AI Daily · Oct 6, 2026: Mistral Large 4 open-weights release: 1T-param natively multimodal model, half price on OpenRouter for two weeks
Mistral released Large 4 (Le Chonk), a 1T-parameter, 49B-active natively multimodal open-weights model, now in public preview on OpenRouter at 50% off for two weeks ($0.68/M input, $2.09/M output). Reflection AI also announced Beam, a 501B-param agentic model with weights open-sourcing this month.
Pick a date
-
Mistral Large 4 arrives on OpenRouter in public preview with launch discount
Mistral Large 4 (Le Chonk) is now on OpenRouter in public preview with 1T parameters (49B active), natively multimodal, 512K context and up to 256K output, priced 50% off for two weeks at $0.68/M input, $2.09/M output, and $0.07/M cached.
-
Reflection AI introduces Beam agentic open model with 501B total parameters
Reflection AI announced Beam, an efficient agentic open model with 501B total parameters and 23B active parameters, trained from scratch, focused on coding and agentic tasks, with full weights release planned this month. Ollama confirmed plans to host the model.
-
OpenAI speeds GPT-6 Astra and GPT-6.1 Sol by about 50 percent
OpenAI says default speed for GPT-6 Astra and GPT-6.1 Sol via subscription across products and partners using Sign in With ChatGPT has been optimized to about 50% faster, rolling out within two hours with no user action needed.
-
OpenAI makes Codex Auto-review free for all ChatGPT users
OpenAI has enabled Auto-review free of charge for everyone signed in with a ChatGPT account, toggleable under settings > permissions > auto-review. It loosens the default sandbox that demands approval for every action, reducing the need to hand-configure fine-grained rules.
-
Command Code releases Agr open decision model family with 31B and 360M variants
Command Code launched Agr, an open decision model family including Agr (31B) and Agr-flash (360M), scoring 58.15 on Decision Index 0.2.1 and optimized for tool calls and routing. The models are available on Hugging Face and GitHub with TypeSafe SDK support.
-
Cline adds Pareto 26.10 Preview routing model at 56x lower cost
Cline now offers Pareto 26.10 Preview, which routes queries across frontier and open models, picks the best answer while preserving prompt cache, and costs $0.24 per task versus $13.41 for Fable at the same DeepSWE score — about 56x cheaper.
-
OpenAI extends content provenance to text under EU rules
OpenAI is expanding content provenance to cover text to meet EU regulatory requirements, embedding an invisible statistical signal in generated text. Only approved researchers will initially access the detector, as watermarks can be removed by rewriting or translating and often go undetected in short passages.
-
Requesty benchmark compares Clef and Opus 5.5 routing costs
Requesty tested 22,000 decisions across 3 decision models and 8 LLMs: Clef (27B) hit 83.2% accuracy at $0.13 per 1K items, while Opus 5.5 reached 86.6% at $3.80 per 1K items on human-labeled gateway tasks.
-
Tencent open-sources Octop multi-user AI workspace with isolation
Tencent released Octop as an open-source, multi-user AI workspace where each account logs in separately and gets isolated data, with multiple agents per user. Each agent ships with a browser, terminal, and swappable models, storage, memory, and plugins.
-
GMI Cloud offers GPT-6.1 Sol at a fifth of Astra price
GMI Cloud announced that GPT-6.1 Sol is now available to all GMI users, positioned as near-Astra Intelligence capability at a fifth of the price.