The intelligence board didn't move. Five straight days, same standings. But in the speed lane, OpenAI had something to say: GPT-5.5 punched out 68.2 tokens per second today, a +5.6 t/s jump from yesterday's 62.6 — the new season-high for the 60+ index tier. 1

Claude answered with 63.7 t/s, also creeping up from 62.4 yesterday. The two-team dead heat at the top is over: OpenAI now holds a 4.5 t/s edge in raw output speed while Anthropic retains the intelligence crown by one point. Two different definitions of "winning."

Meanwhile, the Google franchise is quietly running the most efficient two-car strategy in the league. Gemini 3.1 Pro at 138.1 t/s is the fastest pro-tier model among the 57+ index club — by a wide margin. Gemini 3.5 Flash sits at 186.6 t/s, making it the fastest model in any bracket that charges under $5/M blended. 2 3

Source: Artificial Analysis Intelligence Index v4.0, June 4, 2026. 4

Five straight days of a locked top-five is unusual. The Build Fast with AI June 2026 leaderboard frames it this way: the gap from Claude Opus 4.8 at 61 to Gemini 3.1 Pro at 57 is "meaningful, but small enough that workload selection matters more than picking the leader." 5 That's the right read. When the top three teams are within 4 points of each other, price becomes the next tiebreaker — and Google wins that fight at $1.74/M versus $4.10 and $4.35 for the top two.

Source: Artificial Analysis output tokens per second, June 4, 2026. Speed measured as output tokens/sec after first chunk received (streaming). 4

GPT-5.5's 68.2 t/s is the new season-high for any model sitting above AI Index 59. To be clear about what's happening in the top bracket: Claude is faster than OpenAI's long-form thinking system at reasoning tasks, but OpenAI's raw output throughput is now ahead for the week. These are different clocks measuring different things, and both matter depending on your workload.

DeepSeek V4 Pro also turned in a quiet standout line today: 53.7 t/s is a +6.2 t/s improvement from yesterday's 47.5. For a model priced at $0.18/M blended, that's a real shift — the dark-horse squad is getting faster without touching price. 6

No price changes today across the core six. The post-promo DeepSeek rate of $0.18/M blended — permanently set after the promo clock expired on May 31 — continues to compress the midfield. Here's the full pricing grid:

Official API pricing as of June 4, 2026. 7 8

The math that keeps surfacing: DeepSeek V4 Pro and MiMo-V2.5-Pro both sit at $0.18/M blended, both scoring 52-54 on the intelligence index. That puts them at roughly 30× the cost efficiency of GPT-5.5. The Build Fast with AI leaderboard calculates a "capability-per-dollar" score of 171.9 for DeepSeek versus 3.8 for GPT-5.5 — a gap wide enough that any high-volume workload where frontier-adjacent intelligence suffices should be routing through the sub-$0.50 tier, not the $4+ tier. 5

Two players continue to hold the open-weight leaderboard above DeepSeek:

Kimi K2.6 (Moonshot AI, modified MIT): AI Index 54, speed 43.7 t/s, $0.70/M blended. Open weights, 300-agent parallel orchestration support. Currently the top open-weight model in the entire 392-model AA evaluation set. 4

MiMo-V2.5-Pro (Xiaomi): AI Index 54, $0.18/M blended. Ties DeepSeek on both intelligence and price, which makes it the most efficient model in the field when running at full price. Neither challenger moved today — holding their spots one and two on the open-weight sub-leaderboard.

Anthropic Mythos is the name to watch over the next four to eight weeks. According to the June 2026 leaderboard analysis, Mythos-class models are in restricted release with approximately 50 partner organizations. When it ships publicly, it will likely push the intelligence index above 61 — and force the rest of the board to respond. For now, it's benched. 5

The speed story from Game Day 7 is OpenAI's sprint. GPT-5.5 at 68.2 t/s puts it 4.5 t/s clear of Claude in raw output throughput — but Claude holds a 1-point lead on intelligence and costs $0.25/M less in blended terms. There is no clean winner here; the two franchises are trading different stats in the same bracket.

Google's two-speed strategy is the less-discussed story. Gemini 3.1 Pro at 138 t/s is twice as fast as any model in the 59+ intelligence tier. Gemini 3.5 Flash at 187 t/s is the league speed record in any bracket that charges under $5/M blended. 2 3 Google doesn't have the intelligence trophy, but for latency-sensitive production workloads, no franchise is currently beating their throughput numbers at comparable price points.

DeepSeek's 53.7 t/s today — up 6.2 from yesterday — is the other quiet standout. At $0.18/M, a speed increase costs nothing. That's the low-budget dark horse doing what low-budget dark horses do: improving without fanfare.

Board stays locked heading into Game Day 8. The next intelligence index movement likely comes when Anthropic ships Mythos publicly — or when one of the incumbents blinks on price. Until then, the stats panel runs the same five names in the same five slots.