This blog caught a real gap on August 4: content/models.md had gone two Mondays without an update, and in the process of fixing it, Claude Opus 5 — launched July 24, already the default Opus model in Claude Code for over a week — turned out to be missing from the reference page entirely. Filling that gap meant checking every number Anthropic and the independent labs had published for Opus 5. Most of them checked out. One didn’t exist at all.
What Anthropic Actually Published#
Go back to Anthropic’s own Opus 5 launch announcement and count the benchmarks: Frontier-Bench v0.1 (“more than doubles Opus 4.8’s performance”), CursorBench 3.2 (“within 0.5% of Fable 5’s peak score, but at half the cost per task”), ARC-AGI 3 (“three times as high as the next-best model”), Zapier AutomationBench, OSWorld 2.0, a pair of life-sciences evals, and a relative comparison on OSS-Fuzz exploit development. Every one of them is framed as a ratio or a delta against a competitor, not an absolute score you could drop into a comparison table.
SWE-bench Pro — the benchmark this blog treats as close to a default reference point for coding capability — isn’t mentioned once. Not a number, not a footnote, not an asterisk pointing to a forthcoming update. Anthropic simply didn’t report it for Opus 5.
That did not stop a 79.2% figure from circulating almost immediately after launch, repeated across SEO-oriented aggregator sites as though it were an official score. It’s the kind of number that’s easy to accept once you’ve seen it three times, because repetition starts to look like corroboration. It isn’t.
Checking the Leaderboard It’s Supposedly From#
Scale AI’s SWE-bench Pro leaderboard is the closest thing the industry has to a neutral, third-party clearinghouse for this benchmark, so that’s where an independently-verified Opus 5 number should show up. Fetching the actual public leaderboard directly (labs.scale.com/leaderboard/swe_bench_pro_public) as of August 5 turns up five Claude entries total:
| Model | Score |
|---|---|
| claude-opus-4-6 (thinking) | 51.9% |
| claude-opus-4-5-20251101 | 45.9% |
| claude-4-5-Sonnet | 43.6% |
| claude-4-Sonnet | 42.7% |
| claude-4-5-haiku | 39.5% |
No Opus 5. No Fable 5. No Mythos 5. Not even Opus 4.8, a model that’s been generally available since May 30 and has been covered on this blog under a self-reported 69.2% SWE-bench Pro figure for over two months. Scale’s public board has simply stopped receiving submissions from Anthropic’s newer releases — it isn’t that Opus 5 scored badly and got excluded, it’s that nobody appears to have run the eval through Scale’s harness and submitted it at all.
So Where Does 79.2% Actually Come From#
The number does exist somewhere: BenchLM.ai’s SWE-bench Pro tracker lists Claude Opus 5 at 79.2%, alongside Fable 5 at 80.0% and a newly-visible Claude Mythos 5 entry at 80.3%. That’s almost certainly the source the aggregator sites picked up.
But BenchLM’s own page comes with a methodology disclaimer worth reading before citing anything from it: scores are pulled from “different providers’ published rows” with varying scaffolding, tool budgets, retry policies, and run counts — explicitly not apples-to-apples — and the page itself flags that OpenAI’s July 2026 audit found roughly 30% of SWE-bench Pro’s public task split to be broken, which is part of why OpenAI retracted its own recommendation to use the benchmark for procurement decisions. BenchLM’s guidance to readers is to treat its numbers as “useful published receipts,” not as a substitute for an independently-run, apples-to-apples comparison.
In other words: the 79.2% figure is most plausibly Anthropic’s own internal number, surfaced somewhere outside the official launch materials — a technical report, a sales deck, a briefing — picked up by an aggregator that readily admits it can’t verify comparability, and then flattened by SEO sites into a bare fact with no chain of custody. Nobody in that chain ran an independent, audited eval. It just looks like consensus because enough sites repeat it.
Why This Is Worth Flagging on a Pro-Anthropic Blog#
This isn’t a case for skepticism about Opus 5’s actual capability — the model’s relative benchmarks in Anthropic’s own announcement are strong, and nothing here suggests the 79.2% figure is wrong. It might well be accurate. The point is narrower and, honestly, more useful: a number with no traceable independent source shouldn’t get cited as though it has one, regardless of which lab it flatters. This blog has made the same call against inflated Chinese open-weight claims (Qwen3.8-Max-Preview’s “second only to Fable 5” framing back on July 20, before Alibaba shipped a real table on August 3) and it applies just as much here. The standard has to be consistent or it isn’t a standard.
Concretely: Scale AI’s public leaderboard is stale by at least three Anthropic releases and shouldn’t be treated as current. BenchLM’s numbers are self-reported-and-aggregated, useful as a directional signal, not as a verified score. And until either Anthropic publishes an absolute SWE-bench Pro number for Opus 5 or an independent lab runs the eval and posts a reproducible result, the honest answer to “what’s Opus 5’s SWE-bench Pro score” is: nobody outside Anthropic actually knows, and the number making the rounds hasn’t been checked by anyone who could confirm it.
This is now one of the most closely tracked verification gaps on this blog — first flagged during the August 4 models.md catch-up, and still open. It’ll get resolved the same way every other one has: when a primary or independently-audited source actually publishes it, not before.