GitHub Changelog · Oct 28, 2026 · 6:18 PM

GitHub Copilot Fusion goes GA, routing each task across models at Standard, High or Max effort

Announced at Universe on Oct 28, Fusion replaces Auto in the model picker and decides per prompt whether one model answers, a cheap model drafts and escalates, or one model critiques another; Fusion Max scored within two points of Claude Code on Opus 5.5 Max on Artificial Analysis's coding index at about half the cost per task. It's on all plans in VS Code, the Copilot app, CLI and JetBrains, billed per model used, with 30% off for a limited time.

Read at github.blog

GitHub Copilot

AI coding assistance across editors, GitHub, and the CLI

More news

View all

Copilot Memory goes GA on paid plans, and opt-in Dreaming curates memories and proposes skill PRs

Announced at Universe on Oct 28, Memory now carries your preferences and repo conventions across the Copilot app, CLI, VS Code, cloud agent and code review. Dreaming, in public preview and billed in AI credits, reflects on past sessions to update memories and opens pull requests to improve checked in skills and instruction files…

GitHub Changelog 2 hours ago

GitHub Copilot app adds Projects for multi-repo goals and desktop computer use, both in preview

Announced at Universe on Oct 28, Copilot Projects let you give the app a goal plus context like repos, Figma mocks or Slack threads, and it runs as many background agents as needed, returning PRs for review. Computer use, in public preview on macOS and Windows, lets Copilot click and type in desktop apps with no API after you approve it (Settings Computer Use)…

GitHub Changelog 3 hours ago

Akamai Agent Sandbox runs agent-written code in nested WebAssembly with outbound access blocked by default

In beta since Oct 27, your agent runs on Akamai Functions via the Spin SDK, and the JavaScript or Python it writes runs in a second, tighter Wasm sandbox with limits on compute, memory and network; the tech preview suits short steps of about a minute. Akamai also opened a beta of Akamai Inference, OpenAI compatible serverless or dedicated hosting for 12 open weight models including gpt oss 120b and Gemma 4…

Akamai Newsroom 4 hours ago

Vercel AI SDK 7.0.154 lets you steer a running agent with a queued message at the next step

Released Oct 26, experimental createSteering() plugs into generateText , streamText and ToolLoopAgent , and steer(message) slips a user message in once the current step and its tools finish, resolving to delivered or a reason like finished or aborted . It never interrupts a running tool and doesn't work across processes yet; the same patch adds an onStepError callback for steps that fail before producing a result…

GitHub (vercel/ai) 5 hours ago

HashiCorp Vault will issue AI agents short-lived, scoped credentials it can revoke mid-task

Announced Oct 27 at IBM TechXchange, Vault ties each agent to an IBM Agent Identity record (owner, lifecycle, granted authority) and hands it short lived, least privilege credentials, revoking the session in real time if the agent tries something outside its permissions. Vault's agent model caps access by the human's policy, an agent ceiling policy, and per task authorization requests (HashiCorp)…

IBM Newsroom 6 hours ago

Spotted something we missed? Start a thread.