Opus 5 beats Fable 5 by 150 Elo, 20% cheaper

Opus 5 opened a nearly 150 Elo lead over Fable 5 on Artificial Analysis's benchmark, per Latent Space's July 25 AINews roundup. Fable 5 edges it out on a separate index.

That Elo gap came with a 20% cut in cost per task, per Artificial Analysis. For a team picking a default coding model, that's the pitch: less spend, no capability given up. Except the benchmarks don't agree on that last part.

On the Epoch Capabilities Index, Fable 5 leads Opus 5, 161 to 159. Narrow the index to SWE-ECI, the software engineering subset, and the two tie at 161 apiece, per Latent Space's breakdown. General capability tilts to Fable 5. Coding capability is a dead heat.

Then there's the anomaly. Users benchmarking Opus 5 on FrontierCode found medium reasoning effort scoring higher than high effort. That's not supposed to happen if the benchmark tracks a real capability curve. Latent Space flags it as evaluation instability rather than a genuine result, and says it deserves more attention than it's getting.

The two scoring systems don't agree on who's ahead. Fable 5 leads by two points on ECI. They're dead even on SWE-ECI. Opus 5 is well ahead on Elo and cost. That spread is the story, not either headline number. Until someone explains the FrontierCode inversion, the SWE-ECI tie is the number worth trusting if you're picking a model to write code.