This brief covers AI news from 2026-10-31 UTC.
Money got tighter and models got cheaper on the same day, with OpenAI’s raise shrinking while its new computer-use model undercuts its own price list.
OpenAI trims funding target by $10B as tech stocks slide
What happened: The Financial Times reported, and Reuters relayed, that OpenAI cut its latest round target by $10 billion to about $20 billion, with its desired valuation down to $1.25 trillion. A UAE consortium including MGX is expected to be the largest participant, and the first close may land below $20 billion.
Why it matters: A smaller raise and a lower valuation change how much compute OpenAI can commit to, which feeds directly into API pricing and capacity for anyone building on its models.
OpenAI ships GPT-6.1 Terra for computer use at $1 input
What happened: OpenAI introduced GPT-6.1 Terra, a fast model for computer use, visual reasoning and everyday writing, available in ChatGPT Work, Codex and the API as gpt-6.1-terra at $1 per million input tokens and $5 per million output. OpenAI says that is half the price of GPT-5.6 Terra and GPT-6.1 Sol.
Why it matters: Cheap computer-use tokens make screen-driving agents affordable to run at volume, though OpenAI still points to Sol and Astra for the hardest coding work.
New lab Mirendil launches Fern with a 10M-token window
What happened: Mirendil, a new lab led by Behnam Neyshabur, released Fern 0.1, pretrained from scratch with up to 10 million tokens of context, roughly half a million lines of code. The team says 10M context is not yet as usable as 1M, and its own charts put Fern behind GLM 5.3 and Kimi K3 on agentic coding tests.
Why it matters: A 10M-token window is enough to hold a large codebase in one prompt, but the team’s own benchmarks say to test it on your repo before trusting it with agentic work.
Salesforce to buy Rune for a reported $375M
What happened: Salesforce agreed to acquire Rune Technologies, which builds AI software to help the U.S. military forecast supply needs and manage logistics from the field, for $375 million in cash and stock, per Resilience Media. Salesforce said its Missionforce defense unit has grown 80% year over year.
Why it matters: Salesforce keeps absorbing agent and vertical AI startups, so builders targeting its platform should expect more of the workflow to arrive pre-built.
Ottawa takes equity in Cohere as lab seeks C$7 billion
What happened: Canada said it will invest up to C$1.6 billion for equity in Cohere, which agreed to keep its headquarters in Canada as it seeks up to C$7 billion.
Why it matters: Sovereign equity stakes give AI labs capital outside the venture market, and Cohere’s Canadian anchor may shape where its enterprise models are hosted and sold.
Google Cloud launches Agent Compute Engine sandboxes
What happened: Google Cloud introduced Agent Compute Engine, a compute service providing hardware-isolated microVM sandboxes with sub-second activation and copy-on-write memory sharing, plus agent governance through Agent Gateway and Agent Registry. It also announced A5X instances on NVIDIA Vera Rubin NVL72.
Why it matters: Isolated, fast-starting sandboxes are the missing piece for running untrusted agent code in production without building your own isolation layer.
NVIDIA ships NemoClaw blueprints for self-hosted agents
What happened: NVIDIA published NemoClaw open blueprints that combine an agent harness, Nemotron models, NeMo optimization tools and the OpenShell secure runtime, supporting OpenClaw, Hermes Agent and LangChain Deep Agents. NVIDIA says LangChain tuned Deep Agents for Nemotron 3 Ultra to near-frontier performance at about 10x lower cost per run.
Why it matters: OpenShell enforces policy on file, network and credential access, so agents can touch real systems without the harness being the only guardrail.