5 fresh signals — Microsoft Build reframes Windows, Anthropic splits the frontier, OpenClaw hits its security wall, Kimi K2.6 leads open weights, and humanoids leave the demo floor
Seven in-house MAI models shipped at Build — including MAI-Thinking-1 (matches Opus 4.6 on SWE-bench Pro) and MAI-Code-1-Flash for Copilot. New Project Solara is a chip-to-cloud agent platform. Windows AI APIs moved from NPU-only to CPU + GPU. Microsoft Execution Containers (MXC) sandbox agent blast radius. Files "survived because the OS enforced the rule, not because the model behaved." Microsoft is no longer an AI customer — it's a builder.
Opus 4.8 (May 28) hits 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro. Mythos Preview (restricted to Glasswing) takes the Pro crown at 77.8% — 9 points ahead. New capability: dynamic workflows that spawn hundreds of parallel sub-agents per call. Mythos priced at 25/125 dollars per M tokens (5x Opus). Mythos GA in weeks. Every coding-agent startup priced on Opus 4.x economics has a 90-day clock.
Antiy CERT confirms 1,184 malicious skills on ClawHub (22% contamination). Trend Micro finds 492 MCP servers exposed with zero authentication. Six CVEs in 2026. 42,000 exposed servers. Pentagon designated Anthropic a "supply-chain risk" — first American company classified that way. Yet OpenClaw gained 186 stars yesterday (now 377,144). The 2015-npm moment for agent skills is here, with the agent's full permission scope as the attack surface.
Moonshot's 1T MoE / 32B active model, 256K context, 300 parallel sub-agents, modified MIT license, fits on 4x H100 INT4. Top open-weight on SWE-Bench Pro. Qwen 3.6 27B (April 16) is the consumer-hardware winner at 77.2% SWE-bench in 24GB Q4. The open-weight race is now Kimi K2.6 (frontier) vs Qwen 3.6 27B (consumer) — Mythos-vs-K2.6 at the top tier.
Tesla killed Model S and X (Jan 28) to fund Optimus. Musk admitted on Q4 2025 earnings Optimus is "not in usage in our factories in a material way." Figure 03 is the only humanoid with a paying enterprise customer — BMW Plant Leipzig, plus a 39B dollar valuation. NVIDIA's Physical AI stack (Isaac + Omniverse + Jetson Thor) is the platform layer. The race is over for demos; it's deployed units now.
Coding-agent evaluation — percentage of real-world GitHub issues solved
*Kimi K2.6 estimated from SWE-Bench Pro open-weight leaderboard. Mythos Preview restricted to Project Glasswing.
Antiy CERT audit of ClawHub registry as of June 2026
Trend Micro scan of publicly reachable MCP endpoints
Snyk for ClawHub/MCP doesn't exist yet. The trust layer is the biggest unfilled gap in agent infrastructure.
SWE-bench Pro scores across 2026 releases
Threat: Coding-agent startups priced on Opus 4.x economics have a 90-day clock. Dynamic workflows × Mythos price collapse resets the cost curve.
Microsoft is now a model builder, not a customer. Seven proprietary MAI models plus Project Solara agent platform.
Key insight: Files "survived because the OS enforced the rule, not because the model behaved." The OS is the new agent primitive. TTL should treat Microsoft as a model competitor, not a customer.
Five open-weight models, ranging from consumer laptops to multi-GPU server rigs
Key insight: Kimi K2.6 owns the frontier open-weight slot, Qwen 3.6 27B owns consumer hardware. Mythos-vs-K2.6 is the closed-vs-open frontier race for the next 90 days.
BMW Plant Leipzig is the only enterprise deployment in production. NVIDIA Isaac + Omniverse + Jetson Thor stack is the platform.
80% of Tesla strategic focus, but "not in usage in our factories in a material way." Model S and X killed Jan 28 to fund it.
First open-weights physical-AI foundation model with native action modality. Shifts robotics pretraining bottleneck from data to fine-tuning compute.