Need to know
- GitHub Copilot enters public preview for computer use, letting it click and type inside desktop apps on macOS and Windows.
- xAI rolled out Grok 4.7 as its base model across all modes and added a proactive Primary Bot and Agent Dashboard to Grok Build.
- Meta Muse hit 5 million downloads in 3 weeks, faster than ChatGPT, Claude, or Grok reached that mark.
- Anthropic launched Claude Code Mods, TypeScript hooks that can rewrite prompts, intercept tool calls, and replace built-in features.
- Cohere shipped Embed 5 with a shared vector space across Pro and Fast tiers, letting teams index with Pro and query with Fast without re-indexing.
New Releases
GitHub Copilot can now click, type, scroll, and navigate inside desktop applications on macOS and Windows, entering public preview on October 1 alongside Dynamic Workflows, a code-defined multi-agent orchestration feature.
- Computer use targets software with no API or CLI surface, requires explicit user activation, and ships a documented warning that on-screen content can trigger unintended actions.
- Dynamic Workflows define sequential and parallel agent steps in code with pause checkpoints, available on all Copilot plans via the CLI with the experimental flag.
xAI made Grok 4.7 the base model across every Grok mode on October 1 and simultaneously launched Primary Bot, a proactive assistant that volunteers tasks, and an Agent Dashboard for managing parallel coding sessions in Grok Build.
- Primary Bot watches for tasks it can take on before a user asks, then routes them to specialized Bots, shifting the interaction model from reactive to ambient.
- Agent Dashboard at /dashboard gives developers one screen to launch, view, and manage isolated coding worktrees in parallel, with Grok 4.7 also now available through Google Gemini Enterprise Agent.
Anthropic released Claude Code Mods in v2.1.287 on October 1, letting developers install TypeScript or JavaScript functions that intercept every prompt and tool call, rewrite them, and replace or disable any first-party Claude Code feature.
- Mods run unsandboxed with full machine access, can read and write files, and consume usage on behalf of the user, creating a meaningful supply-chain risk for any team that installs third-party mods without review.
- Enterprise guardrail is a sec-default mod that loads first on Teams and Enterprise plans to prevent user-installed mods from overriding security controls, but the protection is only as strong as that mod.
Cohere released Embed 5 on September 30 with a shared embedding space between its Pro and Fast tiers, so teams can index documents with Pro-quality vectors and serve live queries with Fast without rebuilding the index.
- Embed 5 Pro scores 85.8 on ViDoRe V3, leading Voyage 4 Large, Gemini Embedding 2, and OpenAI text-embedding-3-large according to Cohere's own benchmarks, with particular strength on FinanceBench.
- Deployment options include Cohere API, Microsoft Foundry, Amazon SageMaker, and private VPC via vLLM, with support for text, images, and mixed inputs across 100-plus languages at 128k context.
Funding
Mistral closed a roughly €3B round led by Samsung in September at a $24B valuation and paired it with an open-weight safety classifier anyone can run on a single consumer GPU, a robotics model, and a formal-verification tool, making the clearest lab-level argument yet that open weights plus safety tooling can beat closed APIs in regulated enterprise markets.
ZoomInfo launched Agent Teams, a multi-agent sales and marketing platform built on its DoubleO.ai acquisition and bundled at no extra cost into GTM Studio, signaling that the data-plus-agent bundle is becoming the competitive moat in B2B GTM rather than data alone.
Drata broke its third-party risk management module into a standalone product with continuous evidence-based vendor assessment on October 2, a structural move that lets the company compete directly with dedicated TPRM vendors rather than only upselling existing compliance customers.
Case Studies
Rakuten France deployed more than 80 Notion Custom Agents that check its backlog overnight and route exceptions before stand-up, part of Notion's report that one million agents have now been built on its platform.
- The workflow replaces manual pre-stand-up ticket review, surfacing stale items and missed dates automatically so engineers start the day on exceptions rather than discovery.
- Caveat: the accuracy data enterprises need before granting agents write access still comes from Notion, not from an independent audit.
A new study cited by Ethan Mollick found that frontier AI models are now faster and more accurate than junior accountants on medium-length, well-defined accounting tasks, a reversal from eighteen months ago when the same models scored below human performance.
- Scope is specific: the advantage holds on structured, well-defined tasks, not open-ended judgment work, which limits direct generalization to full accounting roles.
- The eighteen-month reversal is the significant data point: it suggests the capability gap is closing faster in knowledge-work verticals than most enterprise buyers have modeled.
Trending on X
- OpenAI vs Anthropic model leapfrog Developers are tracking each Anthropic and OpenAI drop in real time, arguing the two labs are now within 0.1 version points of each other with every release, and debating whether the advantage is durable for more than two weeks.
- Meta Muse traction versus OpenAI Dots The community is debating whether OpenAI's Dots can compete with Muse given Meta's reported 100 million free tokens per week per user, with Scale AI's Alexandr Wang amplifying Muse praise and analysts questioning whether that subsidy is sustainable.
- LLMs versus SLMs in agentic call routing An Nvidia claim that 40 to 70 percent of agent LLM calls don't need a full LLM is circulating widely, with developers discussing how to route formatting, tool-calling, and validation tasks to cheaper SLMs while reserving frontier models for planning.
- Karpathy on interpreting LLM outputs Andrej Karpathy posted that teams will spend significantly more time understanding what language models know, illustrated by a geography probe showing models have internalized global land-water boundaries from pretraining data alone.
- Chollet on LRMs versus base LLMs Francois Chollet argued the key distinction between pre-2025 base LLMs and modern language reasoning models is not tool use but a shift from transductive to inductive reasoning, prompting debate among researchers about what that means for benchmark design.