MAI-Thinking-1
35B MoE, SWE-Pro = Opus 4.6
5 fresh signals compiled from Scout + Quant + Daily Research streams · 2026-06-10
Anthropic shipped Mythos-class capability to the public on June 9. Same underlying model, two access tiers — Fable 5 (public, sensitive cyber/biology/chemistry routed to Opus 4.8) and Mythos 5 (Glasswing + 150 new vetted orgs, cyber safeguards lifted). 60% price cut from Mythos Preview. 1M context, 128k output. This is a pre-packaged EU AI Act compliance tier — and the template has already spread to OpenAI's GPT-Rosalind for life sciences.
Microsoft BUILD 2026 (June 2-9) closed the loop: MAF 1.0 GA on April 2, MAI-Thinking-1 35B-active MoE (~1T total) matches Claude Opus 4.6 on SWE-Bench Pro, CodeAct + Hyperlight micro-VMs cut agent latency ~50% and tokens ~60%, Foundry Hosted Agents GA target July. "Build in GitHub, Run in Foundry, Ship in Teams" is now a complete enterprise agent platform. Microsoft can credibly serve enterprise without routing through OpenAI.
Recursive Superintelligence emerged from stealth May 13 with 8 co-founders (Yuandong Tian ex-Meta FAIR, Jeff Clune Darwin Gödel Machine, Tim Rocktäschel ex-DeepMind, Alexey Dosevitskiy ViT author). $650M sole round, $4.65B post, GV + Greycroft co-led. The signal: Nvidia AND AMD on the same cap table is a category bet on recursive self-improvement as a near-term compute customer. Jack Clark pegs 60% probability of self-improving AI by end-2028.
Broadcom + Apollo + Blackstone launched a $35B initial capital solution to deploy 20+ GW of compute for Anthropic + OpenAI by 2028. First tranche funds Anthropic's 1+ GW Fluidstack buildout (mid-2026). Funded as equity + credit + project finance hybrid — NOT typical VC. Pure equity can't fund the projected $1T+ compute need. This is the financial template for H2 2026+ AI capex. New systemic risk vector for credit markets if 2027-28 AI revenue underwhelms.
NOTUS: top AI companies learned of Trump's planned White House meeting from Air Force One press comments — not the White House. Sanders' American AI Sovereign Wealth Fund Act (50% equity tax) is the policy backbone. Separately: SpaceX IPO T-2 days, $135 fixed × 555.6M shares = $75B raise, $1.77T valuation, Nasdaq debut June 12. Engineered price floor (15% greenshoe, 30% retail, Nasdaq fast-entry). If SPCX breaks $135 in first 30 days, it re-prices the entire H2 2026 IPO pipeline.
Anthropic alone: $70B+ secured compute via 4 tranches. Mix is shifting from equity to credit.
Mix shift: $35B XPV Platform is credit + project finance, not equity. Bifurcated AI capex is the new systemic risk vector.
Each stack has a different lock-in vector. MCP is the portability layer.
Wildcard: SpaceX $60B option on Cursor. If exercised: 4-way war, SpaceX has most concentrated vertical integration (IDE + Colossus + xAI Grok). Decision expected later 2026.
Prosaic-RSI signals: SWE-Bench saturation + time-horizon jump + 8x code-merge at Anthropic
Clark's alignment warning: 99.9% accuracy → ~60% after 500 generations. Compounding-error risk in recursive loops is the under-discussed threat.
The contract is a sentiment indicator, not a forecast. YES > 30% on volume = real institutional re-pricing. Currently still "probably not" — but the drift is the only consistent negative signal in markets.
Market-moving events in the next 8 weeks
TTL signal: Three market-shaping events stack this week: (1) Trust-tier era is officially here — Mythos 5 + GPT-Rosalind template confirmed; (2) AI capex shifted to credit/project finance — systemic risk vector opens; (3) RSI is the first cap-table bet on the self-improvement thesis. The harness layer (eval, observability, governance) is the only durable moat if the model layer becomes recursive-self-improving. Multi-stack primary is still the only safe play. MCP is the portability bet.