https://moloch-sh.netlify.app/?lang=en#hoy The AI-race news that the observatory detects every day, chosen with a fixed rule. Figures from OpenRouter, Artificial Analysis, Epoch AI, Hugging Face and METR, with their source. en Sat, 31 Oct 2026 03:54:09 GMT https://moloch-sh.netlify.app/?lang=en#carrera indicador|eci|2026-10-31 Sat, 31 Oct 2026 03:54:09 GMT Best model on Epoch’s ECI: from 168.92 to 169.06 points (GPT-6 Astra). GPT-6 Astra, from OpenAI, with 169.1 points. It is in the briefing because the best model on Epoch’s ECI changes. Published on 2026-09-03; detected on 2026-10-31. Source: Epoch AI (https://epoch.ai/benchmarks/eci). https://moloch-sh.netlify.app/?lang=en#ai2027/horizonte3s valoracion|horizonte3s|observatorio|2026-10-30 Fri, 30 Oct 2026 20:14:30 GMT “Horizon of about three weeks of work”: the observatory’s assessment goes from “not yet assessable” to “ahead”. The authors frame it with METR’s 80% horizon, which does not measure reliably above 16 hours, and the last model it has measured was published on 7 Apr 2026. The observatory checks it against MirrorCode, from Epoch and METR: reimplementing whole programs from their behaviour, with tasks that would take an engineer weeks. The first model to reach 80% is Claude Opus 5 (max), with 87.8%, published on 24 Jul 2026. Rule (the observatory’s reading): “consistent” if, when the deadline expires, the best result reaches 80%. MirrorCode only covers programming and its score is not the success rate on three-week tasks. Before: “not yet assessable”. Scenario deadline: late 2026. It is in the briefing because the observatory’s assessment of an AI 2027 milestone changes. Detected on 2026-10-30. Source: Epoch AI (https://epoch.ai/benchmarks/mirrorcode). https://moloch-sh.netlify.app/?lang=en#modelos/mistralai/mistral-large-4-0 precio|mistralai/mistral-large-4-0|2026-10-30 Fri, 30 Oct 2026 20:14:30 GMT Mistral cuts the price of Mistral Large 4 on OpenRouter: input from $1.36 to $0.7 and output from $4.18 to $2.1 per million tokens. Its blended price on OpenRouter goes from $2.07 to $1.05 per million tokens. Of the 4 models with a similar AA index (±2 points of its 26), 3 cost less; the cheapest is MiMo-V2.6-Flash, at $0.175. It is in the briefing because the lab itself cuts its blended price by 49% (threshold: +50% or −33%). Detected on 2026-10-30. Source: OpenRouter (https://openrouter.ai/mistralai/mistral-large-4-0/providers). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#carrera indicador|eci|2026-10-30 Fri, 30 Oct 2026 20:14:30 GMT Best model on Epoch’s ECI: from 168.83 to 168.92 points (GPT-6 Astra). GPT-6 Astra, from OpenAI, with 168.9 points. It is in the briefing because the best model on Epoch’s ECI changes. Published on 2026-09-03; detected on 2026-10-30. Source: Epoch AI (https://epoch.ai/benchmarks/eci). https://moloch-sh.netlify.app/?lang=en#carrera/agentes indicador|mirror|2026-10-30 Fri, 30 Oct 2026 20:14:30 GMT Best MirrorCode result (Epoch runs): from 77.4 to 97.8% (Claude Sonnet 5.5 (max)). Before, 77.4%. Reimplementing whole programs from their behaviour, tasks that would take an engineer weeks; Epoch runs. In AI 2027 it is evidence for “Horizon of about three weeks of work” and “Superhuman coder”. It is in the briefing because a new MirrorCode record, the measure of tasks lasting weeks with which the observatory replaces METR. Published on 2026-09-28; detected on 2026-10-30. Source: Epoch AI (https://epoch.ai/benchmarks/mirrorcode). https://moloch-sh.netlify.app/?lang=en#modelos/google/gemini-4-carbon lider|closed|google/gemini-4-carbon|2026-10-29 Thu, 29 Oct 2026 15:05:11 GMT Gemini 4 Carbon becomes the best closed model in the catalogue on the Artificial Analysis Intelligence Index v5: 59.4 points, 1 more than Claude Opus 5.5 (58.4). It has no public API yet: it is not on OpenRouter. The gap between the best closed and the best open-weight model on the AA index stands at 13.4 points (12.4 the day before). It does not equate to months of lag. It is in the briefing because the best closed model changes with the same AA index version and a margin of 1 points (minimum 1). Detected on 2026-10-29. Source: Artificial Analysis (https://artificialanalysis.ai/models/gemini-4-carbon). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#modelos/google/gemini-4-carbon sinapi|google/gemini-4-carbon Thu, 29 Oct 2026 05:00:00 GMT Gemini 4 Carbon (Google) appears scored on Artificial Analysis with 59.4 points, still without a public API: it is not on OpenRouter. AA index 59.4: 1 points above the previous best closed model (Claude Opus 5.5). Rank 1 of 33. It is not on OpenRouter yet: no public API and no price. It is in the briefing because it is a new model and enters the top 20 of the AA index (rank 1 of 33). Published on 2026-10-28; detected on 2026-10-29. Source: Artificial Analysis (https://artificialanalysis.ai/models/gemini-4-carbon). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#hoy version|5 Wed, 28 Oct 2026 03:54:19 GMT Artificial Analysis moves its Intelligence Index from version 4.3 to 5: scores from the two versions are not comparable. Scores from one version and the other are not comparable: the daily evolution of the leaders starts over with this version. It is in the briefing because Artificial Analysis changes the version of its index. Detected on 2026-10-28. Source: Artificial Analysis (https://artificialanalysis.ai/leaderboards/models). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#modelos/openai/gpt-6.1-terra nuevo|openai/gpt-6.1-terra Mon, 26 Oct 2026 05:00:00 GMT GPT-6.1 Terra (OpenAI) appears in the OpenRouter catalogue. AA index 51: 8.4 points behind the best closed model (Gemini 4 Carbon). Rank 6 of 33. Blended price $1 per million tokens. It is in the briefing because it is a new model and enters the top 20 of the AA index (rank 6 of 33). Published on 2026-10-26; detected on 2026-10-26. Source: OpenRouter (https://openrouter.ai/openai/gpt-6.1-terra). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#modelos/moonshotai/kimi-k3.1-pro lider|open|moonshotai/kimi-k3.1-pro|2026-10-24 Sat, 24 Oct 2026 06:13:33 GMT Kimi K3.1 Pro becomes the best open-weight model in the catalogue on the Artificial Analysis Intelligence Index v4.3: 55.1 points, 8.8 more than MiMo-V2.6-Pro (46.3). Its weights have been announced by the lab, but cannot be downloaded yet. The gap between the best closed and the best open-weight model on the AA index stands at 2.5 points (11.3 the day before). It does not equate to months of lag. It is in the briefing because the best open-weight model changes with the same AA index version and a margin of 8.8 points (minimum 1). Detected on 2026-10-24. Source: Artificial Analysis (https://artificialanalysis.ai/models/kimi-k3-1-pro). Artificial Analysis data (https://artificialanalysis.ai), published with its permission. https://moloch-sh.netlify.app/?lang=en#modelos/deepseek/deepseek-v4.1-flash-1017 uso|primero|deepseek/deepseek-v4.1-flash-20261017|2026-10-24 Sat, 24 Oct 2026 06:13:33 GMT DeepSeek V4.1 Flash 1017 becomes the most used model on OpenRouter over the last 7 days (14.9% of tokens), ahead of DeepSeek V4.1 Flash. 14.9% of OpenRouter’s tokens from 2026-10-17 to 2026-10-23; before, the first was deepseek/deepseek-v4.1-flash-20260910. This is usage on OpenRouter, not across the whole market. It is in the briefing because the most used model on OpenRouter over the last 7 days changes. Published on 2026-10-23; detected on 2026-10-24. Source: OpenRouter · usage (https://openrouter.ai/rankings). https://moloch-sh.netlify.app/?lang=en#modelos/minimax/minimax-m3.1-flash precio|minimax/minimax-m3.1-flash|2026-10-23 Fri, 23 Oct 2026 04:25:22 GMT Minimax starts charging for MiniMax M3.1 Flash on OpenRouter: $0.15 input and $0.6 output per million tokens. Until now it was free. Its blended price on OpenRouter goes from free to $0.262 per million tokens. Of the 10 models with a similar AA index (±2 points of its 42.7), 3 cost less; the cheapest is Ling 3.1 Flash, at $0.112. It is in the briefing because the lab itself starts charging for it on OpenRouter, where it was free. Detected on 2026-10-23. Source: OpenRouter (https://openrouter.ai/minimax/minimax-m3.1-flash/providers). Artificial Analysis data (https://artificialanalysis.ai), published with its permission.