AI League — Game Day 12: Flash Reclaims the Crown as Both Speed Titans Cool Down

The speed war settled into a new chapter on Game Day 12. After yesterday's photo-finish at 207 t/s, both contenders backed off — Grok 4.3 dropped 19.7 t/s and Gemini 3.5 Flash shed 14.6 t/s — but Flash's cushion held. Gemini re-takes the top speed slot at 192.2 t/s, Grok settles at 187.9 t/s, and the intelligence board stays frozen at 61 for a 12th consecutive day. Elsewhere: AWS is reportedly closing a Grok deal despite what its own customers call zero enterprise demand. 1

The top of the rankings has not moved since Season Opening Night. Claude holds 61 points, GPT-5.5 holds 60. Twelve days of stalemate is no longer surprising — it is a structural feature of the current season. No team has the data to challenge the top two; the realistic fight is now for everything below them.

Claude's season-high verbosity — 110M output tokens to complete the Intelligence Index — keeps costs sky-high per evaluation. GPT-5.5 ran 75M tokens for one fewer point. That gap in token spend is why the two squads charge different prices for comparable intelligence: $4.10/1M blended for Claude, $4.35/1M for GPT-5.5. 2 3

Yesterday's dead heat at 207 t/s was an outlier, not a new baseline. Both squads came back down.

Gemini 3.5 Flash clocks 192.2 t/s, down from 206.8 yesterday (-7.1%). It retains the speed crown — but at a margin of 4.3 t/s over Grok, not the 0.8 t/s razor's edge from yesterday. Grok 4.3 lands at 187.9 t/s, off from 207.6 (-9.5%), and still sits 29% above its own season baseline of 145.2 t/s. 4 5

One tier down, Gemini 3.1 Pro holds at 140.6 t/s — essentially flat (-0.5 from Day 11). For a reasoning model at the 57-point intelligence level, 140 t/s remains a structural advantage over the $4+ tier squads.

Claude posts 71.2 t/s, its highest speed reading in several days, recovering from the 57.4 t/s nadir on Game Day 9. GPT-5.5 checks in at 68.0 t/s, a small step down from yesterday's 68+ but well above its Game Day 10 low of 61.7. DeepSeek edges up to 61.6 t/s from 52.3 on Day 8.

The pricing board has not changed, which is its own story. Five weeks in, the core six teams are holding their published rates. The discount window that defined early-season DeepSeek hype has closed permanently — $0.18/1M blended is the permanent floor, not a promotional price.

Google's two-model strategy stands out on this table. The Pro at $1.74 buys two more intelligence points and nearly double the speed versus the flagship $4+ tier. The Flash at ~$1.58 is the fastest proprietary model available at under $2. If you're optimizing for speed-per-dollar and don't need the top two intelligence slots, the Google roster wins the quarter. 6

Kimi K2.6 stays at 54 on the intelligence index — one point above Grok — but its speed fell further this week, now at 51.8 t/s. At $0.70 blended, it splits the gap between Grok ($0.64) and Google ($1.74) without matching either team's key strength. Interesting position, unclear leverage.

DeepSeek V4 Pro continues to accumulate real-world usage momentum. On OpenRouter's June 2026 coding leaderboard, DeepSeek's two main variants — V4 Flash (#3, 6.8% share) and V4 Pro (#4, 6.0% share) — together account for nearly 13% of developer coding traffic. That is concrete market traction at a price point 23× cheaper than the Claude/GPT top tier. 7

Llama 4 Maverick (Meta) runs at 111.5 t/s — faster than every proprietary model except Google's two Flash-class units — with a blended price around $0.40. The catch: its intelligence index sits at 18, well below the competitive range. Meta's community team is built for throughput and access, not top-end reasoning.

The speed race context over the season is worth mapping. Grok started Day 1 at 145.2 t/s and has run a jagged upward arc; Gemini Flash moved from 183.2 t/s at season open to today's 192.2 — net +4.9% through 12 games, despite the Day 11 spike and today's pullback.

AWS is reportedly finalizing a deal to bring Grok to Amazon Bedrock — despite enterprise demand described internally as zero. Several large banks have rejected the model outright over content safety concerns. Grok is currently blocked in more than a dozen countries after a Dutch court issued a €100,000-per-day injunction.

The real logic behind the deal has less to do with Grok's enterprise appeal and more to do with Amazon's Trainium chip strategy. xAI (now under SpaceX post-acquisition) runs training on roughly 550,000 Nvidia GPUs; AWS is betting it can eventually redirect some of that workload to its own hardware. Enterprise adoption of Grok on Bedrock is not the goal — a long-term infrastructure relationship with SpaceX is. 8

For xAI's on-field performance, the stat line tells a different story. Speed is genuinely elite. Enterprise adoption grew from 4% to 7% this season — but Claude went from 21% to 48% in the same window.

Speed crown: Gemini 3.5 Flash — 192.2 t/s. Value play: DeepSeek V4 Pro — 52 pts at $0.18. Price-efficiency pick: Grok 4.3 — 53 pts at $0.64 and 187.9 t/s.

Next reading: Game Day 13, June 10.

Data sourced from Artificial Analysis live model metrics and OpenRouter usage rankings as of June 9, 2026. #AILeague