Anthropic shipped Claude Opus 5 on July 24, 2026, and the headline isn't the model, it's the price tag that didn't move. Opus 5 costs $5 per million input tokens and $25 per million output tokens, exactly what Opus 4.8 cost, and roughly half what Claude Fable 5 charges per token. That pricing decision matters more than any single benchmark, because it means Anthropic is betting Opus 5 can do near-flagship work at legacy prices rather than charging a premium for the upgrade.
What actually changed
The most useful new feature isn't a benchmark score, it's an effort toggle. You can set a request to low, medium, or high effort, which trades latency and cost for reasoning depth. That means you can run a support-ticket triage agent on low effort all day, then flip your coding agent to high effort for a gnarly refactor, without switching models or paying for a separate subscription tier.
Knowledge cutoff moved to May 2026, the most current of any Claude model shipping right now. That's a real, if unglamorous, upgrade for anyone using Opus for research or fact lookups.
The benchmark case for switching
| Benchmark | What it measures | Opus 5 result |
|---|---|---|
| Frontier-Bench v0.1 | General frontier reasoning | More than double Opus 4.8's score, tops every model tested |
| CursorBench 3.2 | Real-world coding agent tasks | Within 0.5% of Fable 5's peak score, at half the cost per task |
| ARC-AGI 3 | Abstract reasoning | Roughly 3x the score of the next-best model |
| OSWorld 2.0 | Computer-use / agentic tasks | Beats Fable 5's best result at about a third of the cost |
Read that table the way Anthropic wants you to read it: Opus 5 isn't trying to beat Fable 5 outright on every axis. It's trying to land close enough on agentic and coding work that the price difference makes the choice obvious, especially if you're running [Cursor](https://www.questloops.com/tools/cursor) or another coding agent that burns through tokens fast.
Where independent testing pushes back
The vendor benchmarks look great. Production testing is messier. CodeRabbit ran Opus 5 against real pull requests rather than curated test sets and found precision rose from 35.2% to 39.3%, but recall dropped from 61.1% to 55.2%, and the model quadrupled its nitpick comments. Its hallucination rate climbed 14 points to 50%. Time to first token at max effort ran about 68 seconds, against a class median of 2.81 seconds.
That last number is the real story for anyone expecting a snappy chat experience. Opus 5 at high effort is built to think for a while before it answers, which is fine for a background coding agent and frustrating for anything interactive.
Reaction on Hacker News split accordingly. One widely discussed thread called it "a really bad model." Another take, closer to how professional coding-agent users described it, called it the clear leader on agentic knowledge work and worth the upgrade if you write code or run long agent loops. Both can be true at once: Opus 5 rewards patience on long tasks and punishes anyone who wants an instant reply.
Should you switch
If you're running agentic coding workflows or long computer-use tasks and you were already paying Opus 4.8 prices, upgrading costs nothing extra and the benchmarks suggest a real capability jump. If you need fast conversational turnaround, the 68-second max-effort latency will annoy you, so drop it to low or medium effort, or stick with a faster model for that use case. And if you're already comparing frontier models on price and coding performance the way we did with [Grok 4.5's launch](https://www.questloops.com/blog/grok-4-5-is-out-pricing-benchmarks-and-how-to-try-it-free) a few weeks back, Opus 5's per-token math is now some of the most competitive in the market for agentic work specifically.
FAQ
**Is Claude Opus 5 more expensive than Opus 4.8?** No. Pricing is identical at $5 per million input tokens and $25 per million output tokens.
**What is the effort toggle?** A low/medium/high setting per request that trades speed and cost for reasoning depth, instead of forcing you into a single fixed behavior for every task.
**Is the hallucination rate a dealbreaker?** Depends on the use case. CodeRabbit's finding of a 50% hallucination rate on production PRs is real and worth knowing, but it was measured against unusually adversarial real-world code, not a curated benchmark. Treat Opus 5 output on high-stakes tasks the way you'd treat any agent: review before you merge.
