Summary

July 2 was dominated by a frontier-model reset and the fast-industrializing agent stack around it. Anthropic made Claude Sonnet 5 the Claude Code default with a native 1M-token context, restored Fable 5 to global availability, and consolidated the Claude API into three higher-limit tiers, while OpenAI previewed a tiered GPT-5.6 (Sol/Terra/Luna) that set a new Terminal-Bench SOTA. Coding agents kept spreading across new surfaces — Claude Code 2.1.198 took Chrome GA and pushed background agents toward auto-opened PRs, Cursor shipped an iOS app, and the Windsurf-to-Devin transition completed with Cascade's end-of-life. Underneath the products, agent-platform plumbing and interoperability standards advanced (Vercel AI SDK 7, GitHub's ARD-based agent finder, Devin's ACP adoption), and a distinct thread of supply-chain and trust hardening emerged around installable skills and plugins (ZeroClaw ZIP limits, QwenPaw version routing, Qwen Code credential redaction). Anthropic and Vercel also pushed AI into vertical, standards-aware surfaces with Claude Science and v0 Design Systems 2.0.

Key themes

  • Frontier-model reset and pricing competition: Claude Sonnet 5 becomes the Claude Code default with a 1M-token context, OpenAI previews tiered GPT-5.6 (Sol/Terra/Luna) with a Terminal-Bench SOTA, Anthropic restores Fable 5 globally and consolidates the Claude API into three higher-limit tiers.
  • Coding agents expand to new surfaces and more autonomy: Claude Code 2.1.198 makes Chrome GA and moves background agents toward committing and opening draft PRs, Cursor launches on iOS, Qwen Code adds configurable auto-compaction, and the Windsurf-to-Devin transition completes as Cascade reaches end-of-life.
  • Agent-platform infrastructure and interoperability standards mature: Vercel ships AI SDK 7 as a production agent platform, GitHub launches an agent finder implementing the open Agentic Resource Discovery (ARD) spec, Devin Desktop adopts the Agent Client Protocol (ACP), and LangChain's OpenWiki targets agent-maintained codebase context.
  • Supply-chain and trust hardening for installable skills and plugins: ZeroClaw bounds skill ZIP extraction against zip bombs, QwenPaw routes plugin-market downloads by major version, and Qwen Code redacts credentials from daemon worker logs and adds leader approval for plan-required spawned agents.
  • AI moves into vertical and design-system-aware surfaces: Anthropic launches Claude Science as an auditable research workbench, and Vercel's v0 Design Systems 2.0 lets AI generation conform to a team's own components and design tokens.

Notable items

  • Claude Sonnet 5 becomes the default model in Claude Code (v2.1.197) with a native 1M-token context window and promotional pricing of $2/$10 per million tokens through August 31.
  • OpenAI previews GPT-5.6 in three tiers — Sol ($5/$30), Terra ($2.50/$15), Luna ($1/$6) — with Sol setting a new Terminal-Bench 2.1 SOTA, an 'ultra' subagent mode, and Cerebras runs up to 750 tokens/sec slated for July.
  • Claude Code 2.1.198 makes Claude in Chrome generally available, adds agent_needs_input/agent_completed hooks and a /dataviz skill, and has finished background agents commit, push, and open draft PRs for worktree code work.
  • Anthropic launches Claude Science, a vertical AI workbench integrating PubMed, Jupyter, R, and cluster terminals with auditable artifact history.
  • Vercel ships AI SDK 7, adding provider-agnostic reasoning control, typed scoped context, MCP Apps, a durable WorkflowAgent, sandboxed command execution, a HarnessAgent layer for Claude Code/Codex/OpenCode, and OpenTelemetry GenAI conventions.
  • The Windsurf-to-Devin transition completes: Cascade support ended July 1, Devin Local (a token-efficient Rust rewrite) is the default local agent, and Devin Desktop adopts the open Agent Client Protocol.
  • GitHub ships an agent finder across all Copilot plans that indexes MCP servers, skills, and tools and returns ranked matches, implementing the open Agentic Resource Discovery (ARD) specification.
  • Anthropic restores Fable 5 to global availability after export controls are lifted and consolidates the Claude API into three tiers (Start, Build, Scale) with raised rate limits.
  • Cursor for iOS leads Product Hunt's AI launches with 565 votes, signaling demand for asynchronous, mobile supervision of coding agents.
  • Trust hardening across agent ecosystems: ZeroClaw bounds skill ZIP extraction against zip-bomb inflation, QwenPaw routes plugin downloads by major version to prevent incompatible installs, and Qwen Code redacts API keys from daemon worker stderr.

Source coverage

Source rows used: 17