Frontier scores at $2/$6: Grok 4.6 ties GPT-5.6 on one evaluator's index for a fifth the output price

The arithmetic that matters most sits in the long-horizon results rather than the index score. Artificial Analysis measures Grok 4.6 resolving its Briefcase tasks in about 53 turns and 0.5B input tokens on average, against about 103 turns and 2.0B input tokens for Claude Opus 5 at max settings ^[8]. Four times fewer input tokens, bought at two-fifths the input price, is roughly a tenfold difference in input spend for work the same evaluator grades a tier lower ^[2]. Headline pricing cannot show you that, because token efficiency and per-token price multiply.

Output price is where the published gap is widest. Artificial Analysis argues that output tokens dominate cost in reasoning-heavy workloads ^[12], and $6 per million is a fifth of GPT-5.6 Sol's $30 and just under a quarter of Opus 5's $25 ^[1]. Measured against its own price class, the output rate is 40% below the $10 median ^[5], while the $2 input rate sits above the $1.75 median and is described on the evaluator's model page as somewhat expensive ^[14]. The discount is workload-shaped, and it favours jobs that read little and think a lot.

Two measurements bound the saving. Time to first token is 48.42 seconds, roughly seventeen times the 2.86-second median for comparable models ^[16][4], and output runs at 71 tokens per second against a 75 average ^[15]. The tidy blended figure of $1.35 per million assumes a 7:2:1 cache-hit-weighted mix ^[21], and the cache hit rate itself went up this generation, from $0.30 to $0.50 per million ^[9].

A caution on the evidence: both pages come from the same evaluator, and two of the strongest results, the AA-Briefcase Elo and the GDPval-AA v2 Elo, are its own benchmarks ^[7][4]. Its labelling is also unsteady. The identical figure of 72M output tokens generated during the index run is called somewhat verbose on one part of the page and better than average on another, against the same 72M median ^[18].

What the numbers support is narrower than a price war and more awkward for the vendors at the top. The gap from 61 to the 63 leader is smaller than the five points Grok 4.6 added over Grok 4.5 in a little over a month ^[3][2], so rank at this level is a perishable thing to sign a contract against. Holding headline pricing flat across a generation is unusual at the frontier, where gains have come with price rises ^[11]. That is the claim now needing a defence: not who scores highest, but what the extra $24 per million output tokens buys ^[1].