Five dollars and twenty five dollars. Anthropic shipped Claude Opus 5 on Friday at exactly what Opus 4.8 cost, to the cent, and claims it more than doubles the older model's benchmark score at a lower cost per task. The price card did not move. The bill for a finished job did.
Google's Gemini 3.6 Flash dropped output to $7.50 a million from $9.00 and says it needs 17% fewer output tokens to do the work, discounting the same job twice. AMD launched its Helios rack on Thursday advertising up to 30% more tokens per dollar than the rival cabinet, on its own lab's estimate. Nvidia's Vera Rubin went into production at five clouds, sold on tokens per megawatt. Not one of them is competing on dollars per hour.
Our basket agrees by standing still. All seven models we track came back at last week's price, and no neocloud list rate changed. The one number that moved was the B200 ceiling on the spot marketplace, $10.63 an hour down to $7.48. Which leaves an awkward column on the buyer's comparison sheet. If price per million tokens is the constant now, what is it still measuring?
All seven models in the token basket and all four neocloud GPU ranges came back unchanged from the 20 July snapshot; the only move was the marketplace B200 ceiling, $10.63 to $7.48 per GPU-hour.
