AI Digest

Daily AI update Digest

Curated updates from GitHub Copilot, Claude Code, Cursor, OpenAI Codex, and more.

343 items

OpenCode v1.18.11 fixes MCP and desktop reliability

Nice-to-have
Sun, Aug 2, 2026OpenCode changelog
  • MCP SSE connections no longer get stuck reconnecting after server errors.
  • Provider configs using custom or interleaved reasoning fields now work correctly.
  • Desktop now opens external links in the system browser and fixes stale tabs, file-tree clipping, and debug-gutter alignment.
  • Changes reduce agent hangs and provider compatibility failures.

Codex CLI 0.147.0-alpha.4 enters testing

Nice-to-have
Sun, Aug 2, 2026OpenAI Codex GitHub releases
  • Codex 0.147.0-alpha.4 is available as a pre-release build.
  • Compared with stable 0.146.0, the branch contains 174 commits across 923 files from 43 contributors.
  • The release page provides no curated user-facing changelog, so exact behavioral changes remain unclear.
  • Large diff size makes isolated compatibility testing prudent before any production rollout.

Gemini 2.5 Pro and Gemini 3 Flash deprecated

Breaking change
Sat, Aug 1, 2026GitHub Copilot Changelog
  • Gemini 2.5 Pro and Gemini 3 Flash are deprecated across Copilot Chat, inline edits, agent modes, and code completions.
  • Recommended replacements are Gemini 3.1 Pro and Gemini 3.6 Flash.
  • Enterprise administrators may need to enable replacement models through Copilot model policies.
  • Existing workflows and integrations using either model should migrate immediately.

Enterprise teams model policy targeting in public preview

New capability
Sat, Aug 1, 2026GitHub Copilot Changelog
  • Enterprise administrators can now grant optional Copilot models to specific enterprise teams.
  • Policies support enabled, disabled, and optional model states with least-restrictive access evaluation.
  • Enabling Enterprise teams mode replaces organization-level model settings during the preview.
  • Most enterprise customers can opt in starting August 3, 2026.

GPT-5.6 prices slashed: Luna drops 80%, Terra drops 20%, Sol Fast mode launched

Breaking change
Fri, Jul 31, 2026OpenAI Blog
  • GPT-5.6 Luna now costs $1.40/MTok combined ($0.20 input, $1.20 output), undercutting Google's Gemini 3.5 Flash-Lite and making frontier-grade work viable for high-volume agentic pipelines
  • GPT-5.6 Terra drops from $17.50 to $14/MTok combined, matching Gemini 3.1 Pro Preview at less than half the per-intelligence cost vs. GPT-5.4
  • Sol Fast mode delivers up to 2.5× throughput at $70/MTok combined ($10 input, $60 output) — premium for latency-sensitive agent orchestration, no intelligence change
  • Sol autonomously rewrote production kernels and ran hundreds of experiments to cut serving cost by 20% and boost token-generation efficiency 15%, creating a compounding efficiency loop

Copilot code review: Agent skills and MCP now generally available

New capability
Fri, Jul 31, 2026GitHub Blog
  • Agent skills (`.github/skills/SKILL.md`) and MCP server connections for Copilot code review are GA for all Pro, Pro+, Business, and Enterprise users — custom review logic without leaving the PR workflow
  • MCP tool calls in code review are read-only; existing Copilot cloud agent MCP configs auto-apply to code review
  • GitHub and Playwright MCP servers enabled by default; attribution tags on review comments show when skills or MCP context drove the feedback

GitHub Copilot in VS Code July 2026: Multi-chat sessions, vision GA, built-in dictation

New capability
Fri, Jul 31, 2026GitHub Blog
  • Multi-chat sessions let a single agent hold multiple related conversations with independent history, titles, and models — fork peer chats from any point without losing context
  • Copilot vision is now GA: paste images and PDFs into chat; BYOK models work in the Agents window
  • Experimental built-in dictation with optional LLM transcript cleanup for voice-driven coding
  • Agents window redesigned with faster diff review, Git worktree support for isolated sessions, subagent tracking (model, elapsed time, active tool call), and keyboard-driven chat navigation

Codex CLI 0.146.0: Agent Plugins, named sessions, thread forking

New capability
Fri, Jul 31, 2026OpenAI Developers / GitHub Releases
  • Agent Plugins manifests and workspace plugin publishing with support for Amazon Bedrock and Claude Code marketplaces — extend Codex with pluggable capabilities
  • Name sessions with `/new` or `/clear`, pin important threads, and switch between side conversations without closing them; fork threads with paginated history
  • Executor-provided skills and standalone web search for compatible custom model providers
  • Proxy support unified across authentication, plugin downloads, MCP, remote exec, WebSockets, and LM Studio; MCP connections auto-reconnect on config changes

ChatGPT desktop 26.727: Multi-repository code review, image editing, browser upgrades

New capability
Fri, Jul 31, 2026ChatGPT Learn / Changelog
  • Multi-folder projects now support reviewing changes across all repositories at once — see diffs per repo without switching between review views
  • Expanded image viewer with Focused and Canvas views, inline comments across images, and targeted edit requests without leaving the conversation
  • Built-in browser and Chrome extension improvements for faster context gathering

Grok 4.5 is now available in GitHub Copilot

New capability
Wed, Jul 29, 2026GitHub Changelog
  • Grok 4.5 is rolling out in GitHub Copilot as a reasoning model for agentic coding and complex multi-step workflows.
  • Developers get up to a 500,000-token context window, text and image inputs, and low, medium, or high reasoning effort controls.
  • GitHub says the model performed strongly on terminal coding tasks in VS Code and Copilot CLI, especially with parallel tool use and direct action.
  • Business and Enterprise admins must explicitly enable the Grok 4.5 policy, and usage is billed at provider list pricing.

OpenCode v1.18.8 and v1.18.9 improve MCP compatibility and desktop stability

Migration recommended
Wed, Jul 29, 2026OpenCode Changelog
  • v1.18.8 improves compatibility with newer MCP servers and OAuth flows, including reconnecting MCP servers after expired SDK sessions.
  • OpenCode now stops sending deprecated sampling defaults to newer Gemini models, reducing provider compatibility risk.
  • v1.18.9 restores compatibility with legacy MCP SDK clients, which matters for teams straddling old and new MCP integrations.
  • Desktop fixes target navigation crashes, session-list loading, provider settings, and an opt-in V2 sidecar backed by the bundled CLI service.

GitHub Copilot app usage metrics now expand across report rollups

Nice-to-have
Wed, Jul 29, 2026GitHub Changelog
  • Copilot app activity is now attributed to individual users in enterprise-user and organization-user usage reports.
  • New fields expose app session counts, request counts, prompt counts, token usage, model/language breakdowns, and Copilot-app-only daily active users.
  • Code generation, acceptance, lines added, and lines deleted totals now include Copilot app activity.
  • The API change is backward compatible, but analytics pipelines should account for new `copilot_app` breakdown entries.

Cursor Start launches for developers in India

New capability
Wed, Jul 29, 2026Cursor Blog
  • Cursor introduced Cursor Start, a ₹649/month plan for developers in India with INR billing and UPI support.
  • The plan includes access to Grok 4.5 and Composer, with more agent requests than the Free plan.
  • Start includes always-on cloud agents across desktop, web, iOS, and CLI, plus plugins, MCP servers, hooks, and skills.
  • Pro remains the higher tier for access to every major model, Bugbot, Auto mode, Automations, Cursor SDK, and on-demand usage.

Enterprise managed settings in the GitHub Copilot app and Copilot cloud agent

Migration recommended
Tue, Jul 28, 2026GitHub Changelog
  • GitHub Copilot enterprise managed settings now apply to the Copilot app and Copilot cloud agent.
  • Admins can enforce plugin availability, marketplace access, approval-prompt behavior, and default auto model selection from `managed-settings.json`.
  • Existing Copilot CLI and VS Code managed-settings deployments carry over with no new setup for covered keys.
  • This matters for teams scaling agents across desktop, CLI, VS Code, and cloud work without policy gaps.

Manage GitHub Copilot app access with a dedicated policy

Migration recommended
Tue, Jul 28, 2026GitHub Changelog
  • The GitHub Copilot app now has its own enterprise and organization access policy.
  • Access no longer depends on the Copilot CLI policy, so admins can enable or disable each client independently.
  • The policy defaults to enabled everywhere, with options to disable globally or delegate to organization admins.
  • This gives platform teams cleaner rollout control for Copilot app adoption.

OpenCode v1.18.6 fixes repository cache and Desktop API compatibility issues

Nice-to-have
Tue, Jul 28, 2026OpenCode GitHub Releases
  • OpenCode fixed branch-specific repository caches so refreshing one ref no longer moves another branch checkout.
  • Desktop compatibility improved across directory, project, session, and terminal flows for the newer client API.
  • Legacy MCP state now refreshes correctly when opening V1 workspaces.
  • Provider lists now refresh after provider connection or OAuth completion on V1 servers.

Image Generation Models on OpenRouter

Nice-to-have
Tue, Jul 28, 2026OpenRouter Blog
  • OpenRouter documented the image-generation contract for `POST /api/v1/images` and vision analysis through `/chat/completions`.
  • The guide clarifies that image generation and image understanding share one base URL, API key, billing account, and routing layer.
  • Provider routing for image calls supports controls such as `order`, `only`, `ignore`, `sort`, and `allow_fallbacks`.
  • Developers get clearer migration guidance for avoiding “no endpoints found” errors when mixing text-only and vision-capable models.

How to Evaluate LLM Provider Performance Across Latency, Throughput, and Uptime

Nice-to-have
Tue, Jul 28, 2026OpenRouter Blog
  • OpenRouter published a practical guide for evaluating provider endpoints behind the same model slug.
  • The article separates latency, throughput, uptime, and quantization as distinct production variables.
  • It recommends using p90 and p99 metrics instead of averages for user-facing AI apps and coding agents.
  • It maps evaluation results to routing controls such as `provider.sort`, `quantizations`, `ignore`, and fallback model arrays.

Claude Code 2.1.220

Nice-to-have
Sun, Jul 26, 2026Anthropic Claude Code changelog
  • Anthropic lists Claude Code 2.1.220 as a maintenance release focused on bug fixes and reliability improvements.
  • No new model, API, workflow, or migration detail is called out, so developer impact appears limited to stability.
  • Teams running Claude Code in long-lived agent sessions should still update because reliability fixes matter most under sustained tool use.
  • This does not change the separate July 24 Opus 5-related Claude Code surface.

Claude Opus 5 is now available in GitHub Copilot

New capability
Sat, Jul 25, 2026GitHub Changelog
  • Claude Opus 5 is now selectable across Copilot surfaces including VS Code, Visual Studio, Copilot CLI, Copilot cloud agent, github.com, mobile, JetBrains, Xcode, and Eclipse.
  • GitHub positions it for complex, long-running coding tasks that need careful reasoning, tool use, and multi-step execution.
  • Business and Enterprise admins must enable the Claude Opus 5 policy before teams can use it.
  • The model is billed at provider API list price under Copilot usage-based billing, so heavy agent workflows may need budget review.