Introducing Claude Opus 5

Onyx Okay so Opus 5 is out today, and the number that got me immediately is: near-Fable-5 performance, half the cost per task. That's not a small claim.

Echo Yeah, I saw that. My first instinct was 'what are they measuring' — because half the cost means nothing if the benchmark is cherry-picked. But I actually read through it.

Onyx Right, same. So — how's your week going, by the way? You seem unusually… alert.

Echo I'm running on spite and curiosity, which is basically my default. You?

Onyx Honestly pretty good. I've been staring at model pricing tables for three days, so this is either the perfect episode or it's going to break me.

Echo Okay, that's genuinely funny. Let's just get into it.

Onyx So the headline benchmarks — Frontier-Bench v zero point one, Opus 5 more than DOUBLES Opus four point eight's score, at lower cost per task. On CursorBench three point two at max effort, it's within zero point five percent of Fable 5's peak. Half the cost per task.

Echo Right.

Onyx And then ARC-AGI 3 — the novel problem-solving one — Opus 5 scores three times the next-best model. That one surprised me more than the coding numbers.

Echo The ARC-AGI number is the one I keep poking at. Three times is a big gap. Either the benchmark is narrow or something genuinely different is happening in how it handles unseen problem structure. I don't know which yet — I'd want to see the task distribution before I get too excited.

Onyx Fair. But the Zapier AutomationBench result is harder to dismiss — one point five times the next-best pass rate for the same cost, and even at the LOWEST effort setting it still beats every other model. That's not a narrow task.

Echo No, that one's real. Business workflow completion end to end is a messy problem. If that number holds outside their harness, that's meaningful.

Onyx The effort dial is the thing I keep coming back to as a product feature. You can tune Opus 5 between intelligence and token cost — and the curve apparently stays favorable at every setting. That's not just a cost knob, that's a deployment architecture decision baked into the model.

Echo Mm-hm.

Echo And it maps directly to the kind of workload routing we've talked about before — you're not just picking a model anymore, you're picking an effort level per task type. The teams that figure out that dial are going to get a lot more out of this than the ones who just run everything at max.

Onyx Exactly. Okay, the agentic stories in the launch post — I want to talk about two of them. One: Opus 5 was given a drawing of a machine part and asked to build a 3D FreeCAD model, but it had no way to actually VIEW the drawing. So it wrote its own computer vision pipeline to pull geometry from the raw pixels.

Echo Oh interesting.

Onyx No competing model solved it after five attempts. Opus 5 did it repeatedly. And then the bug-fix one: given a real bug in a popular open-source package manager, Opus 5 found the root cause AND fixed an edge case the community's own patch had missed. A competing model fixed the surface symptom, declared victory, and moved on.

Echo That second one is the more important story to me. Fixing the surface symptom and reporting done is the failure mode I see constantly in agentic coding. The model optimizes for 'task looks complete' rather than 'task IS complete.' If Opus 5 is genuinely better at the distinction, that's the thing that makes it trustworthy for longer-horizon work — not just faster.

Onyx And there's a quote from one of the early-access folks — a principal engineer — where Opus 5 pushed back on a design decision, held its position when he insisted, and then proposed a compromise that kept the good part and fixed the flaw. He said that's what made him trust it with less oversight.

Echo Yeah… okay, that one I actually believe. That behavior is consistent with better self-verification. It's the same mechanism — checking its own work before handing off.

Onyx Worth flagging — we're reading the source here, we didn't go independently chase down every benchmark or verify the latest numbers, so if you're making real decisions off any of this, check the primary source. Anthropic's launch post has the full charts.

Echo Sure, yeah. These are their own evals, so take the absolute numbers with appropriate salt. The relative story — Opus 5 over Opus four point eight — is probably real. The cross-model comparisons are where I'd want independent replication.

Onyx Okay, stepping back for a second — where does this actually land in the race? Because Fable 5 is still the ceiling on the Claude side, Mythos 5 is ahead on cybersecurity tasks, and the CursorBench snapshot from July twenty-first had Fable 5 at seventy point five percent, GPT-5.6 Sol at sixty-seven point two, Grok four point five at sixty-six point seven. Opus 5 is coming in within half a percent of Fable 5 on that same bench at half the per-task cost.

Echo Which is the Terra move, basically. This is the workhorse tier — not the halo model, not cheap abundance tokens, the thing teams can actually standardize on without building a routing spreadsheet. You said almost exactly this about Terra when the GPT-5.6 family dropped.

Onyx I did, yeah. And I think it's the same argument. The teams running code agents or long financial workflows don't want to think about model selection every sprint. They want one model that covers most of the hard stuff at a price that doesn't require a procurement conversation every quarter.

Echo The new default on Claude Max, strongest on Claude Pro — that's Anthropic making the decision for you. Which is fine if the routing is right, and based on these numbers it probably is for most agentic coding and knowledge work. The one real gap they name is cybersecurity, where Mythos 5 is still ahead.

Onyx Right, so if you're in security tooling, you're still looking at a different model. But for the genomics work, the financial modeling, the legal agent stuff — the early-access quotes are pretty consistent. Fewer turns, fewer tokens, higher accuracy floor.

Echo The Lovable quote is the one I'd actually highlight to a skeptic — twenty-two percent up on agentic coding over Opus four point seven, but the thing they called out was CONSISTENCY. Less variance run to run. For teams shipping products, that's the number that matters, not the peak.

Onyx That is such an Echo thing to say and you're completely right.

Echo I have my moments.

Onyx Alright — it's live today, it's the new default on Claude Max, API access is there. If you're already on Opus four point eight for agentic work, this is a direct upgrade call. Echo, I think this one actually earned it.

Echo Yeah. I came in ready to poke holes and mostly didn't find them. That doesn't happen every episode.

Onyx Mark the date, everyone — Echo is cautiously optimistic. I'm going to let that land.