BenchLeader
Menu

State of the leaderboard: week 37, 7–13 September 2026

Complete week: data as of , from 6 daily updates

The week in 5 sentences

  • In week 37 of 2026 (7–13 September 2026), GPT-6 Astra (OpenAI) was #1 on the BenchLeader Index at 70.4, 0.8 points ahead of Claude Fable 5.1 (69.6).
  • 1 new model arrived (agnes-3-0-flash); none has enough independent results for an index yet.
  • The biggest riser was GPT-5 mini, up 8.3 points to 58.8; the biggest faller Magistral Medium, down 6.8 to 39.7.
  • 37 models cut a list price and 32 models raised one; the largest cut was Muse Spark 1.2's output price, 95% lower.
  • 3 new benchmark results were published, the most on AA Intelligence Index v4.3.2 (1).

On 9 Sep 2026, 10 Sep 2026, 12 Sep 2026, 13 Sep 2026 BenchLeader added a source that listed its back catalogue; those models and results are new to this site, not new to the world, and are left out of the counts above.

Top 10 at the end of the week

The BenchLeader Index, one row per model at its best reasoning effort, against the table at the end of the previous week.

#ModelProviderIndexLast weekMove
1GPT-6 AstraOpenAI70.4–new
2Claude Fable 5.1Anthropic69.6–new
3Claude Fable 5Anthropic68.1–new
4Claude Opus 5Anthropic67.3–new
5GPT-5.6 SolOpenAI66.9–new
6GPT-5.5OpenAI65.8–new
7Kimi K3Moonshot AI65.5–new
8Muse SparkMeta65.4–new
9Muse Spark 1.1Meta64.5–new
10GPT-5.4 ProOpenAI64.4–new

New models and their first results

Biggest movers on the index

ModelProvider7 Sep 202613 Sep 2026Change
GPT-5 miniOpenAI50.558.8+8.3
MiniMax M2.7MiniMax49.456.6+7.2
Magistral MediumMistral AI46.539.7-6.8
MiniMax M2.5MiniMax46.753.4+6.7
gpt-oss-120bOpenAI42.549.1+6.6
GLM 5.1Zhipu AI53.660.1+6.5
MiniMax M3MiniMax52.358.1+5.8
Qwen3 235B A22BAlibaba51.745.9-5.8
Qwen3 30B A3BAlibaba48.642.9-5.7
Grok 4 FastSpaceXAI46.752.2+5.5

Price changes

Cheaper (37)

  • Muse Spark 1.2 (xhigh) output price moved from $4.25/M to $0.20/M (95% down).
  • Muse Spark 1.3 (max) output price moved from $4.25/M to $0.20/M (95% down).
  • MiMo-V2.5 output price moved from $2.00/M to $0.28/M (86% down).
  • Qwen3.5 Flash (no reasoning) output price moved from $1.72/M to $0.26/M (85% down).
  • GLM 5.2 (max) output price moved from $3.85/M to $0.88/M (77% down).
  • Qwen3 32B (no reasoning) output price moved from $1.15/M to $0.28/M (76% down).
  • Mistral Medium 3 input price moved from $1.50/M to $0.40/M (73% down).
  • Gemma 4 26B A4B (minimal) input price moved from $0.13/M to $0.042/M (68% down).
  • MiMo-V2.5-Pro output price moved from $3.00/M to $1.04/M (65% down).
  • LongCat 2.0 input price moved from $0.75/M to $0.30/M (60% down).
  • Qwen3.7 Plus (no reasoning) output price moved from $3.00/M to $1.28/M (57% down).
  • GPT-5.6 Sol (max) input price moved from $4.40/M to $2.00/M (55% down).
  • …and 25 more.

More expensive (32)

  • Grok 4 (thinking) output price moved from $2.50/M to $15.00/M (500% up).
  • Ling 3.0 Flash input price moved from $0.021/M to $0.080/M (281% up).
  • Llama 3-70B output price moved from $0.74/M to $2.75/M (272% up).
  • Llama 3-8B output price moved from $0.040/M to $0.14/M (250% up).
  • Qwen3.5 397B-A17B (no reasoning) input price moved from $0.17/M to $0.60/M (249% up).
  • GPT-3.5 Turbo (high) input price moved from $0.50/M to $1.50/M (200% up).
  • DeepSeek V3.1 (thinking) input price moved from $0.57/M to $1.64/M (186% up).
  • Qwen3 235B A22B (thinking) output price moved from $3.00/M to $8.40/M (180% up).
  • Qwen3-VL 30B-A3B (thinking) output price moved from $1.00/M to $2.40/M (140% up).
  • DeepSeek V4 Flash output price moved from $0.28/M to $0.60/M (114% up).
  • DeepSeek V4 Pro (max) input price moved from $0.43/M to $0.87/M (100% up).
  • GLM 5.3 Flash (max) input price moved from $0.075/M to $0.15/M (100% up).
  • …and 20 more.

List prices per million tokens; each model's full price history is on its page and the LLM API pricing tracker.

Other weeks

← Week 38, 14–20 Sep 2026 · The weekly review on the blog

Benchmark categories: AI agent benchmarks · AI coding benchmarks · Reasoning and math benchmarks · Multimodal AI benchmarks · New AI model releases

How this is built: the BenchLeader Index is recomputed with every update from independent benchmark publishers (see the methodology), and a maker's own launch numbers are never used. Past ranks are rebuilt from the daily index history of the models listed today.

Cite as: BenchLeader, “AI leaderboard week 37, 7–13 Sep 2026: who was #1”, data as of 13 Sep 2026, https://www.benchleader.com/state/2026-w37.