Welcome to this week’s roundup of AI news that actually matters if you code with AI. As always: every item below was checked against its primary source, and anything we couldn’t confirm — or that won’t matter in a month — got cut. New here? Start with our AI for Developers hub, and catch up on last week’s roundup.
Models
GPT-5.6 is generally available — and you can finally use it
Last week we flagged GPT-5.6 as a preview you probably couldn’t touch yet. That changed on July 9, when OpenAI moved the whole family to general availability across ChatGPT, Codex, and the API. You get three tiers — Sol (flagship), Terra (balanced, roughly GPT-5.5-level), and Luna (fastest and cheapest) — at $5/$30, $2.50/$15, and $1/$6 per million input/output tokens. For coding specifically, OpenAI claims a new state of the art on the Artificial Analysis Coding Agent Index (80 with max reasoning) and on Terminal-Bench 2.1 and DeepSWE, with early praise from Cursor, Cognition, and Qodo in the launch post. Two things worth trying: ultra, a setting that fans work out across four parallel agents by default, and Programmatic Tool Calling in the Responses API, which lets the model write a small program to coordinate tools instead of round-tripping every result through the model — handy for token-heavy agent loops. If you benchmarked the preview and shelved it, now’s the time to wire it into real workflows.
Tools & pricing
A big Copilot month lands in VS Code
GitHub published its June 2026 Copilot roundup for VS Code (versions 1.123–1.127) on July 8, and it’s heavy on agent workflow. You can now run agent sessions side by side and split one session into multiple focused chats — implementation in one, tests and docs in another — while keeping the whole task in one place. Autopilot got more independent and better at knowing when a task is actually done, session cost is now visible across a whole chat (including delegated subagents), and you can browse and install model providers straight from the Marketplace. Rounding it out: 1M-token context windows with compatible Anthropic and OpenAI models, and pull requests generated from session context. If you manage your default coding stack, our guide to the best AI coding tools in 2026 is a good place to slot these in.
The GitHub Copilot desktop app opens to everyone
On July 7 GitHub made its Copilot desktop app available on every plan, including Copilot Free and GitHub Education, for macOS, Windows, and Linux. The more interesting detail for tinkerers: you can bring your own key and run agent sessions against your own model provider with no Copilot subscription at all. It’s a low-friction way to try agent-driven development from the desktop without committing to a paid tier.
Agents & frameworks
Codex shows up as an agent provider inside JetBrains
If you live in IntelliJ, PyCharm, or another JetBrains IDE, GitHub’s July 7 update adds Codex as a selectable agent provider in public preview — install the Codex CLI, point Copilot at it, and pick it from the agent picker alongside the built-in and Claude agents. The same release adds real substance around it: Hooks and MCP server management directly in Agent Customizations (including workspace-level servers via .github/mcp.json), approval modes for Copilot CLI sessions, permission modes and debug logs for the Claude agent, and Inline Chat going generally available. It’s a meaningful step toward mixing and matching agents without leaving your editor.
Open-source projects
Mistral’s Leanstral 1.5: open-weight proof engineering for Lean 4
A little outside the usual coding-assistant beat, but genuinely useful if you touch formal verification: Mistral released Leanstral 1.5, an Apache-2.0 model (119B total, ~6B active) built for Lean 4, the proof assistant used in software verification. Weights are on Hugging Face and there’s a free API endpoint. Beyond saturating math benchmarks, the more relevant result for developers is that its automated pipeline flagged 47 violated properties across 57 repositories and confirmed 11 real bugs — five of them previously unreported, including an integer-overflow edge case in a varint library that fuzzing had missed. If you’ve been curious whether formal methods are practical yet, this is a low-cost, fully open way to experiment.
Worth a look
Two smaller notes. Kimi K2.7 Code — the open-weight model that joined the Copilot picker earlier this month — is now available for Copilot Business and Enterprise, so teams on managed plans can try it too (an admin has to enable it). And the broader theme this week is pricing pressure: GPT-5.6’s GA tiers and its efficiency claims reset the cost math again. Our guide on cutting your LLM API costs covers the durable tactics — caching, routing, effort levels — that survive every reprice.
That’s it for this week. Everything above links to a primary source — no rumors, no vaporware. See you next Friday.