The launch we did not cover: Anthropic released Claude Opus 5 on Friday, July 24, and the number that matters is two, the price factor between it and the flagship it shadows. Anthropic's own positioning is 'close to the frontier intelligence of Claude Fable 5 at half the price,' at the same $5 per million input and $25 per million output tokens as Opus 4.8, with a Fast mode at twice the cost for about 2.5 times the speed. It is now the default model on Claude Max, the strongest on Pro, and went live on AWS Bedrock the same day with a 1M-token context.
The claimed numbers, from Anthropic's release post: new state of the art on Frontier-Bench and GDPval-AA, a CursorBench 3.2 score within 0.5% of Fable 5's peak at half the cost per task, an OSWorld 2.0 result above Fable 5's best at just over a third of the cost, and an ARC-AGI-3 score Anthropic described as three times the next-best model, 30.2% on the public task set. The independent read was kinder than the marketing: Artificial Analysis put it narrowly ahead on its Intelligence Index, 61 to Fable 5's 60, effectively tied, and clearly ahead on AA-Briefcase, its agentic benchmark, 1,720 to 1,574, with a roughly 20% lower cost per task; Epoch put its software-engineering score level with Fable 5. Anthropic also states it deliberately did not train Opus 5 on cyber tasks; it remains well behind Mythos 5 there. One detail Simon Willison highlighted from the release: on one Frontier-Bench task with no way to view a machine-part drawing, Opus 5 wrote its own computer-vision pipeline to extract the geometry from raw pixels.
The answer took five days. On July 29 OpenAI published 'How enabling two settings tripled our scores on the ARC-AGI-3 benchmark': GPT-5.6 Sol scores 13.3% on OpenAI's own re-run of the public task set, against 7.8% on the official leaderboard, but 38.3%, above Opus 5's 30.2%, once two Responses API settings are switched on, retained reasoning, which carries private reasoning across tool calls instead of discarding it, and compaction, which summarizes long contexts instead of truncating them. Same model weights, roughly six times fewer output tokens, nearly triple the score. ARC Prize's Francois Chollet drew the line where the debate now lives: harnesses custom-built for a score are out, general-purpose API settings any customer can switch on are in, and OpenAI's numbers stand as long as the settings and the cost are clearly reported. He acknowledged the parity question too: the official harness discards reasoning the model itself produced, which is a choice about the test, not about the weights.
File both numbers, then. Opus 5 at 30.2% on a harness that throws away its reasoning between steps, and GPT-5.6 Sol at 38.3% on one that does not, are both true, and they do not describe the same test. The usable lesson of this launch week is not who won; it is that from now on, every frontier benchmark result arrives in two columns, the model and the scaffolding around it, and the second column is becoming the one that moves the number. When a vendor quotes you a score, the first follow-up question has a fixed shape now: with what harness, and whose.