Production systems arrive while consumer surfaces shift
Company knowledge, Mistral AI Studio and Vercel Agent, and The Deploy Log for this week. Every price and claim read off the publisher's own page, with a call on each.
OpenAI
Company knowledge
DEPLOYif you run ChatGPT Business, Enterprise, or Edu
WHAT SHIPPEDCompany knowledge for ChatGPT Business, Enterprise, and Edu brings context from connected apps like Slack, SharePoint, Google Drive, and GitHub into answers with citations, powered by a version of GPT-5 trained to look across multiple sources.
THE CATCHWhen company knowledge is on, ChatGPT cannot search the web or create charts and images, and you must manually select it for each new conversation to get the full depth and citations.
IF YOU RUNa team where answers live across Slack, docs, tickets, and email and you need cited, permission-aware summaries
SOURCEOpenAI, Work smarter with your company knowledge in ChatGPT
Mistral
Mistral AI Studio
WATCHif your AI work is stuck at prototypes
WHAT SHIPPEDMistral AI Studio packages Observability, Agent Runtime, and AI Registry into one platform for building, evaluating, and running AI in production, with hybrid, dedicated, and self-hosted deployments.
THE CATCHAccess is a private beta, so you cannot deploy it into a production system today without signing up and being admitted.
IF YOU RUNan enterprise AI team blocked by missing evaluation harnesses, versioning, and governed deployment
SOURCEMistral, Introducing Mistral AI Studio
Vercel
Vercel Agent
DEPLOYif you ship on Vercel and use GitHub
WHAT SHIPPEDVercel Agent is available in Public Beta with Code Review and Investigations, running simulated builds inside a Vercel Sandbox to verify pull request recommendations and analyzing telemetry to explain production anomalies.
THE CATCHCode Review and Investigations are the only two skills available today; the agent cannot yet create pull requests or propose fixes.
IF YOU RUNa Vercel team that wants tested pull request feedback and faster root cause analysis on anomalies
SOURCEVercel, Introducing Vercel Agent
The Deploy Log
Models and APIs
| DEPLOYif you train robot policies LeRobot v0.4.0. LeRobot v0.4.0 introduces Datasets v3.0 with chunked episodes and streaming, integrates PI0, PI0.5, and GR00T N1.5 policies, adds LIBERO and Meta-World simulation support, and launches a plugin system for third-party hardware. Hugging Face |
| DEPLOYif you need UK data storage UK data residency. OpenAI is introducing UK data residency on Friday 24th October for API Platform, ChatGPT Enterprise, and ChatGPT Edu customers, and a new agreement with the UK Ministry of Justice will provide 2,500 employees with ChatGPT Enterprise access. OpenAI |
| DEPLOYif you work in life sciences Claude for Life Sciences. Claude Sonnet 4.5 scores 0.83 on Protocol QA against a human baseline of 0.79, and Anthropic is adding connectors to Benchling, BioRender, PubMed, Scholar Gateway, Synapse.org, and 10x Genomics plus a single-cell-rna-qc Agent Skill. Anthropic |
| WATCHif you depend on SBERT embeddings Sentence Transformers. Sentence Transformers is transitioning from the UKP Lab at TU Darmstadt to Hugging Face, with over 16,000 models on the Hub serving more than a million monthly unique users, and Tom Aarsen continuing as maintainer. Hugging Face |
| DEPLOYif you download models from the Hub VirusTotal scanning. Every one of the 2.2M+ public model and dataset repositories on the Hugging Face Hub is being continuously scanned with VirusTotal, comparing file hashes against its threat-intelligence database without sharing raw file contents. Hugging Face |
| WATCHif you build document pipelines OCR open models. Hugging Face published a guide comparing open OCR models, with Chandra scoring 83.1, OlmOCR-2 scoring 82.3, dots.ocr scoring 79.1, and DeepSeek-OCR scoring 75.4 on the OlmOCR Benchmark. Hugging Face |
Tools and products
| DEPLOYif you run Next.js Server Actions Server Actions firewall. Vercel Firewall and Observability Plus now support custom rules targeting specific Server Action names, such as rate limiting app/auth/actions.ts#getUser to 100 requests per minute per IP address, starting with Next.js 15.5. Vercel |
| DEPLOYif you pay for Vercel and build monorepos Turbo build machines. Turbo build machines are available for all paid plans with 30 vCPUs and 60GB of memory, aimed at Turbopack builds and large monorepos that run tasks in parallel. Vercel |
| DEPLOYif you run long agentic tasks Sandbox extendTimeout. Vercel Sandbox now has an extendTimeout method to extend a running sandbox beyond its initial timeout, with Pro and Enterprise plans supporting up to 5 hours and Hobby up to 45 minutes. Vercel |
| DEPLOYif you run microfrontends on Vercel Microfrontends preview links. Preview links between microfrontends projects now serve all routes in the group, eliminating 404s and automatically linking deployments built from the same commit or branch. Vercel |
Breakthroughs
| WATCHif you work with geospatial data Google Earth AI. Google Earth AI introduces new Remote Sensing Foundations and Population Dynamics Foundations models, and a Gemini-powered Geospatial Reasoning agent that achieved 0.82 accuracy on a Q&A benchmark versus 0.50 for Gemini 2.5 Pro. Google Research |
| WATCHif you track quantum computing Quantum Echoes. Google Quantum AI demonstrated a verifiable quantum advantage with the Quantum Echoes algorithm on the Willow chip, where experiments took approximately 2 hours versus an estimated 13,000 times longer on a classical supercomputer. Google Research |
| WATCHif you need private synthetic data Private synthetic albums. Google Research introduced a method for generating differentially private synthetic photo albums using an intermediate text representation and hierarchical generation, tested on the YFCC100M dataset. Google Research |
| WATCHif you classify scientific images Gemini astronomy. Gemini achieved 93% accuracy across three astronomical datasets by learning from just 15 annotated examples per survey, and improved to 96.7% on MeerLICHT through an iterative human-in-the-loop process. Google Research |
Left out
- Vercel Agent investigations changelog: Duplicate of the Vercel Agent lead item with the same product facts.
- Vercel Agent blog post: Already covered as a lead; the changelog item adds only pricing details already captured.
- Agent investigations: the publisher's page does not show it available yet
- OpenEnv Hub: the publisher's page does not show it available yet
- ChatGPT WhatsApp exit: a figure in the draft could not be found on the publisher's own page
- ChatGPT Atlas: it did not carry a number that moved
- AI Sheets vision: it did not carry a number that moved
- AI Chat in docs: it did not carry a number that moved
- Vercel Python SDK: it did not carry a number that moved
- Workflow Development Kit: it did not carry a number that moved
The pattern
The week splits into two movements. On one side, production systems arrive: Mistral AI Studio packages Observability, Agent Runtime, and AI Registry for building, evaluating, and running AI in production, Vercel Agent runs simulated builds inside a Vercel Sandbox to verify pull request recommendations, Vercel Firewall and Observability Plus target specific Server Action names like app/auth/actions.ts#getUser at 100 requests per minute per IP address, and Turbo build machines offer 30 vCPUs and 60GB of memory for Turbopack builds and large monorepos. On the other side, consumer surfaces shift under constraint: when company knowledge is on, ChatGPT cannot search the web or create charts and images, and you must manually select it for each new conversation to get the full depth and citations.
The same tension appears in the numbers. Claude Sonnet 4.5 scores 0.83 on Protocol QA against a human baseline of 0.79, Gemini achieves 93% accuracy across three astronomical datasets from just 15 annotated examples per survey, and Google Earth AI's Gemini-powered Geospatial Reasoning agent reaches 0.82 accuracy versus 0.50 for Gemini 2.5 Pro, while Chandra scores 83.1, OlmOCR-2 scores 82.3, dots.ocr scores 79.1, and DeepSeek-OCR scores 75.4 on the OlmOCR Benchmark. Yet access remains gated: Mistral AI Studio is a private beta, Vercel Agent has only Code Review and Investigations, and LeRobot v0.4.0 launches a plugin system for third-party hardware while Datasets v3.0 introduces chunked episodes and streaming. The pattern is not capability but admission: every new production surface arrives with a manual switch, a beta door, or a named skill you must select.