5 – 11 October 2026

A 15% score sounds like a failure. Per VentureBeat, though, it's the best open-weight result on Harvey's Legal Agent Benchmark, and it belongs to Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter model that outputs text only.

VentureBeat's report puts the same model at 62% on DeepSWE, ahead of GLM-5.3. It also scores 67% on Finch, a set of financial tasks, which is again the top open-weight result. TNW reports that the weights are planned for October 27. When Mistral's CEO spoke to Reuters on announcement day, the remarks named no model and gave no scores.

The catch is that these numbers come from press coverage, and there's no technical report showing how they were measured. Would you trust them before the weights are out?

"The narrative that Europe cannot compete is something that is not true," Mistral's CEO told Reuters on the day the lab is announcing a new model. The CEO also said the model is "above the Chinese models on certain aspects, including cyber." The quoted remarks don't give the model's name or benchmark scores, and they don't say whether the weights will be open.

Google Docs, Sheets and Slides now have Claude inside them, according to a post tracking Anthropic's releases, and those files also open directly in Claude. The update follows Anthropic adding Claude Docs and Claude Slides as tools inside the chat. The post calls it a small change with a big quality-of-life payoff, but it doesn't say which plans get it or when the rollout finishes.

According to a post quoting OpenAI's feature description, everyone who signs in with a ChatGPT account can now use ChatGPT Auto-review for free. It runs a separate background agent that watches the main agent and checks whether its actions match what the user intended. The post's author names the catch: approving every small step without handing over full access is still a chore, and they hope Auto-review makes it less distracting.

OpenAI's new Meetings plugin for ChatGPT takes notes during a meeting and saves a summary to ChatGPT Space, the shared workspace for people and agents. The beta is open to Pro and Business users in the ChatGPT desktop app on macOS, and OpenAI says Enterprise is "coming soon." The announcement doesn't say which meeting apps the plugin supports or how it captures the call.

One 4096x4096px image uses 2520 tokens. At $30 per 1,000,000 output tokens, that works out to $0.076 an image. Google is rolling out Nano Banana 2.1 in Google AI Studio, its APIs and Gemini.

The model's description promises better factuality, visual fidelity and multi-turn consistency "at Flash scale." Multi-turn consistency means the picture stays coherent across several rounds of edits. The post that spotted the rollout reads the price differently and guesses that it looks like a Pro model.

The post doesn't give a price for the previous version, so we can't compare the two directly. Is $0.076 a fair price for a 4K image?

If you run out of quota halfway through a slide deck, the job can now wait in a queue instead of waiting on you. An ICYMI recap of the update lists the changes to the notebook tool's limits: • Limits refresh every 5 hours, so new allocation arrives throughout the day. • One flexible Notebook-specific limit replaces the per-feature caps, and you decide how to spend it. • Artifact generation now shows a usage estimate before it runs. • A background queue holds videos, slides and other generations you defer once you're out of limits, and they generate automatically later.

The recap doesn't say how big the new shared limit is. Under the old system, limits refreshed every 24 hours.

A video titled "Anthropic Whistleblower's Testimony on Camera" has appeared on NDTV Profit, and the title is almost all the page offers: no name and no quotes we could check. Separately, senior leaders from Anthropic, OpenAI, Google and Meta testified under oath on AI risks at a New York City Council hearing held on Monday, October 5. According to a MindStudio analysis, a viral Anthropic resignation post has drawn questions about funding and timing coordination, though those claims and the whistleblower's identity are unverified.