# Microsoft’s MAI Launch, Codex’s Expansion, and the New Compute Crunch *By AI High Signal Digest • June 3, 2026* Microsoft’s seven-model MAI release led the day, while OpenAI broadened Codex into a larger work platform and warned of deep compute constraints ahead. This brief also covers new research on scientific discovery and agent reliability, plus notable desktop-agent launches and funding moves. ## Top Stories *Why it matters: model competition is shifting from raw capability to deployability, cost, and broader adoption.* - **Microsoft made its biggest MAI push yet.** At Build, Microsoft launched seven new MAI models across reasoning, code, image, transcribe, and voice, led by **MAI-Thinking-1**, a 35B active-parameter MoE with 256K context that scored **97% on AIME 2025** and **53% on SWE-Bench Pro**. Microsoft also said the model delivers **30% better performance per dollar** and **1.4x performance-per-watt** on its **MAIA 200** chip versus GB200. A separate post linked a **109-page** technical report. [^1][^2][^3] - **OpenAI pushed Codex beyond coding.** New **Sites** let teams turn plans and work into interactive websites or apps with deployed URLs, authentication, static files, and database-backed dynamic data, with rollout starting on Business and Enterprise plans. OpenAI also added role-specific plugins for sales, product design, creative production, data analytics, and public equity investing. A report first shared with Axios said Codex passed **4M weekly users**, up **5x since February**, with knowledge workers now one-fifth of users and growing faster than developers. [^4][^5][^4][^6][^7] - **Compute scarcity is becoming a strategic constraint.** OpenAI CFO Sarah Friar said demand is rising “almost like a vertical wall,” that materially more compute in **2026** will be hard to source, and that supply constraints are likely to remain severe in **2027**. She also said Nvidia remains OpenAI’s top partner while the company is also working with AMD, Cerebras, and Broadcom on diversified chip supply. [^8] ## Research & Innovation *Why it matters: the most useful research today focused on scientific discovery, agent reliability, and where current frontier systems still fail.* - **Google DeepMind launched Co-Scientist**, a Gemini-based multi-agent system that generates, debates, verifies, and ranks thousands of hypotheses for complex scientific problems. DeepMind said evaluations surfaced new targets for liver fibrosis, fresh ALS approaches, and genetic leads for reversing aging, and the system is now available to individual researchers through **Gemini for Science**. [^9][^10][^11][^12] - **Harness-1 proposed a different way to build agents.** The core idea is to keep reliable working memory—candidate pools, evidence links, verification records, and budget-aware context—outside the policy, leaving the 20B model to decide what to search, keep, verify, and when to stop. Across eight retrieval benchmarks, it reached **0.730 average curated recall**, beating the next-best open search agent by **11.4 points**. [^13] - **AutoMedBench showed how far medical research agents still have to go.** The benchmark covers **24 tasks** across CT, X-ray, pathology, question answering, report generation, and segmentation; the authors said six tested frontier models remained far from reliable medical AI researchers, with validation the weakest stage and engineering failures dominating. [^14] ## Products & Launches *Why it matters: the product layer kept moving toward desktop-native agents and hybrid local/cloud workflows.* - **Devin Desktop** launched as a workspace for managing fleets of local and cloud agents from one surface, with a full IDE, support for any ACP-compatible agent, and cloud handoff so work can continue after a laptop is closed. [^15][^16][^17][^18] - **GitHub Copilot app** debuted as a desktop home for agent-native software development on GitHub, with continuity across desktop, CLI, mobile, and web, plus agent-native issue and PR workflows. [^19][^20][^21] - **Perplexity announced hybrid agentic inference** for Perplexity Computer, splitting tasks between on-device local models and frontier cloud models to keep private data local and improve token efficiency. The company said it is coming to Windows laptops, Macs, and Linux machines. [^22][^23][^24] ## Industry Moves *Why it matters: labs are extending from model releases into enterprise control, vertical partnerships, and infrastructure bets.* - **Microsoft’s enterprise strategy is moving from models to customization.** Its new **Frontier Tuning** lets companies build company-specific agents they control themselves; Microsoft said an early McKinsey tuning outperformed GPT-5.5 on quality at **10x lower cost**. [^1] - **Microsoft and Mayo Clinic are jointly training a frontier healthcare model**, extending Microsoft’s MAI effort into a high-value vertical. [^1][^25] - **OpenRouter raised $113M in Series B funding** led by CapitalG to scale its multi-model inference routing platform. [^26] ## Policy & Regulation *Why it matters: Washington’s AI posture is still being shaped in real time, and lab reactions matter.* - **A new White House executive order on AI drew immediate support from major labs.** Anthropic called it “an important step” for U.S. AI leadership and said it was ready to support implementation, while Sam Altman said the order gets the balance right between leading on model development, safety, and cyber defense. [^27][^28] ## Quick Takes *Why it matters: a few smaller updates still sharpen the competitive picture.* - **Alibaba’s Fun-Realtime-TTS** took the top spot on Artificial Analysis’ Speech Arena with a **1,219 Elo**, ahead of Gemini 3.1 Flash TTS and Inworld Realtime TTS-2. [^29] - **MiniMax-M3** became the new open-weight SOTA on the **Vals Index** and **Vals Multimodal Index**, ranking **#6 overall**. [^30] - **Unsloth**, with NVIDIA and Microsoft, said users can now train **120B+ parameter models locally** on the 128GB unified-memory **RTX Spark** laptop. [^31][^32] - **Krea 2 Medium** debuted at **#6** on Artificial Analysis’ text-to-image leaderboard, ahead of its larger Krea 2 Large variant. [^33] --- ### Sources [^1]: [𝕏 post by @mustafasuleyman](https://x.com/mustafasuleyman/status/2061880164498428188) [^2]: [𝕏 post by @MicrosoftAI](https://x.com/MicrosoftAI/status/2061887500541366489) [^3]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2061889624847343825) [^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2061845949170045346) [^5]: [𝕏 post by @TheRohanVarma](https://x.com/TheRohanVarma/status/2061872164442403139) [^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2061887650391625870) [^7]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2061820756250722474) [^8]: [𝕏 post by @Hangsiin](https://x.com/Hangsiin/status/2062026881315258848) [^9]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2061857539977842793) [^10]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2061857547095597517) [^11]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2061857550438392094) [^12]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2061857553076920643) [^13]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2061825437693841651) [^14]: [𝕏 post by @iScienceLuvr](https://x.com/iScienceLuvr/status/2061772890316698048) [^15]: [𝕏 post by @cognition](https://x.com/cognition/status/2061889596703551926) [^16]: [𝕏 post by @windsurf](https://x.com/windsurf/status/2061889084541509922) [^17]: [𝕏 post by @cognition](https://x.com/cognition/status/2061889649858040131) [^18]: [𝕏 post by @cognition](https://x.com/cognition/status/2061889654878609576) [^19]: [𝕏 post by @pierceboggan](https://x.com/pierceboggan/status/2061868635241828688) [^20]: [𝕏 post by @lukehoban](https://x.com/lukehoban/status/2061905448287322243) [^21]: [𝕏 post by @lukehoban](https://x.com/lukehoban/status/2061905445623833023) [^22]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2061861293569765847) [^23]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2061875858542096520) [^24]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2061877326494027778) [^25]: [𝕏 post by @mustafasuleyman](https://x.com/mustafasuleyman/status/2061903347129418227) [^26]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2061885795779199052) [^27]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2061924580222968183) [^28]: [𝕏 post by @sama](https://x.com/sama/status/2061973280655904815) [^29]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2062016529848222073) [^30]: [𝕏 post by @ValsAI](https://x.com/ValsAI/status/2061926305722200226) [^31]: [𝕏 post by @UnslothAI](https://x.com/UnslothAI/status/2061671421643039138) [^32]: [𝕏 post by @danielhanchen](https://x.com/danielhanchen/status/2061682922357506550) [^33]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2061859279905370468)