AI Coding News: July 30, 2026 — GitHub Models Goes Dark, Kimi K2.7 Lands in Copilot

July 30, 2026 is the kind of day that matters less for what shipped and more for what disappeared. GitHub Models — the free model playground and inference API GitHub built into every repository — goes dark today after a months-long wind-down. At the same time, the rest of the AI coding stack kept moving: Kimi K2.7 Code graduated to general availability across Copilot's Business and Enterprise tiers, Claude Code's July stabilization branch kept shipping reliability and accessibility fixes, OpenAI's Codex CLI made Remote Control generally available, and Cursor rolled out cost-aware model routing. Here's everything worth knowing.

GitHub Models Officially Retires Today

GitHub confirmed on July 1 that GitHub Models is being fully retired on July 30, 2026, closing the playground, model catalog, inference API, and bring-your-own-key (BYOK) endpoints for every customer — not just the trial users who were cut off back in June. Two scheduled brownouts on July 16 and July 23 gave teams a preview of the outage ahead of today's permanent shutdown. Anyone still pointing playground prompts, eval harnesses, or production calls at the Models inference API needed to have migrated to GitHub Copilot or another provider before today's cutover — there's no grace period after this point.

GitHub Copilot: Kimi K2.7 Code Reaches Business and Enterprise

Kimi K2.7 Code, Moonshot AI's open-weight coding model, is now generally available in Copilot for Business and Enterprise plans, expanding a rollout that started July 1 for Pro, Pro+, and Max subscribers. It's the first open-weight model selectable in Copilot's model picker, hosted by GitHub on Azure and billed at provider list pricing under usage-based billing. The model ships off by default for Business and Enterprise — plan admins have to enable the Kimi K2.7 Code policy before anyone in the organization can select it.

Claude Code: Nested Subagents, Background Reviews, and a Pile of Fixes

Anthropic's July stabilization releases, tracked in the official Claude Code changelog through v2.1.220 on July 24, focused less on flashy features and more on making long agentic sessions safer and less flaky:

• Subagents can now spawn nested subagents up to depth 3 by default, up from a hard limit of 1.
• The /code-review command now runs as a background subagent, so a review no longer fills up the main conversation.
• New session-wide caps default WebSearch calls and subagent spawns to 200 each, tunable via environment variables, to stop runaway loops.
• Screen-reader announcements for deleted text and a live elapsed-time counter on collapsed tool calls improve accessibility for --ax-screen-reader mode.
• Fixes landed for corrupted Windows paths with unicode-prefixed segments, reversed right-to-left text rendering in VS Code, and several long-session memory leaks.

OpenAI Codex CLI: Remote Goes GA, Smarter Guardrails

Codex Remote is now generally available, letting developers drive Codex from the ChatGPT mobile app against a connected Mac or Windows host via authenticated one-to-one QR pairing — any pairing older than June 8, 2026 needs to be redone. July's releases also shipped a richer codex doctor diagnostic command, vim-mode text-object editing, tighter detection of dangerous shell commands (including more rm variants) with clearer rejection messages, and support for routing auth and API traffic through corporate PAC/WPAD proxies.

OpenCode v1.18.9 and Cursor's New Cost-Aware Modes

OpenCode, the open-source, provider-agnostic coding agent, pushed v1.18.9 on July 28 with desktop stability fixes and MCP compatibility patches, continuing a fast release cadence that's kept pace with model and provider churn all year. Cursor, meanwhile, spent July rolling out three optimization modes — Intelligence, Balance, and Cost — that let teams trade frontier quality for lower token spend, plus admin controls to restrict modes and models per group, a plan-before-execution flow that surfaces the agent's intended steps before it acts, and, as of July 29, a native iPad app.

The Takeaway

The throughline across today's news is consolidation: GitHub is retiring its standalone inference playground in favor of funneling everyone through Copilot, Copilot itself is diversifying its model roster with cheaper open-weight options like Kimi K2.7, and the agentic tools — Claude Code, Codex CLI, OpenCode, Cursor — are all converging on the same concerns: safer autonomy, cost visibility, and fewer surprises in long-running sessions. If anything you own still points at the GitHub Models inference API, today was the deadline to already be off it.

References & Further Reading

• GitHub Models is being fully retired on July 30, 2026 — GitHub Changelog
• Kimi K2.7 now available for Copilot Business and Enterprise — GitHub Changelog
• Kimi K2.7 Code is now available in GitHub Copilot — GitHub Changelog
• Claude Code Changelog — Anthropic
• Codex Changelog — OpenAI Developers
• OpenCode Changelog
• Cursor Changelog