Most weeks in AI tooling are noise. This one wasn’t — not entirely. A new top-end Claude model turned up in the tools you already use, a cheaper Google model slipped into Copilot, and the protocol a lot of us now build on quietly changed shape. Everything here is checked against its primary source. Links included, so you don’t have to take my word for it.
New models showing up in your editor
Claude Opus 5 arrives
Anthropic shipped Opus 5, and it hit the tools developers actually work in this week. In Claude Code (v2.1.219, July 24) it’s now the default Opus model — 1M-token context, plus a /fast mode. Pricing is spelled out on the changelog; check it there before you kick off a big run, because per-token math adds up quickly on this tier.
Same day, it showed up in GitHub Copilot’s model picker, billed at provider API list price under usage-based billing. GitHub’s pitch is long-running agent work: autonomous code changes, regression checks, jobs that string several tools together. The sweet spot is a model that makes a targeted edit and then goes back to verify its own work. One caveat worth knowing up front — Copilot says Opus 5 carries extra safeguards around high-harm cyber content, so a security-adjacent prompt might get blocked. If it does, add some benign context or switch models.
Gemini 3.6 Flash joins Copilot
Opus 5 is the expensive heavy lifter. Gemini 3.6 Flash (July 21) is the other end of the dial. Google’s newest Flash model is rolling out in Copilot with configurable reasoning effort and parallel tool use, and Google claims better task-completion rates and token efficiency than Gemini 3.5 Flash on coding and agent tasks. It’s provider-list priced, and it lands in the picker across VS Code, Visual Studio, the Copilot CLI, the cloud agent, JetBrains, Xcode, and Eclipse. Try it as a cheap default for agent loops — the steps where you don’t need to spend big-model money on every turn.
The MCP spec is going stateless — mark July 28
If you build anything on the Model Context Protocol, this is the item to read twice. The spec goes stateless on July 28, 2026. The GitHub MCP Server already runs the new version, ahead of the official release.
What actually changes: sessions and the initialize handshake are gone. Clients connect faster, can run the handshake in parallel, and servers get far easier to scale sideways — there’s no per-session state to hang onto anymore. There’s also a new “extensions” concept, which is what makes things like MCP apps and Enterprise Managed Auth possible (both already work in VS Code), and you’ll see more remote servers doing elicitation over multi round-trip requests. Here’s the reassuring part. Every tier-1 SDK kept backwards compatibility and already shipped beta support, so most people won’t have to touch a thing. Newer to this? Our beginner’s guide to MCP walks through the basics.
Agents and workflows
Codex gets a voice — and multi-folder projects
OpenAI’s Codex lives inside the ChatGPT desktop app now, and it picked up two things on July 23. First, ChatGPT Voice, powered by GPT-Live. You talk through the work and ask it to start, check, or steer tasks running in other threads across Chat, Work, and Codex. On macOS there’s a Screen context toggle that hands it a snapshot of your frontmost window, so it can see what you’re staring at. Second, and smaller but genuinely useful: local projects can now hold multiple folders, with one marked as the primary folder for chats and Git operations. If your “project” was really three repos in a trench coat, that one’s for you. Voice ships on the Plus, Pro, Business, Edu, and Enterprise plans.
Copilot’s cloud agent goes GA for Linear
GitHub’s asynchronous cloud agent is generally available inside Linear as of July 23. Assign it an issue. It reads the ticket, opens a draft pull request, and does the work in a throwaway environment on GitHub Actions, streaming progress back to the Linear timeline and pinging you for review when it’s finished. The GA release adds real control: pick the model, point it at a custom agent from your repo, set the base and working branches, or steer a run by mentioning it in a comment. It’s the pattern from our AI agents explainer, made concrete — the tracker is the interface, and the agent lives in CI.
Claude Code tightens up its subagents
Across a string of releases this week (v2.1.215 through v2.1.219), Claude Code made background agents behave. /code-review runs as a background subagent now, so its output stops burying your conversation. Subagents can nest three levels deep. And there’s finally a cap on how many run at once — 20 by default — so one message can’t spawn an unbounded swarm of workers in the background. /deep-research also stopped kicking off on its own; it waits until you ask. No single one of these is a headline. Together, they make a long multi-agent session something you can actually trust.
Worth a look
Rethinking your setup after all that? Our Best AI Coding Tools in 2026 comparison is a decent gut-check on where each tool stands. Last week’s roundup is here if you missed it, and the running list lives on the Coding with AI hub. See you next week.