June 2, 2026 (Tue)

[01]

Top Stories

1. Microsoft Build 2026 — MAI model family unveiled + full-stack agent platform (score 2.60)

MAI-Code-1-Flash: 5B-class coding model, trained inside the Copilot harness. 85.8% on MS adversarial coding benchmark, ~51% on SWE-Bench Pro, 60% token savings on hard tasks. Deployed instantly to all Copilot tiers on Build day (Free/Pro/Pro+/Max model picker). (source)
MAI-Thinking-1: 35B active params, 128K context, reasoning-specialized. Trained from scratch without distillation from third-party frontier models — a declaration of OpenAI IP independence. (source)
Project Polaris confirmed specs: MoE architecture + per-language specialized submodules, Maia 200 in-house chip. Exceeds GPT-4 Turbo on HumanEval and MBPP (MS internal claim, awaiting independent verification). Automatic Copilot switchover in 2026-08, with a 3-month fallback option.
Windows Agent Framework 1.0 open-sourced under MIT, Azure Agent Mesh GA confirmed for Q4 2026, GitHub Copilot App + VS Code multi-agent GA, Azure AI Foundry with first-party support for Claude, Mistral, Llama, and DeepSeek.
Why it matters: Microsoft pushed MAI-Code-1-Flash to tens of millions of Copilot users instantly on Build day — an immediate adoption pipeline no competitor possesses. The front line with OpenAI is now formally "partner → competitor." Windows Agent Framework (MIT open source) emerges as a new option in the agent-framework ecosystem.
→ MAI-Code-1 / MAI-Code-1-Flash · MAI-Thinking-1 · Project Polaris · Microsoft

2. Anthropic: Agent SDK moved to a separate credit pool — effective 2026-06-15 (score 1.30)

Programmatic usage in Claude subscriptions (Agent SDK, claude -p, Claude Code GitHub Actions, third-party agents) is split out into a separate monthly credit pool. Amounts: Pro $20 / Max 5x $100 / Max 20x $200. No rollover; requests fail once the pool is exhausted.
Chat (claude.ai), terminal Claude Code, and Claude Cowork keep their existing subscription limits.
Why it matters: Direct impact on heavy Claude Code and Agent SDK users. Workflows that ran unlimited on a subscription now shift to a paid-credit basis — changing the adoption-cost math for independent developers and small teams. Coming alongside Microsoft Build, it adds another variable to developers' platform-choice calculus.
→ Anthropic (source)

[02]

No new papers today (HuggingFace Daily 403, WebSearch fallback — non-Tier-1 orgs, below threshold)

[03]

MAI-Code-1-Flash independent benchmarks: The 85.8%/51% figures are MS's own. External SWE-bench submissions will reveal true competitiveness. Worth watching the efficiency-vs-quality balance as a small model against Claude Opus 4.8 (69.2%).
Grok V9-Medium release imminent: Musk's 5/25 "~2 weeks out" remark → mid-June timeline. 1.5T parameters (3× V8's 500B), fine-tuned on Cursor data. Set to join the coding-AI race after Build 2026.
First-party Claude support in Azure AI Foundry: Microsoft directly supports Anthropic Claude on Azure. xAI Colossus lease ($1.25B/month) + Azure Claude deployment — a complex relationship mixing competition and cooperation continues.

[04]

New model pages created today — under appropriateness review

MAI-Code-1 / MAI-Code-1-Flash (NEW — Microsoft 5B-class coding model, GA on Build day. Independent benchmark verification pending)
MAI-Thinking-1 (NEW — Microsoft 35B-active reasoning model. Some specs unannounced)

[05]

Major changes to existing pages

Project Polaris: Confirmed specs updated (MoE architecture, Maia 200, HumanEval benchmark, 3-month fallback, MAI model family table added)
Microsoft: Full Build 2026 keynote reflected (7 MAI models, WAF 1.0 MIT, Azure Agent Mesh, Copilot App)
Anthropic: 6/15 billing change added
Agents (LLM Agents): WAF 1.0/Azure Agent Mesh confirmed specs, GitHub Copilot App, Azure AI Foundry multi-model support reflected