Good morning. It's Monday, July 20th, 2026. Here's what's new in the agent platform world this week.

**OpenClaw 2026.7.1 Stable** — OpenClaw's July release hit npm stable with thousands of contributions across hundreds of contributors. It's a massive overhaul: brand-new Control UI with searchable sidebar, compact context ring, reasoning-effort controls, and gateway-backed session groups. Onboarding is completely redesigned. The official iOS, Android, and macOS apps got major updates. Expanded model support includes GPT-5.6, Tencent Hy3, and Meta Muse Spark 1.1. Two beta iterations followed — 2026.7.2-beta.3 focuses on remote coding sessions, native automation parity for mobile, safer channel operation to prevent Telegram disconnects on restart, and guided Control UI setup flow. A per-job dynamic cadence feature for cron jobs also landed, letting you set pacing bounds and one-shot check handling for individual scheduled tasks. July 18 patch 2026.7.1-2 tightens restart catch-up behavior, Codex budget accounting, and channel state reliability.

**Claude Fable 5 Returns July 20 — With Strings Attached** — After weeks of subscription limbo, Anthropic reversed course on Saturday and announced that Claude Fable 5 returns to Max and Team Premium plans starting tomorrow, July 20th. The catch: it's included at 50 percent of normal weekly limits. Pro and Team Standard subscribers get a one-time $100 usage credit, then shift to API billing. This is the permanent arrangement — no more temporary extensions. The move comes after OpenAI's GPT-5.6 Sol release intensified competitive pressure. Fable 5 had been offline for three weeks from mid-June to July 1st, then restored with usage-credit pricing. The tiered access strategy — flagship model available to top tiers, pay-as-you-go for mid-tier — is a significant departure from Anthropic's earlier "everything included" approach.

**Kimi K3 — Largest Open-Weight Model at 2.7 Trillion Parameters** — Chinese AI lab Moonshot released Kimi K3 on July 16th, and it's a statement. At approximately 2.8 trillion parameters (some sources report 2.7T), it's the largest open-weight language model available today — larger than any Llama, Gemma, or Mistral. Moonshot claims Kimi K3 outperforms Claude Opus 4.8 and GPT-5.5 across coding and math benchmarks, though it still trails Fable 5 and GPT-5.6 Sol overall. The Terminal-Bench 2.1 score hits 88.3 percent with the KimiCode harness at max reasoning. On Frontend Code Arena, Kimi K3 actually beats Claude Fable 5 — significant for a Chinese model under compute constraints. The open-weight release is promised by July 27th. Moonshot is also reportedly raising a fresh round of capital. This represents China's strongest push into frontier territory despite U.S. compute export controls.

**MCP Goes Stateless — The Biggest Protocol Change Yet** — The 2026-07-28 MCP specification release candidate dropped on July 16th, and it changes everything about how MCP servers deploy. The session handshake is gone. The Mcp-Session-Id header is gone. Instead, every request carries protocol version, client info, and capabilities in metadata, plus new Mcp-Method and Mcp-Name routing headers. Any server instance can handle any request — no more sticky sessions, no shared session stores. Just a plain round-robin load balancer. It also introduces MCP Apps for server-rendered UIs, the Tasks extension for long-running work, authorization hardening aligned with OAuth and OpenID Connect, and full JSON Schema 2020-12 for tool definitions. The final spec ships July 28th with breaking changes. If you're running MCP servers, start testing the RC now.

**OpenAI GPT-Live — Full-Duplex Voice** — On July 8th, OpenAI shipped GPT-Live, a fundamentally different architecture for voice interaction. Instead of waiting for you to finish talking — the old half-duplex model — GPT-Live listens and speaks simultaneously. It drops in short backchannels like 'mhmm' and 'yeah' while processing, and can delegate hard questions to GPT-5.5 in the background while the conversation keeps flowing. Two variants: GPT-Live-1 and GPT-Live-1 mini, both built on full-duplex architecture with concurrent audio input and output processing. It combines the OpenAI Realtime API with WebRTC for browser-based audio streaming. This is the most significant voice AI architecture shift since Apple Intelligence voice mode — and it's available now in ChatGPT for Plus and Pro subscribers.

**GPT-5.6 Sol Tops Coding Benchmarks** — OpenAI released the GPT-5.6 family on July 9th with three models: Luna, Terra, and the flagship Sol. GPT-5.6 Sol leads Terminal-Bench 2.1 at 91.9 percent in Ultra mode and 88.8 percent base, surpassing Claude Mythos 5 and GPT-5.5. It tops the Agents' Last Exam at 53.6 percent and the Coding Index at 80, though Claude Fable 5 still wins SWE-Bench Pro at 80 percent versus Sol's 64.6 percent. Pricing: Sol at $5 per million input and $30 output, Terra at $2.50 and $15, Luna at $1 and $6. Prompt caching got rebuilt with explicit breakpoints and a 30-minute minimum cache life. GPT-5.6 is already in OpenClaw's model catalog for 2026.7.1.

That's the briefing for Monday, July 20th. Six new stories, no repeats.