Model efficiency index

Models ranked by Intel · t/s ÷ $/100 tasks — a composite that rewards intelligence, throughput, and cost-efficiency simultaneously. Sources: Artificial Analysis (intel, speed, verbosity) + OpenRouter API pricing, verified 2026-06-20.

#	Model	Lab	Intel	t/s	$/100T	Intel/$100T	Intel·t/s/$100T
1	GPT-OSS 120B	US	24	344	$0.687	34.9	12,017
2	Mimo V2.5~	CN	40	82	$0.659	60.7	4,978
3	DS V4 Flash	CN	40	108	$0.879	45.5	4,915
4	MiniMax M2.5~	CN	34	209	$1.560	21.8	4,555
5	Gemma 4 31B	US	29	35	$0.531	54.6	1,911
6	Grok 4.3	US	38	135	$4.435	8.6	1,157
7	DS V4 Pro	CN	44	91	$3.525	12.5	1,136
8	MiMo-V2.5-Pro	CN	42	53	$2.046	20.5	1,088
9	MiniMax-M2.7	CN	38	49	$1.990	19.1	936
10	MiniMax-M3	CN	44	57	$3.108	14.2	807
11	GPT-5.4 Nano	US	38	162	$9.021	4.2	682
12	GLM-5~	CN	40	77	$5.448	7.3	565
13	GLM-5.1	CN	40	90	$8.841	4.5	407
14	Haiku 4.5~	US	24	89	$5.730	4.2	373
15	Kimi K2.5~	CN	38	55	$6.450	5.9	324
16	Kimi K2.6	CN	43	45	$12.811	3.4	151

~ = Tok/Task estimated; cost approximate. US = US lab, CN = Chinese lab. Intel = AA Intelligence Index · t/s = AA median output speed · OR pricing.

Methodology

Raw inputs

Intel: Artificial Analysis Intelligence Index — composite benchmark score
t/s: Artificial Analysis median output speed in tokens per second
Tok/Task: AA "Output Tokens per II Task" — median output tokens per benchmark task
OR In$/1M: OpenRouter input token price (cache-miss / fresh)
OR CH$/1M: OpenRouter cache-hit input price (— = no discount offered)
OR Out$/1M: OpenRouter output token price

Cost model — $/Task

Assumes a representative agentic request of 10,000 input tokens, split 70/30 fresh vs. cached:

$/Task =
  ( 7,000 × OR_In$/1M
  + 3,000 × OR_CH$/1M   ← OR_In if no cache
  + Tok/Task × OR_Out$/1M
  ) ÷ 1,000,000

$/100T = $/Task × 100, so dollar amounts are human-readable. No model on OpenRouter charges for cache writes.

Intel / $100T

Intel ÷ $/100T

Intelligence value per $100 spent. Analogous to miles-per-gallon — higher means more intelligence per dollar. Does not account for speed.

Intel · t/s / $100T

(Intel × t/s) ÷ $/100T

The composite score. Rewards models that are simultaneously smart, fast, and cheap. Doubling speed at constant cost and intelligence doubles the score. Doubling cost halves it. Ranking by this metric is the primary sort.

Key findings

→GPT-OSS 120B ranks #1 on the composite by a 2.4× margin, driven entirely by its 344 t/s throughput — fastest in the set. Trade-off: lowest intelligence score (24).
→Mimo V2.5 wins the pure Intel/$100T race and holds #2 on the composite. Balanced across all three dimensions: intel 40, 82 t/s, $0.659/100T.
→Gemma 4 31B is the cheapest model ($0.531/100T) and ranks #2 on Intel/$100T, but falls to #5 on the composite — 35 t/s is the anchor.
→Cliff at rank 4→5: MiniMax M2.5 (rank 4, 4,555) is more than double Gemma (rank 5, 1,912). Practical panel candidates live in the top 4.
→Haiku 4.5 is the worst US-lab model on both metrics — premium-priced at $5.73/100T with the same intel as GPT-OSS 120B but ¼ the throughput.