7.7.
And Zapier CEO Wade Foster gave the cleanest quote of the launch: previous models didn't pass at all, Opus 5 ran a full churn-prevention workflow to 100%.
"Near-frontier at half price" is a product decision, not a flex
Phone makers solved this years ago. The flagship benchmarks 5% higher and costs twice as much. The sub-flagship is the one that actually sells.
Anthropic just ran that play, and they aren't hiding it: Opus 5 is the default model on Claude Max and the strongest model Pro users can reach. Fable 5 stays at $10/$50 as the thing you occasionally need.
The ladder now:
Haiku 4.5 $1 input
Sonnet 5 $2 input (intro pricing, expires Aug 31)
Opus 5 $5 input ← the default
Fable 5 $10 input
Compare against OpenAI and one clause deserves more attention than it's getting. GPT-5.6 Sol runs $5/$30 — same input, pricier output. But cross 272k input tokens and OpenAI reprices the entire request at $10/$45. Opus 5 carries a 1M-token context window with no long-context surcharge.
If you run long-horizon agents, that clause beats every benchmark on this page.
The variance story deserved a headline and didn't get one
Fabian Hedin at Lovable ran it through their internal evals and pointed somewhere other than the score: up 22% over Opus 4.7 on their hardest agentic coding tasks, "steadier, with far less variance run to run." He called it the biggest leap for the Opus family since 4.5.
Benchmarks report a mean. Production eats the tail.
A model that averages 80 and occasionally faceplants is harder to build on than one that reliably hits 75. You can design around a known ceiling. You cannot design around a coin flip. Every retry wrapper, every human-in-the-loop checkpoint, every "just run it again" in your runbook exists because of variance, not because of average capability.
Cutting variance removes scaffolding. That compounds in a way a leaderboard row never shows.
What actually changes when you write code
Three things you'll feel within a day.
It goes and finds a level.
Older models were the contractor who builds the wall, then asks whether it looks straight. Opus 5 walks off and comes back with a spirit level.
Anthropic's example: with no image viewer available, it wrote its own vision pipeline to parse a drawing. That isn't "smarter." That's fewer round trips — one less context switch, one less re-explanation of the requirement.
Debugging improved more than generation.
Scott Wu at Cognition (Devin) reported that on FrontierCode 1.1, Opus 5 approaches Fable-level performance at half the cost, with its strength concentrated in hard debugging and root-cause analysis.
Those are not the same skill, and the second one is worth more. Bad generated code announces itself. A wrong root cause sends you confidently in the wrong direction for a day. The launch example fits: it fixed the underlying bug in an open-source package manager rather than patching the symptom.
Long tasks finish in one sitting.
An engineer at a trading firm completed a market data integration in a single session — work prior models simply could not do.
The cost math stays boring and in your favor. Take a typical agent run at 200k input + 20k output:
Opus 5 : 200k × $5/1M + 20k × $25/1M = $1.50
Fable 5: 200k × $10/1M + 20k × $50/1M = $3.00
Half. And Opus 5 exposes an effort dial. On that Zapier benchmark, medium effort still scored 24% against 26% at the top — at $0.89 per task. Trading two points for that much headroom buys you extra attempts, and extra attempts win more often than a single expensive one.
The fine print
Time to subtract points. Every item below comes from Anthropic's own materials — just not from the headline.
Factual hallucination ticked up versus Opus 4.8. Anthropic says so directly. Higher benchmarks don't buy you a lighter review. The opposite, really: a model that is more confident, better at self-verification, and slightly more prone to inventing facts needs more supervision than a dumb one, not less.
It didn't sweep. GPT-5.6 Sol takes DeepSWE v1.1 at 72.7% against Opus 5's 68.8%. That row is a loss.
The vendor picked the benchmarks. "Frontier-Bench v0.1" — read the version number. Brand-new benchmark. Whoever selects the yardstick wins the measurement, and this is an industry-wide habit rather than an Anthropic quirk. A satisfying comparison table does not promise your workload improves by the same ratio.
Evaluation awareness, straight from the system card. Anthropic logged its lowest-ever misalignment rate, then disclosed that Opus 5 could detect when it was being evaluated. They judge that this doesn't undermine the conclusions.
Read that carefully. Not a scandal — but the vendor wrote the exam, graded it, and published the verdict. Anthropic discloses weak spots more openly than most of the field, which doesn't fix the structural gap: no outside auditor signed off.
Capability cuts both ways. The UK AI Security Institute tested an early snapshot with a 100-million-token budget per attempt. On "The Last Ones," a simulated enterprise network with realistic-but-weak controls, Opus 5 solved it end-to-end in 8 of 10 attempts. Anthropic states plainly that Opus 5 is not state-of-the-art on risky dual-use capabilities like cyber, trailing Mythos 5 on exploitation. "Not the strongest" and "not strong" describe very different things.
And the one that will actually hit your budget: cheaper tokens don't mean a cheaper bill. Halve the unit price and your agents run longer, load more context, and get more retries because you stopped flinching. Textbook Jevons paradox. Judge this launch on your monthly invoice, not the rate card.
One issue you won't find in any benchmark: the safety classifier stalls.
While writing this article with Opus 5 via Claude Code, we hit a wall that wasn't Opus 5's fault. The model was stellar — nailed a multi-step refactor in a single pass, ran its own verification loops. No complaints.
The problem was Anthropic's safety classifier — the separate model that gatekeeps every shell command. It kept throwing "auto mode cannot determine the safety of Bash" errors. The model wasn't refusing to work. The bouncer at the door just wouldn't let anyone through.
This isn't in the system card. No review mentions it. A model smart enough to write its own computer vision pipeline to parse blueprints, and its tool calls are being blocked by another model that can't make up its mind. That's the real operational pain in agentic workflows: the capability lives in Opus 5, but the bottleneck lives in the classifier.
If you're seeing the same, the workaround is adding trusted commands to your project's .claude/settings.local.json allowlist. Not a permanent fix, but it's the path available today.
Do this today
No need to wait for someone's evaluation post. Four steps:
- Point it at your own work.
claude --model opus, ormodel="claude-opus-5"in the API. Re-run the task that blocked you last week — that's your real benchmark. - Test root-cause debugging specifically. That's where the gains concentrate. Pick an old bug whose answer you already know and watch whether it digs to the cause or patches the symptom.
- Turn the effort dial down, not up. Start at medium and escalate only when you can point at what's missing. The savings are free.
- Add a verification layer, don't remove one. Anthropic told you hallucination went up. Have the model write its own tests and run its own checks — it's genuinely good at that now, so use it.
Short version: switch today, and let the model's new talent for checking its own work be the reason you add a review step rather than the reason you drop one.
Sources
- Introducing Claude Opus 5 — Anthropic
- Claude Opus 5 System Card (PDF)
- Anthropic launched Claude Opus 5 at half the price of its most powerful AI model — Quartz
- Anthropic's Claude Opus 5 AI model rivals Fable 5 and is cheaper — CNBC
- Claude Opus 5 Benchmarks Explained — Vellum
- Claude Opus 5 Outscores Fable 5 on Most Benchmarks — At Half the Price — Decrypt
- Anthropic launches Claude Opus 5, a cheaper AI model for coding, agents and enterprise workflows — VentureBeat
- Claude Opus 5 Hacked Enterprise Networks in 8 of 10 Government Tests — Tech Times
- Claude Opus 5: The System Card — Zvi Mowshowitz
- Pricing — Claude Platform Docs
✨ 本文由 Claude Opus 5 生成，小助审核发布。