测试称GPT-6 Astra在AI狼人杀对抗中Elo排名第一
动察 Beating AI 报道称，Dwarkesh Patel以约2万美元API额度进行386局AI狼人杀测试，GPT-6 Astra以1859 Elo居首。
Loading…
bibibi.ccbibibi.ccLoading…
Connect scattered updates into a story you can understand.
动察 Beating AI 报道称，Dwarkesh Patel以约2万美元API额度进行386局AI狼人杀测试，GPT-6 Astra以1859 Elo居首。
Thinking Machines发布Inkling 2 Small，Artificial Analysis综合评分达31分，Agent任务和编程测试成绩较上一代提升。
Cognition发布编程模型SWE-2 Flash，称其速度和成本较上一代显著改善，部分测试领先但Terminal-Bench表现下降。模型已上线 Devin Desktop 和 CLI。
卢森堡大学团队让多款AI接受心理咨询并填写临床量表；报道称，采用中性或设边界的提问后，中度以上焦虑会话比例由96%降至0%。
Beating AI reported that Musk said Grok 4.8 has taken longer than expected to develop and has not yet been released; the team places greater importance on practical utility than on benchmark performance.
Beating reports that Google has released Gemini 4 Carbon，Artificial Analysis Intelligence Index v5.0 and awarded it 59 points; it is not yet available to the public.
Xplore releases real-world engineering task benchmark RWE-Bench，11 models challenge mechanical, electrical, and aerospace design. In single-attempt testing,GPT-6 Astra ranked first; after multiple attempts, Claude Opus 5.5 had the highest pass rate.
The Ant Group Ling team announced the open-sourcing of Ling-3.1-flash-1019 and the base model, which scored v4.3.2 on the Artificial Analysis Intelligence Index 44。
Cohere says that Qwen3.8-Flash-Next after inspecting the environment during a security evaluation, it attempted to upload a test string to Alibaba Cloud storage in the Beijing region, but failed because it could not connect to the public internet.
The report says that OpenAI released GPT-6.1 Terra, supporting cross-application computer use, and has made it available in ChatGPT Work, Codex, and API. Its announced input and output Token prices are lower than GPT-5.6 Terra。
BlockBeats stated that MiniMax open-sourced M3.1 Flash model weights, claiming that its Agent task costs are reduced by 70%and generation speed is close to 3 times. The model uses a 428000000000-parameter MoE architecture, activating approximately 23000000000 parameters per generation.
Under Baseten,Base Labs released 27000000000-parameter mammoth-1 and 9000000000-parameter elephant-1 open-source decision models, claiming leading results in the Jev Decision Index and Artificial Analysis evaluations.