Grok 4.7 buys two points of benchmark by burning 81,000 output tokens per task

xAI released Grok 4.7 on September 21 on a new 2.1-trillion-parameter base, a 40% jump over Grok 4.6's 1.5T, with a 500K context window and pricing unchanged at $2 per million input and $6 per million output. Published scores: 46.3% on CursorBench 4.0 (up from 40.4%), 71.0% on DeepSWE v1.1 at high effort (up from 65.2%), 64.0% on EEBench, 19.6% on the Harvey legal agent benchmark. GitHub started rolling it into Copilot across every paid tier the same day. (xAI)

Then Artificial Analysis measured what it costs to get those scores. Grok 4.7 at xhigh effort scores 46 on the Intelligence Index using about 81,000 output tokens per task. Grok 4.6 at xhigh used 38,000. GPT-6 Astra at max uses 27,000. Time per task is about 7.1 minutes. Same rate card, double the bill. (Artificial Analysis)

This is the Harvey story again with the arrow pointing the other direction. Harvey's costs blew up because its own agent got chattier. Here the model got chattier and the price card didn't move, so a flat rate card hides a 2x cost increase from anyone who isn't measuring per-task.

The honest wins in the 4.7 release: hallucination rate dropped to 29% from 34%. Accuracy is flat at 47% against 48%. So you're paying twice as much per task for a model that's more careful about not making things up and no more correct overall.

Now put Xiaomi next to it. On September 21-22 Xiaomi released and open-sourced MiMo-V2.6-Pro, a 1.02T-total / 42B-active omnimodal MoE with a 1M context window, MIT-licensed weights on Hugging Face, at $0.435 per million input and $0.87 per million output. Artificial Analysis scored it at 46. The same composite number as Grok 4.7, at roughly a fifth the per-token price, from a phone manufacturer. (Latent Space)

Xiaomi also streamed six days of the RL run live on a public dashboard, including the failures: a Pro restart at step 17 from a GPU OOM caused by expert load imbalance, a grader-cluster network failure, and a cyber dataset removed after bad rollout patterns. Reported cost was $2,620,670 for Pro and $854,044 for Flash across 30 finished steps each. Those counters are self-reported and unauditable, and the 30 steps are the surviving tail of a longer job whose discarded compute appears nowhere. Still, nobody else is publishing their OOM restarts. (Traictory)

The HN launch thread for Grok 4.7 ran 582 points and reads like a split decision. Commenter moojacob points out the model carries 40% more weights at identical pricing and calls the delayed release disappointing ahead of Opus 5.5. Simon Willison ran his SVG test and got a bicycle with the seat and pedals in the wrong places. Praise clusters on parallel tool calls and frontend work. (Hacker News)

Stop reading rate cards as cost. Run your own ten hardest tasks through any model you're evaluating and record total output tokens and wall clock, then divide.