Daily digest, 27 Sept 2026

BenchLeader refreshed 1521 models from 22 of 22 sources. 15 price changes, 155 new or updated benchmark results across 22 sources.

Price changes

• GLM 5.2 (max) cost per task moved from 1 to 1.5 (53% up). (details)
• Qwen3.7 Plus cost per task moved from 0.3 to 0.2 (26% down). (details)
• Grok 4.3 (high) cost per task moved from 0.2 to 0.2 (24% up). (details)
• Nemotron 3 Ultra 550B A55B (thinking) cost per task moved from 0.5 to 0.6 (9% up). (details)
• GPT-5.4 Mini (xhigh) cost per task moved from 0.4 to 0.4 (9% up). (details)
• Mistral Medium 3.5 cost per task moved from 0.4 to 0.5 (15% up). (details)
• Claude Haiku 4.5 (thinking) cost per task moved from 0.2 to 0.3 (33% up). (details)
• Mercury 2.5 cost per task moved from 0.1 to 0.1 (82% up). (details)
• Grok 4.3 (no reasoning) cost per task moved from 0.1 to 0.2 (24% up). (details)
• Mistral Large 3 cost per task moved from 0.1 to 0 (42% down). (details)
• Mistral Small 3.1 output price moved from $0.17/M to $0.30/M (76% up). (details)
• Mixtral 8x7B input price moved from $0.45/M to $0.70/M (56% up). (details)
• Mistral Medium 3.5 (high) cost per task moved from 0.4 to 0.5 (15% up). (details)
• Mercury 2.5 (high) cost per task moved from 0.1 to 0.1 (82% up). (details)
• Qwen3.7 Plus (no reasoning) cost per task moved from 0.3 to 0.2 (26% down). (details)

New benchmark results

• AudioMC (audio output): 15 new results, including Qwen3 Omni 30B A3B Instruct, Gemini 2.5 Flash Native Audio 12 2025, Gemini 2.5 Flash Native Audio 12 2025… (board)
• DTBench: 6 new results, including Mistral Medium 3.5, Magistral Small 1.2, Mistral Small 3.1… (board)
• Epoch Capabilities Index: 6 new results, including Baichuan1-7B, Falcon 2 11B, Stable Beluga 2… (board)
• EnigmaEval: 5 new results, including GPT-5.6 Sol, Claude Fable 5, Gemini 3.5 Flash… (board)
• SWE-bench Verified (any scaffold): 5 new results, including Claude Sonnet 4.5, GPT-5, Claude Opus 4.5… (board)
• SciCode: 4 new results, including Mistral Medium 3.5, Magistral Small 1.2, Devstral Small 2… (board)
• LiveCodeBench: 4 new results, including Devstral 2, Magistral Small 1.2, Devstral Small 2… (board)
• AA Intelligence Index v4.3.2: 3 new results, including MiMo-V2.6-Flash, Magistral Small 1.2, Mistral Small 4 (board)
• Humanity's Last Exam (AA): 3 new results, including MiMo-V2.6-Flash, Magistral Small 1.2, Mistral Small 4 (board)
• GDPval-AA v2.1: 3 new results, including MiMo-V2.6-Flash, Motif 3, Magistral Small 1.2 (board)
• LMCA: 3 new results, including Mistral Medium 3.5, Command R+, Mistral Small 4 (board)
• MATH Level 5: 3 new results, including Mistral Small 3, Phi 3, Mistral Small 3.1 (board)

Score revisions

• Gemini 3 Pro on SWE-bench Verified (any scaffold): 77.4% → 74.2%. (details)
• Motif 3 on GPQA Diamond (AA): 83.4% → 86.9%. (details)
• Motif 3 on Humanity's Last Exam (AA): 37% → 40.4%. (details)
• Claude Opus 4 on SWE-bench Verified (any scaffold): 73.2% → 67.6%. (details)
• Kimi K2 on SWE-bench Verified (any scaffold): 65.4% → 53.4%. (details)
• Claude Sonnet 4 on SWE-bench Verified (any scaffold): 76.8% → 74.8%. (details)
• Claude 3.7 Sonnet on SWE-bench Verified (any scaffold): 70.4% → 52.8%. (details)
• Qwen3 Coder 480B on SWE-bench Verified (any scaffold): 69.6% → 55.4%. (details)
• GPT-4o on SWE-bench Verified (any scaffold): 38.8% → 21.6%. (details)
• Gemini 2.0 Flash on SWE-bench Verified (any scaffold): 52.2% → 13.5%. (details)
• Qwen2.5 Coder 32B Instruct on SWE-bench Verified (any scaffold): 47% → 38%. (details)
• Mistral Small 3.1 on GPQA Diamond: 41.9% → 47.5%. (details)
• Mistral Small 3.1 on OTIS Mock AIME: 3.9% → 5.8%. (details)
• Mixtral 8x7B on MATH Level 5: 9.3% → 10%. (details)
• Baichuan2-13B on MMLU: 55.1% → 59.2%. (details)
• Baichuan2-13B on GSM8K: 45.7% → 52.8%. (details)
• Baichuan2-13B on BIG-Bench Hard: 47.2% → 49%. (details)

Speed changes

• Claude Opus 5.5 output speed changed from 61 tok/s to 95 tok/s. (details)
• Claude Fable 5 time to first answer changed from 153.30 s to 96.96 s. (details)
• Claude Opus 5 time to first token changed from 2.87 s to 4.03 s. (details)
• GPT-5.5 time to first answer changed from 65.38 s to 35.33 s. (details)
• GPT-5.5 response time changed from 70.03 s to 39.95 s. (details)
• Claude Opus 4.7 time to first token changed from 3.48 s to 1.38 s. (details)
• GPT-5.2 time to first token changed from 2.33 s to 3.26 s. (details)
• DeepSeek V4 Pro time to first token changed from 1.52 s to 2.32 s. (details)
• GLM 5.2 time to first token changed from 0.87 s to 1.21 s. (details)
• GPT-5.4 Pro output speed changed from 6 tok/s to 1 tok/s. (details)
• GPT-5.4 Pro time to first token changed from 62.03 s to 5.67 s. (details)
• GLM 5.3 Flash time to first token changed from 1.05 s to 1.54 s. (details)
• Claude Opus 4.8 time to first answer changed from 15.30 s to 9.41 s. (details)
• Grok 4.7 output speed changed from 57 tok/s to 82 tok/s. (details)
• Grok 4.7 time to first answer changed from 2.81 s to 48.09 s. (details)
• …and 73 more.

Today's top five

1. Claude Opus 5.5 — 72.0
2. Claude Fable 5.1 — 70.8
3. GPT-6 Astra — 70.2
4. Claude Opus 5.5 — 70.2
5. Gemini 4 Argon — 70.0

This digest is generated automatically from the day's data changes. Every line links to the page where you can check the numbers and their source.