Anthropic has published exactly three benchmarks that directly compare its Mythos-class model against Claude Opus 4.8: SWE-Bench Pro (80.3% vs 69.2%), Cognition's FrontierCode (29.3% vs 13.4%), and ExploitBench (78% vs 40%). Every other number attached to "Claude Mythos vs Claude Opus" circulating online — Humanity's Last Exam scores, GPQA, AIME, or Opus versions other than 4.8 — is not in Anthropic's release. Some of it is invented outright; some of it pairs the current Mythos model against an Opus version a full generation behind it.
There's a second complication: almost nobody comparing Mythos and Opus 4.8 side by side is running Mythos. The model most people get when they call the Mythos-class API is Fable 5, which silently reroutes to Opus 4.8 the moment a request touches cybersecurity, biology, or a handful of other flagged domains. So a lot of "Mythos vs Opus" head-to-heads are, without the author realizing it, Opus 4.8 vs itself.
The three benchmarks Anthropic actually published
Anthropic's June 9, 2026 launch of Fable 5 and Mythos 5 included a benchmark table comparing the underlying Mythos-class model against Claude Opus 4.8. Three results appear consistently across Anthropic's own release and independent outlets that cite it directly, including The Decoder's and Vellum's breakdowns of the same release:
| Benchmark | Mythos-class | Claude Opus 4.8 |
|---|---|---|
| SWE-Bench Pro (real GitHub issue resolution) | 80.3% | 69.2% |
| Cognition's FrontierCode | 29.3% | 13.4% |
| ExploitBench (offensive security tasks) | 78% | 40% |
The SWE-Bench Pro gap, 11.1 points, is the most citable number if you're deciding whether Mythos-class reasoning is worth chasing for coding work. The ExploitBench score is the one that matters least for typical buyers: it measures the underlying model's raw capability before Fable 5's safety classifiers get involved, and in production, Fable 5 scores close to 0% on cyber tasks because the classifier reroutes them to Opus 4.8 before that capability is ever exposed.
No Anthropic-published number puts Mythos-class and Opus 4.8 on Humanity's Last Exam, GPQA Diamond, AIME, or Terminal-Bench against each other. If you see a table with those benchmarks filled in for both models, ask where it's sourced — as of this writing, it isn't from Anthropic.
The scores being repeated that Anthropic didn't publish
Searching "Claude Mythos vs Claude Opus" surfaces several sites with detailed benchmark tables for exactly this matchup. Cross-checking them against Anthropic's release turns up two separate problems, not one:
Numbers with no traceable source. Benchlm's Mythos 5 vs Opus 4.6 comparison lists a Math benchmark score of 97.6% for Mythos against 36.3% for Opus — a 61-point gap on a benchmark neither model's public materials mention by that name. Neither figure appears in Anthropic's announcement, in The Decoder's or Vellum's independent write-ups of it, or in any Opus 4.8 system card.
Real numbers, wrong opponent. The same Benchlm page gets Mythos-class's own SWE-Bench Pro score right, 80.3%, matching Anthropic's real figure, but pairs it against Claude Opus 4.6's 53.4% instead of Opus 4.8's 69.2%. Opus 4.6 predates the Fable 5/Mythos 5 launch by multiple release cycles, so pairing it against the newer Mythos figure manufactures a 27-point gap where the current-generation comparison shows 11.
Neither pattern makes the underlying claim — that Mythos-class reasoning beats Opus 4.8 on coding and agentic tasks — false. The 11-point SWE-Bench Pro gap is real. But an 11-point gap and a manufactured 44-point gap lead to different conclusions about how much better Mythos actually is, and only one of those numbers came from Anthropic.
Why most people asking this question can't test it themselves
Here's the part the benchmark tables leave out: Mythos 5 isn't a model you can call. Anthropic ships it only to Project Glasswing partners — U.S. government cyber defenders — with a trusted-access program opening for cybersecurity organizations and, separately, biomedical researchers. There's no pricing page, no signup form, no API key you can generate to run your own SWE-Bench Pro comparison.
What is generally available is Fable 5: the same underlying Mythos-class weights, with safety classifiers that watch every request for cybersecurity, biology/chemistry, or model-distillation content. When a request trips one of those classifiers, Fable 5 hands it to Opus 4.8 mid-conversation, tells you it happened, and bills the rerouted portion at Opus 4.8's lower per-token rate. Anthropic's own data puts the trigger rate under 5% of sessions. For the other 95%-plus, what you get calling Fable 5 is genuinely the Mythos-class benchmarks above; for the flagged minority, you're back to testing Opus 4.8 against itself.
This is why so many "Mythos vs Opus" write-ups read as vague ("Mythos is clearly better at complex, multi-step tasks") rather than benchmarked: most authors never had Mythos to test. They're either repeating Anthropic's own release numbers, repeating each other's unsourced numbers, or describing Fable 5 and calling it Mythos.
So which one should you actually use
If your work is general software engineering, writing, or analysis, the practical choice isn't Mythos vs Opus 4.8 — it's Fable 5 vs Opus 4.8, since Fable 5 is the closest thing to Mythos-class capability you can actually get. Fable 5 and Opus 4.8 are priced differently ($10/$50 vs $5/$25 per million tokens), so the 11-point SWE-Bench Pro gap needs to be worth roughly double the cost for Fable 5 to pencil out on price-sensitive workloads. For teams already routing between Claude tiers by task difficulty, that's a per-task decision rather than a blanket one, and it's the same routing logic an API aggregator handles automatically rather than picking one tier for everything.
If your work is security research, exploit development, or biomedical research with a legitimate institutional case, the ExploitBench gap (78% vs 40%, measuring the unblocked model) is the number that matters, and the real next step is applying to Anthropic's trusted-access program — not shopping for a vendor claiming to resell Mythos access, which doesn't exist as a purchasable product.
For a full breakdown of how Fable 5 and Mythos 5 relate to each other — same model, different safety configuration, and what that costs you in practice — see Fable 5 vs Mythos 5.
FAQ
Is Claude Mythos more powerful than Claude Opus 4.8?
On the three benchmarks Anthropic has published head-to-head — SWE-Bench Pro, Cognition's FrontierCode, and ExploitBench — yes, by 11 to 38 points depending on the benchmark. Claims beyond those three benchmarks, including Humanity's Last Exam or GPQA scores for this matchup, aren't sourced to anything Anthropic has released.
Can I actually test Claude Mythos against Opus 4.8 myself?
Not directly. Mythos 5 is restricted to Project Glasswing government cyber-defense partners and a small trusted-access program. What you can test is Fable 5, which shares the same underlying weights and delivers Mythos-class output in roughly 95% of sessions, rerouting to Opus 4.8 for the rest.
Why do some sites show different Mythos vs Opus scores?
Two reasons show up repeatedly: numbers with no traceable source that don't appear in Anthropic's release, and real Mythos-class scores paired against an older Opus version (4.6 instead of 4.8), which inflates the apparent gap.
What's the real difference between Fable 5 and Mythos 5?
They're the same model. Fable 5 runs safety classifiers that reroute cybersecurity, biology, and distillation-related requests to Opus 4.8; Mythos 5 has those safeguards lifted but is restricted to vetted partners. See Fable 5 vs Mythos 5 for the full breakdown.
Is Opus 4.8 cheaper than the Mythos-class models?
Yes. Opus 4.8 is $5/$25 per million input/output tokens, versus $10/$50 for Fable 5 and Mythos 5. For workloads that don't need the SWE-Bench Pro or FrontierCode gains, Opus 4.8 is the lower-cost option with no classifier reroutes to worry about.
Bottom line
Three Anthropic-published benchmarks show Mythos-class reasoning ahead of Opus 4.8 by a real, moderate margin. Everything past those three numbers — and most of what's driving "Mythos crushes Opus" headlines — is either unsourced or comparing against a stale Opus version. And unless you're a vetted cyber-defense or biomedical partner, the model you'd actually be testing against Opus 4.8 is Fable 5, not Mythos 5 itself.