State of the leaderboard: week 38, 14–20 September 2026
Complete week: data as of , from 7 daily updates
The week in 6 sentences
- In week 38 of 2026 (14–20 September 2026), GPT-6 Astra (OpenAI) held the #1 spot on the BenchLeader Index at 70.0, 0.9 points ahead of Claude Fable 5.1 (69.1).
- 1 model entered the top 10 (Muse Spark 1.3 at #7), pushing out GPT-5.4 Pro; 1 other model changed places.
- 13 new models arrived (k2-1b-final, k2-4b-ph1 and k2-7b-ph2 and more); none has enough independent results for an index yet.
- The biggest riser was GLM 4.5V, up 8.0 points to 51.5; the biggest faller Nemotron 3 Ultra 550B A55B, down 8.4 to 52.2.
- 15 models cut a list price and 10 models raised one; the largest cut was Nemotron Nano 12B v2 VL's input price, 100% lower.
- 6,954 new benchmark results were published, the most on CritPt (436).
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. Left the top 10: GPT-5.4 Pro.
| # | Model | Provider | Index | Last week | Move |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 70.0 | 1 | – |
| 2 | Claude Fable 5.1 | Anthropic | 69.1 | 2 | – |
| 3 | Claude Fable 5 | Anthropic | 68.2 | 3 | – |
| 4 | Claude Opus 5 | Anthropic | 68.1 | 4 | – |
| 5 | GPT-5.6 Sol | OpenAI | 66.8 | 5 | – |
| 6 | GPT-5.5 | OpenAI | 66.1 | 6 | – |
| 7 | Muse Spark 1.3 | Meta | 65.8 | – | new |
| 8 | Muse Spark | Meta | 65.6 | 8 | – |
| 9 | Muse Spark 1.1 | Meta | 65.2 | 9 | – |
| 10 | Kimi K3 | Moonshot AI | 64.9 | 7 | ↓3 |
New models and their first results
- k2-1b-final: not enough independent results for an index yet
- k2-4b-ph1: not enough independent results for an index yet
- k2-7b-ph2: not enough independent results for an index yet
- k2-mova-36b-mid5: not enough independent results for an index yet
- Mercury 2.5 (Inception): not enough independent results for an index yet. First results: 4.5% on Code Migration; 8.8% on Excel Modeling Benchmark; 18.6% on Finance Agent v2; 0.0% on Harvey's Legal Agent Benchmark v1.1.
- qwen3-8: not enough independent results for an index yet. First results: 20.0% on CritPt; 58.2% on GDPval-AA v2.1, normalised score; 51.3% on τ³-Banking (AA); 1498 on Arena Maths.
- solar-pro-4: not enough independent results for an index yet
- Palmyra X4 (Writer): not enough independent results for an index yet
- union-alpha: not enough independent results for an index yet. First results: 59.5% on LiveBench Instruction Following.
- Jamba Large (AI21 Labs): not enough independent results for an index yet
- Jamba Mini (AI21 Labs): not enough independent results for an index yet
- Step 5 Preview (StepFun): not enough independent results for an index yet
- GLM 5.3 FlashX (Zhipu AI): not enough independent results for an index yet
Biggest movers on the index
| Model | Provider | 14 Sep 2026 | 20 Sep 2026 | Change |
|---|---|---|---|---|
| Nemotron 3 Ultra 550B A55B | NVIDIA | 60.6 | 52.2 | -8.4 |
| GLM 4.5V | Zhipu AI | 43.5 | 51.5 | +8.0 |
| Grok 4.3 | SpaceXAI | 50.7 | 58.3 | +7.6 |
| GPT-5.4 Mini | OpenAI | 50.2 | 56.9 | +6.7 |
| Qwen3 14B | Alibaba | 41.2 | 47.7 | +6.5 |
| Qwen3 30B A3B | Alibaba | 42.9 | 48.9 | +6.0 |
| Grok 4.20 | SpaceXAI | 54.3 | 60.0 | +5.7 |
| Qwen3 32B | Alibaba | 46.8 | 52.1 | +5.3 |
| GPT-5.2-Codex | OpenAI | 61.3 | 56.3 | -5.0 |
| GPT-5.6 Terra | OpenAI | 58.7 | 63.4 | +4.7 |
Price changes
Cheaper (15)
- Nemotron Nano 12B v2 VL input price moved from $0.20/M to $0.00/M (100% down).
- Qwen-VL OCR input price moved from $0.72/M to $0.043/M (94% down).
- Deepseek v4 Flash Vision input price moved from $0.44/M to $0.15/M (66% down).
- Qwen3 Coder Plus output price moved from $5.00/M to $2.30/M (54% down).
- Qwen3.5 Plus (thinking) output price moved from $3.44/M to $1.72/M (50% down).
- Qwen3.5 Flash output price moved from $1.72/M to $1.03/M (40% down).
- Qwen3.5-9B (no reasoning) input price moved from $0.14/M to $0.10/M (29% down).
- MiniMax-M1 input price moved from $0.55/M to $0.40/M (27% down).
- Muse Glimmer 30B output price moved from $1.50/M to $1.10/M (27% down).
- Qwen3 Coder Next output price moved from $1.50/M to $1.20/M (20% down).
- Muse Glimmer output price moved from $1.50/M to $1.20/M (20% down).
- Qwen3.8 27B (xhigh) input price moved from $0.50/M to $0.40/M (20% down).
- …and 3 more.
More expensive (10)
- Grok 4 Fast (thinking) input price moved from $0.20/M to $1.25/M (525% up).
- Qwen3 Max (max) output price moved from $3.44/M to $7.75/M (125% up).
- Nemotron 3 Nano Omni 30B A3b (thinking) input price moved from $0.090/M to $0.20/M (122% up).
- GLM 5.3 Flash input price moved from $0.075/M to $0.15/M (100% up).
- Gemini Pro input price moved from $1.25/M to $2.00/M (60% up).
- gpt-oss-20b input price moved from $0.050/M to $0.060/M (20% up).
- Nemotron 3 Nano 30B A3B (thinking) input price moved from $0.050/M to $0.060/M (20% up).
- Qwen3.8 27B (xhigh) output price moved from $2.55/M to $3.00/M (18% up).
- Llama 3.3 70B input price moved from $0.59/M to $0.66/M (12% up).
- Gemma 4 26B A4B input price moved from $0.12/M to $0.13/M (8% up).
List prices per million tokens; each model's full price history is on its page and the LLM API pricing tracker.
Other weeks
← Week 39, 21–27 Sep 2026 · Week 37, 7–13 Sep 2026 → · The weekly review on the blog
- Week 45, 2–8 Nov 2026
- Week 44, 26 Oct–1 Nov 2026
- Week 43, 19–25 Oct 2026
- Week 42, 12–18 Oct 2026
- Week 41, 5–11 Oct 2026
- Week 40, 28 Sep–4 Oct 2026
- Week 39, 21–27 Sep 2026
- Week 38, 14–20 Sep 2026
- Week 37, 7–13 Sep 2026
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 38, 14–20 Sep 2026: who was #1”, data as of 20 Sep 2026, https://www.benchleader.com/state/2026-w38.