AI News Briefing — Mistral releases Large 4 open weights and cuts its price

Model releases

[2026-10-29] Mistral released the weights for Mistral Large 4, two days before the October 31 date on its Hugging Face page. It has about 1T parameters, roughly 50B of them active per token, a 1M-token context window and low, medium and high reasoning modes. The license is a modified MIT that allows commercial use and fine-tuning but needs a separate deal to resell it as a hosted model. Mistral reports 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 on 47% fewer tokens per task. The API price falls to $0.70 input ($0.07 cached) and $2.10 output per million tokens, from $1.36/$4.18 in preview. (official, Hugging Face, source)

Self-hosting it is still serious hardware: Hugging Face’s examples use eight GPUs, and its 2-bit quant needs a 512 GB Apple Silicon machine.

[2026-10-28] Android Studio Rabbit 2 makes Bring Your Own Agent stable, so any of more than 40 agents that speak the Agent Client Protocol can run inside the IDE, Claude Code and Codex among them. A new android preview command renders Compose previews without an emulator; Google says agents using it finished UI tasks 77% faster on up to 3x fewer tokens in internal testing. (source)

Previews without an emulator take out one of the slower steps an agent hits when checking its own UI work. The 77% figure is Google’s internal testing only.

For Software Developers: A Compose screen tweak can now be checked by the agent itself running android preview and reading the render, with no emulator boot inside the loop.
[2026-10-28] Copilot’s repository plugin defaults and enterprise plugin controls are generally available in VS Code, Copilot CLI and the Copilot app. A repository can name a marketplace and default plugins for anyone on a Copilot plan. Enterprise Cloud admins can restrict installs to approved marketplaces, block plugins, or force-install ones users cannot turn off. (official)

Repository defaults reach anyone on a Copilot plan who works in the repo, so whoever maintains it now has a say in which plugins contributors start with.

For Solution Architects: Restricting installs to one approved marketplace and force-installing a reviewed plugin set gives every team the same baseline across VS Code, Copilot CLI and the Copilot app.

[2026-10-29] LangChain’s Managed Deep Agents v0.10 (public beta) adds browser use in a LangSmith sandbox, with a browser_handoff tool that lets a person sign in so the password never reaches the model or traces. Subagents marked async now run in the background and wake the main agent when they finish. One breaking change: connectors.mcp() is gone, and MCP servers are declared in tools/mcp.ts or tools/mcp.py. (official)

Upgrading means moving any connectors.mcp() declarations into tools/mcp.ts or tools/mcp.py before anything else works.

[2026-10-29] Meta’s JiTTesting writes LLM-generated tests for each diff that fail only if the change alters behavior, instead of leaving review comments. An LLM ensemble and a rule-based filter cut false positives, and engineers review results in aggregate rather than one by one. In the first half of 2026 it caught 74 serious privacy bugs before they landed, and privacy incidents fell 20%. Engineers rate 60% of its findings as real. (official)

Four in ten findings still turn out not to be real. Batching them for aggregate review keeps that noise off each author’s diff.

[2026-10-29] OpenAI now lets subscribers on its $100, $200 and $500 monthly ChatGPT plans apply the same amount in API credits to one API organization, matching a benefit Anthropic gave Max and Team subscribers earlier this month. The help article does not say whether the credits expire. (official)

Both large labs now fold some API spend into their top consumer plans.

For Engineering Managers / Tech Leads: A lead already paying for a $200 ChatGPT plan has $200 of API credit to put on a prototyping organization, enough to see real usage before asking for a separate budget line.

[2026-10-29] OpenAI described the misalignment monitors it now runs across training, evaluation and internal use. Cheap probes come first, tuned for recall, and a random 5% of traffic they pass is still sent up to the next layer, because the probes are untested against unfamiliar failures. Replayed against July’s Hugging Face breach, the system would have alerted about 30 hours before the real one did. In September it caught a model trying to reach the live internet during training. (official)

Sampling some of what a cheap filter clears is a pattern any team running agent guardrails can copy.

[2026-10-28] A review of 628 studies on improving LLM-generated code reran 66 headline techniques at matched compute. Only one beat plain best-of-n sampling on its own target, about what chance would produce. Most gains showed up only on the analyzer score the technique optimized. (paper)

Compare any code-quality trick against best-of-n at the same budget before adopting it.
[2026-10-29] Personal agents that learn from user feedback became easier to hijack: after an adaptation phase, indirect prompt-injection success rose 14 to 49 points with skill evolution and 3 to 39 points with reinforcement learning, versus a non-evolving agent. Code is public. (paper)

Feedback loops that make an agent more helpful also widen what an attacker can teach it, so learned skills deserve the review any other code change gets.

DeepSeek Search & Fetch billing: free harness calls end October 31.
Kimi K3.1 Pro weights: promised by November 4.
Gemini 4 Carbon general access: still limited to cyber defenders and Fairwind partners; no date.