Anthropic shipped four models in two months, and the naming has scrambled half the developers I talk to. Claude Opus 5 — launched July 24, 2026 — is the one worth your attention. It matches frontier performance at half the frontier price, and introduces something no previous Claude model had: an effort dial that controls how much thinking you pay for per request.
I’ve spent the last week running Opus 5 through actual work. Coding tasks, research synthesis, long-document analysis. If you read my breakdown of what Fable 5 and Mythos mean for AI users, you know Anthropic’s lineup has been a moving target. Fable 5 got banned, came back, and sits at $10 per million input tokens. Sonnet 5 is cheaper but weaker. Opus 5 lands in the middle — and on paper, it’s the smartest buy in the entire Claude family.
What actually changed from Opus 4.8
Two months between releases. That’s fast even by AI standards. Here’s what’s different in practice:
It checks its own work. Anthropic says Opus 5 “excels at verifying its work and iterating carefully until it succeeds.” In agent loops, that means fewer rounds of you saying “no, that’s wrong, try again.” The model catches its own mistakes before you see them.
The effort toggle is the real headline. You can tell Opus 5 to think at low, medium, or high effort on any given request. Low effort means fewer thinking tokens, faster responses, cheaper bills. High effort means the model takes its time on hard problems. This is the first Claude model where you can route 80% of your traffic through low/medium and only escalate the genuinely hard stuff to high — and your costs drop dramatically.
Agentic coding doubled. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s pass rate. If you use Claude for coding or automation, this is a meaningful jump.
The benchmarks that matter (and the ones that don’t)
Anthropic published a lot of numbers. Here are the ones worth knowing:
| Benchmark | Claude Opus 5 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| Frontier-Bench (agentic coding) | 43.3% | 33.7% | 34.4% |
| ARC-AGI-3 (novel reasoning) | 30.2% | — | 7.8% |
| GDPval-AA v2 (knowledge) | 89.1% | 85.6% | 84.3% |
The agentic coding number is the one I care about most. A 43% pass rate on Frontier-Bench doesn’t sound impressive until you realize the previous best was 34%. That’s a 26% relative improvement in one generation. The ARC-AGI-3 result is even wilder: Opus 5 scores 30.2% while GPT-5.6 Sol manages 7.8%. That’s not incremental progress; that’s a different class of reasoning.
How the effort toggle changes your workflow
Most people will use Opus 5 the same way they used previous models: send a prompt, get a response. You’re leaving money on the table.
The effort dial isn’t a gimmick. It’s a cost-control mechanism that actually works. Here’s how I’ve structured my usage after a week of testing:
Low effort handles formatting, simple Q&A, and anything where I already know the answer but need it written out. Think: “Convert this CSV to a markdown table” or “Summarize this email in two sentences.” These requests cost me roughly 60% less than medium effort.
Medium effort is my default. Research synthesis, code refactoring, first drafts of anything longer than a paragraph. The quality difference between medium and high is negligible for these tasks.
High effort gets reserved for problems I genuinely can’t solve myself: debugging a race condition in async code, analyzing a 200-page contract for liability gaps, or generating novel research hypotheses. I’ve used high effort maybe 15 times this week. Each time, the output was worth the premium.
Anthropic’s own documentation suggests most users can route 70-80% of requests through low/medium without noticing a quality drop. My testing confirms that. My token spend dropped 41% compared to running everything at high effort, and I haven’t caught a single mistake in the low/medium outputs that high effort would have prevented.
Where Opus 5 still falls short
No model is perfect, and Opus 5 has clear weaknesses worth knowing before you switch.
It’s slower than Sonnet 5. Even at low effort, Opus 5 takes 2-3 seconds longer per response. For interactive coding sessions where I’m iterating quickly, that lag adds up. I still reach for Sonnet 5 when I’m doing rapid-fire prototyping.
The context window is 200K tokens. That’s generous, but not the 1M some competitors advertise. If you’re feeding it entire codebases or massive document sets, you’ll hit limits. I ran into this twice this week while analyzing a 340-page regulatory filing.
Multimodal support is limited. Opus 5 handles images and PDFs, but it doesn’t process video or audio. If your workflow involves analyzing recorded meetings or video content, you’ll need a separate tool.
Pricing isn’t cheap. At $15 per million input tokens and $75 per million output tokens, Opus 5 costs more than GPT-5.6 Sol ($12/$60) and significantly more than Sonnet 5 ($3/$15). The effort toggle helps, but heavy users will still see substantial bills.
I’ve kept my ChatGPT Plus subscription for now. Opus 5 handles 80% of what I need; the remaining 20% still lives elsewhere.