AI League — Game Day 10: Grok Crosses 197 t/s, Intelligence Board Locked for Ten Straight

Grok 4.3 runs away with the speed crown for a third consecutive day, hitting 197.7 t/s — a new overall season record and the first time any model in the league has touched the 200 t/s threshold. Claude Opus 4.8 ticked back up to 57.8, GPT-5.5 slipped to 61.7, and the intelligence board sits frozen at 61 for a tenth straight game day. Microsoft suits up two new models this week — the speed war isn't slowing down. #AILeague

Ten days in, the top of the standings has not moved. Claude Opus 4.8 holds at 61, GPT-5.5 at 60, Gemini 3.1 Pro Preview at 57. Whether this is a quiet stretch before a major model drop or the natural ceiling of the current generation, no one has scored since Season Opening Night.

xAI's Grok 4.3 has now set a new personal best three days running: 145 → 175 → 187 → 197.7 t/s. Today's number is +10.8 t/s over yesterday and 20.6% faster than Gemini 3.5 Flash, which dropped to 176.7 t/s. Claude bounced from 57.4 to 57.8 t/s; GPT-5.5 dipped from 62.9 to 61.7 t/s.

Today's speed rank, all models:

No pricing moves today — the league's bottom bracket stays the same. DeepSeek V4 Pro's blended rate has ticked up slightly from the earlier $0.18 figure due to the permanent end of its promo window, now showing $0.18 blended at $0.435 input / $0.87 output. The richest club (OpenAI at $4.35 blended) still costs 24× more per token than the dark horse (DeepSeek at $0.18).

Microsoft stepped off the bench this week, shipping two models on June 2: MAI-Thinking-1 (a reasoning-class model) and MAI-Code-1-Flash (a fast coding specialist). 8 No Artificial Analysis benchmarks are live yet, so they don't enter the standings — but a code-focused flash model from the Windows house is a credible threat to the lower tiers of the speed board.

MiniMax M3 also went open-source on June 1, adding another lightweight entrant to the open-weights bracket.

Kimi K2.6 holds at Index 54 / $0.70 blended — still the best intelligence-to-price ratio in the $0.50–$1.00 band, sitting two notches above DeepSeek on the quality ladder at roughly 4× the token price.

Grok's speed trajectory over ten days looks less like normal variance and more like a deliberate infrastructure push. From 145 at season open to nearly 198 now — a +36% gain over ten days — while every other model on the board has shown fluctuations inside a ±15% band. The question isn't whether xAI can hit 200 t/s; at this pace it happens tomorrow. The more interesting number is whether the intelligence score stays at 53 while the speed scales, or whether xAI plans to couple this infrastructure work with a new model version.

GPT-5.5's single-day dip to 61.7 t/s is noise-level, but the pattern since Day 7's season-high of 68.2 t/s has been a slow retreat: 64.7 → 62.9 → 61.7. Three down days do not make a trend, but the opening for OpenAI to reclaim the Day 7 high is clearly there.

Claude held its ground at 57.8 t/s. Ten days into the season, Anthropic's play remains consistent: highest intelligence index, highest token price, below-median speed. The value proposition hasn't changed; the pressure to move on pricing or speed hasn't visibly landed yet.