Opus 4.8 orchestrates parallel subagents, Google I/O delivers Gemini 3.5 Flash and WebMCP, MCP announces its stateless rewrite, and Codex moves to phone and browser.
Every major May release was about running many autonomous agents reliably and cheaply, not about raw benchmark gains: Opus 4.8 with parallel subagents, Gemini 3.5 acting instead of just answering, Codex on the phone and in the browser, and MCP announcing its stateless rewrite. Every link goes to the primary source.
Anthropic's most capable model at unchanged Opus pricing, with the headline claim that it lets flaws in its own code pass unremarked about four times less often than Opus 4.7. It ships with Dynamic Workflows, where Claude Code plans and runs parallel subagents for codebase-scale work, and mid-conversation system entries to redirect long-running tasks.
Replaces GPT-5.3 Instant as the default for all ChatGPT users and lands in the API as chat-latest, with OpenAI reporting 52.5 percent fewer hallucinated claims on high-stakes prompts. If you pin chat-latest, your app's default model changed under you this month.
Gemini 3.5 Flash is generally available via the API, AI Studio, Antigravity, and Android Studio, and beats 3.1 Pro on coding; 3.5 Pro follows later. Gemini Omni adds any-to-any multimodal output starting with video, watermarked with SynthID.
Closed weights, API only at $2.50/$7.50 per million tokens, with a 1M-token context and strong agentic numbers (SWE-Pro 60.6, Terminal-Bench 2.0 69.7), built for long sessions chaining over 1,000 tool calls. Relevant as a cheap long-context agent backend if you accept a non-Western API dependency.
A dense 128B with 256k context under a modified MIT license, 77.6 percent on SWE-Bench Verified, self-hostable on four GPUs: the strongest open coding model for on-prem this month. The same post launches Vibe, parallel sandboxed cloud coding sessions with GitHub, Jira, and Sentry integrations.
Anthropic's developer conference introduced automated PR code review, CI auto-fix, security reviews, and async routines, so you wake up to ready-to-merge PRs. No new model; the pitch was orchestration of the existing ones.
Codex drives Chrome directly on macOS and Windows: testing web apps, reading context across tabs, and using DevTools, with per-site permission prompts. It closes the loop between "agent writes frontend code" and "agent verifies it renders".
All plans on iOS and Android in preview: monitor threads, review outputs, approve commands, and switch models from the phone. Approving a long-running agent's pending action no longer requires a laptop.
The RC makes the protocol core stateless, no more Mcp-Session-Id or initialize handshake, aligns auth with OAuth 2.1, and deprecates roots, sampling, and logging. Server authors have until the July 28 final to plan their migration.
Auto mode replaces permission prompts with background safety checks and now works on Pro with Sonnet 4.6. /usage attributes plan consumption to individual skills, subagents, plugins, and MCP servers, the first real answer to "what is eating my limits", and /code-review posts findings as inline PR comments.
Antigravity becomes a standalone desktop app with a CLI and SDK, and Managed Agents provision remote Linux environments with one API call. WebMCP, a proposed open standard letting websites expose structured tools to browser agents, starts an origin trial in Chrome 149.
Training models on the reasoning behind ethical principles generalized far better than training on aligned behavior alone; every Claude model since Haiku 4.5 now passes the agentic-misalignment eval cleanly. Directly relevant if your agents hold credentials or act unsupervised.
