AI Daily · Oct 7, 2026: Mistral Large 4 launches with 1T parameters, challenging frontier closed models
Mistral AI launched Mistral Large 4, a 1T-parameter multimodal open-weights model claiming to be the best from the US or Europe, with leading results on CTF, automation, and legal benchmarks. OpenAI also cut its top API rate tier threshold from $1000 to $500.
Showing the latest briefing (Oct 7, 2026). Every day is generated the next morning.
Pick a date
-
Mistral Large 4 launches with 1T parameters and 49B active
Mistral AI announced Mistral Large 4 (Le Chonk), a natively multimodal model with 1 trillion parameters and 49 billion active parameters, described as the best open-weights model from the US or Europe on aggregated benchmarks.
-
Mistral Large 4 Solves 18 Of 19 CTF Challenges
Mistral AI reports Mistral Large 4 solving 18 of 19 CTF challenges in a speedrun benchmark with real tool calls and solve times, positioning it among the strongest models for cybersecurity.
-
OpenAI simplifies API rate limit tiers with lower thresholds
OpenAI is consolidating five paid usage tiers into three (Build, Launch, Grow), with the new top tier Grow requiring $500 in total API payments, down from $1,000 previously.
-
Perplexity releases open weights pplx-decider-v1.1-27b model
Perplexity's pplx-decider-v1.1-27b open weights multimodal decision model is available, topping Hugging Face's Decision Index 0.3 benchmark at $0.02 per million input tokens, half the price of v1.
-
OpenAI launches Decisions API in public beta for all developers
OpenAI's Decisions API is now available to all developers in public beta, letting apps pick the right model, tool, or action in near-real time; it makes decisions up to 10x faster than GPT-based alternatives.
-
Google's Nano Banana 2.1 image model now live on OpenRouter
Google's Nano Banana 2.1 is available on OpenRouter, supporting up to 14 reference images, Google Search grounding, and pricing of $0.0336 per 1K image, $0.0504 per 2K image, and $0.1134 per 4K image.
-
Prompt caching cuts coding agent task cost from $126 to $9.59
LiteLLM reported that harness changes yielded a 13x price difference on the same task with the same model—$126 vs $9.59—driven almost entirely by prompt caching at 96% versus 0%.
-
Cursor launches iOS app to control coding agents from phone
Cursor released an iOS app that lets users check in, reply to, and start new agent tasks from their phone, while the agent itself continues running on their computer even if the phone loses signal. Enterprise customers need an admin to enable it.
-
LiteLLM open sources Moyai coding agent cutting Devin bills 79 percent
LiteLLM open-sourced Moyai, a cloud coding agent compatible with Claude Code, Codex, OpenCode, and running across 100+ providers via LiteLLM; their team reports Devin bills dropped from $101,872 to $21,700, a 79% reduction.
-
DeepMind Releases EmbeddingGemma 2 Multimodal On-Device Model
Google DeepMind announced EmbeddingGemma 2, its first natively multimodal open model for on-device embeddings, supporting code, images, audio, and video in a shared embedding space.