Grok 4.6 sells turn count, not benchmark score

xAI released Grok 4.6 for long-running agents, coding, research and what the announcement calls more ambitious interactive work, and Artificial Analysis scored it 61 overall, five points above Grok 4.5 and level with GPT-5.6 Sol ^[1][2][4]. The pitch has moved: the headline claim is not a higher score but a lower cost to finish a job, which is the only frontier metric that turns up on an invoice.

The supporting number is turn count. On AA-Briefcase, per Artificial Analysis, a Grok run reached what the index calls Fable 5-tier while averaging about 53 turns and 0.5B input tokens, against roughly 103 turns and 2.0B input tokens for Claude Opus 5 Max ^[9]. That is about half the turns and a quarter of the input context ^[1][2]. Artificial Analysis also measured $0.84 per task and put Grok 4.6 on its cost-performance frontier across every agentic evaluation in its index ^[7][8].

Those two figures do not sit together at list price. xAI's API pricing starts at $2 per million input tokens and $6 per million output ^[6], so 0.5B input tokens is roughly $1,000 of input alone, about 1,200 times the quoted $0.84 per task ^[3]. The source does not say which tasks or which price basis each number uses, so read "$0.84 per task" as an index-wide average, not the cost of a Briefcase run. If you are budgeting an agent that runs for hours, the per-run token curve is the number to ask for.

Distribution is broad on day one: Cursor, Grok Build, the API, OpenRouter, Vercel and Cloudflare, with 2x included usage in Cursor and Grok Build for the first week ^[3][4], plus Grok 4.6 on the $30/month SuperGrok plan ^[5]. Double usage with a one-week clock is a switching incentive, and the replacement is already pre-announced: Musk says Grok 4.7 is due in three to four weeks, is already significantly better than 4.6, and is being supplemented with a large amount of SpaceX company data ^[10][11]. Teams standardising an agent stack on 4.6 this week are building on a model whose vendor has given it a four-week shelf life. Worth noting that the roundup credits the release to "SpaceXAI" while attributing the pricing page to xAI; the entity naming is unresolved in the source ^[1][6].

The capital is not chasing the model layer. The only funding figure in the day's roundup went to the app layer, where Lovable raised $400M at a $13.3B valuation ^[12][5]; the other model-layer item was a personnel teaser about why Demis did not leave Google ^[13].

As for agents that survive long tasks, the honest datapoint is a user who gave Opus 5 access to Unreal Engine for 24 hours via AAABench and told it to build GTA 6, with the newsletter's own verdict landing at "GTA 4.5" ^[14][15]. Vendors concede the orchestration tax too: Anthropic's updated Claude Voice guide says Voice can use connected Gmail, Google Calendar, Google Docs and Slack, recommends splitting complex questions, and notes that several tools at once can add delay ^[16][17].

Watch whether Artificial Analysis publishes a per-run cost basis that reconciles $0.84 per task with 0.5B input tokens, whether 4.7 actually lands inside the stated three to four weeks, and whether the 2x Cursor and Grok Build allowance converts into a standing price cut once the promo week ends.