This Week's Releases: Mistral Publishes Large 4's Weights, GitHub Ships Copilot Fusion, Claude Gets Message Threads and Gemini Edits Files in Chat

Title: This Week's Releases: Mistral Publishes Large 4's Weights, GitHub Ships Copilot Fusion, Claude Gets Message Threads and Gemini Edits Files in Chat
Date: 4 November 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1485 words
Reading time: 6 min
Published: 04-11-2026
Mistral released the weights of Mistral Large 4, its largest open model, on 29 October. GitHub made Copilot Fusion generally available on 28 October. Copilot Fusion picks which AI models handle each coding task. Google let Gemini users create and edit documents, PDFs and spreadsheets inside the chat on 29 October. OpenAI released GPT-6 Rosalind, a biology model for approved organisations, on 30 October. Anthropic opened a public beta of message threads in its API on 2 November. Anthropic also began watermarking text from four older Claude models on 3 November. Deepgram agreed on 29 October to buy Daily, the company behind the open-source Pipecat voice framework.
Mistral Large 4's weights
Mistral, the French AI company, published the full weights of Mistral Large 4 on Thursday 29 October. Weights are the numbers that make up a trained model. With them, an organisation can download the model and run it on its own servers. Mistral first offered Large 4 as a public preview through Mistral Studio and its API on 6 October. It had promised the weights by the end of the month.
Large 4 has about 1.05 trillion parameters in total. It is a mixture-of-experts model, which means only part of it runs for each word it produces. Most reports give 49 billion active parameters. Mistral's preview documentation listed 52 billion. The model handles images as well as text.
The weights are on Hugging Face, in FP8 and NVFP4 versions, which are compressed formats that need less memory. The licence is the Mistral Ecosystem Open-Weight License, which reports describe as a modified MIT licence. One report says it allows free commercial use, modification and retraining. Organisations should read the licence text itself before building on it.
Mistral says Large 4 leads all open models on Artificial Analysis' Coding Agent Index. One report puts its Terminal-Bench 4.0 score at about 41 per cent, on Mistral's figures. Neither number had been checked independently in the coverage we found.
GitHub's Copilot Fusion
GitHub made Copilot Fusion generally available on Wednesday 28 October, at its Universe conference in San Francisco. Fusion is a routing layer. It decides which models handle a coding request and how they work together.
Fusion uses three patterns, according to write-ups of GitHub's documentation. In the first, one model answers on its own. In a cascade, a cheaper model drafts an answer and a check either accepts it or passes the task to a stronger model. In a critique, one model drafts, others review the draft, and the first model revises it. GitHub has not published the rules for choosing between them, and it says the set of candidate models may change.
Users pick an effort level: Standard, High or Max. One report says Free and Pro plans get Standard only, High starts at Pro+, and Max comes with the Max, Business and Enterprise plans. Higher levels can call more models per task and use more AI credits.
There is no separate charge for Fusion. GitHub bills each model Fusion calls at that model's standard rate. Paid plans get 30 per cent off those model costs until 30 November. Fusion works in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs.
Artificial Analysis scored Fusion Max at 60 on its Coding Agent Index v2.0, against 62 for Claude Code running Opus 5.5 at maximum effort. On that test, Fusion's average API cost was $5.59 per task, against $11.42 for Claude Code. Those costs come from the benchmark runs.
Gemini edits files in the chat
Google announced on Thursday 29 October that the Gemini app can now create, preview and edit files inside a conversation. The formats are documents, PDFs, presentations, spreadsheets, Markdown, plain text and code. Until now, editing a file Gemini had made meant downloading it and opening another app.
A user can open the file full screen and change it by hand. They can also highlight a passage and ask Gemini to rewrite it. Google says the feature is open to all Gemini users, including free accounts. PDFs work on the web and in the mobile apps. The other formats work on the web only for now.
OpenAI's GPT-6 Rosalind
OpenAI released GPT-6 Rosalind on Friday 30 October. It is a model for life sciences research, such as biology and drug discovery. Only organisations approved through OpenAI's trusted-access programme can use it. OpenAI checks each applicant's research mission, its use case, its governance and its ability to control who uses the model.
Approved organisations can use Rosalind in ChatGPT Work and Codex. API access needs a separate approval, and the model is not in the regular Chat tab. OpenAI's system card addendum says Rosalind uses the same weights as GPT-6 Astra. OpenAI says it added no training data and did no further training for this release. The change is that Rosalind refuses fewer dual-use biology questions, ones that could serve research or cause harm.
One comparison of OpenAI's figures reports that GPT-6 Astra refused 29.2 per cent of the test questions and Rosalind 9.4 per cent. The same source says Rosalind's answers passed OpenAI's safety grader 79.0 per cent of the time, against 99.8 per cent for Astra. OpenAI rates Rosalind 'High' for biological and chemical capability, one step below 'Critical'. One outlet reports a price of $15 per million input tokens and $75 per million output tokens. We could not confirm that on OpenAI's own pricing page.
Also on 30 October, ChatGPT on iOS and Android gained account switching. A user can stay signed in to two accounts and move between them under Settings. Each account keeps its own history, memory and settings.
Message threads in Claude's API
Anthropic opened a public beta of message threads for its Messages API on Monday 2 November. Without threads, a developer sends the whole conversation with every request. With threads, Anthropic's servers hold the conversation. The developer sends only the new message and the ID of Claude's last reply.
Threads make replies start sooner. Coverage of Anthropic's announcement gives a fall of about 35 per cent in the median time to the first word of a reply, on requests of 500,000 to one million tokens. Cost does not change. Claude still reads the full conversation on every turn, so the number of tokens billed stays the same.
A thread expires five minutes after the last request. After that, the old ID returns an error, so applications still need to keep their own copy of the conversation. The beta runs on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry. One outlet reports that it is not open to customers on zero data retention or HIPAA arrangements.
Watermarks on older Claude models
Anthropic began adding an invisible watermark to text from four older models on Tuesday 3 November. They are Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Anthropic told organisations using those models by email. Reports say Amazon Bedrock may take a few more days. Claude models launched on or after 2 August 2026 have carried the watermark from launch.
Reports link the watermark to the transparency rules of the EU AI Act. Anthropic says the mark does not change the meaning or quality of the text. It also says the mark carries no identifying information and cannot be traced to a person, an organisation or a chat. Detection is done by Anthropic. Short passages are harder to detect, and the mark can appear on human writing that Claude has only translated or proofread.
Deepgram buys Daily
Deepgram, a speech recognition company, announced on Thursday 29 October that it is buying Daily. Daily makes Pipecat, an open-source framework for building voice agents. The price was not disclosed. Deepgram says Pipecat will stay free, open source and vendor neutral. Kwindla Hultman Kramer, a Daily co-founder, joins Deepgram as head of open source. Daily's WebRTC and telephony products will carry on.
Deepgram also previewed a Voice Agent Platform, to run on Deepgram's own hosting or in a customer's cloud. It gave no date for general availability. Its Think-Live feature is generally available now through the existing Voice Agent API.
Before switching
Most of this week's benchmark figures come from the companies that built the products. The Artificial Analysis score for Copilot Fusion is the main independent number.
Teams on GitHub's paid plans can try Fusion at the discounted rate this month and compare the bill with a fixed choice of model. Developers using message threads should keep their own transcript, because Anthropic's copy lasts five minutes.
In SAASiQ's view, Mistral Large 4 is the release to look at for organisations that need a large model on their own servers, once their lawyers have read the licence.
Moonshot has promised the weights of Kimi K3.1 Pro by 4 November. GitHub's 30 per cent Fusion discount ends on 30 November. Microsoft Ignite runs from 17 to 20 November in San Francisco.
SAASiQ - Intelligent Solutions for SaaS ©


