# OpenAI’s Math Milestone and Anthropic’s Colossus-Scale Bet *By AI High Signal Digest • May 21, 2026* OpenAI claimed the first autonomous AI solution to a major open math problem, while Anthropic paired a multiyear Colossus compute deal with projected profitability. Also inside: agent research advances, new product launches from Cohere, OpenAI, and Google, plus notable funding and policy moves. ## Top Stories *Why it matters: today’s clearest signals were AI reaching a new research threshold, and frontier-model economics becoming even more tied to massive infrastructure commitments.* - **OpenAI said a general-purpose reasoning model solved the planar unit distance problem**, a famous Erdős question posed in 1946. The model found a new family of constructions that beats the square-grid approach mathematicians had treated as best for nearly 80 years, which OpenAI called the first time AI has autonomously solved a prominent open problem central to a field of mathematics [^1][^2]. > “What’s significant about this moment is that it’s the first really clear example of AI solving — not just an unsolved math problem — but a really well-known math problem.” [^3] The result also comes less than a year after frontier models reached IMO gold-level performance, marking a fast jump from competition math to original research [^4][^5]. - **Anthropic’s scale story sharpened on both compute and finances.** A SpaceX filing says Anthropic is paying **$1.25B per month through May 2029** for capacity across Colossus and Colossus II [^6], while WSJ-reported projections put Q2 revenue at **$10.9B** and first operating profit at about **$559M** [^7]. Together, those numbers suggest frontier AI demand is now supporting both multiyear infrastructure commitments and near-term profitability. ## Research & Innovation *Why it matters: beyond headline model releases, the most useful research updates focused on making agents cheaper, more scientifically useful, and easier to evaluate honestly.* - **Databricks introduced MemEx**, a programmable Python scratchpad that lets agents transform, slice, and persist tool outputs as typed objects instead of flooding the context window [^8]. On enterprise agentic tasks, Databricks says frontier models gained **2–5 accuracy points** at **25–30% lower cost**, while Qwen 122B and 397B nearly doubled accuracy at **40–50% lower cost** [^8]. - **Hugging Face released Carbon**, an open DNA base model with open weights, training code, and data pipeline for downstream biology tasks [^9]. The team says Carbon is **275x faster** than the next best model at its size, can run locally, and can process a whole human genome on a single GPU in under two days [^9]. - **InferenceBench offered a reality check on AI R&D automation.** Its creators say current frontier agents still struggle with system-level engineering and complex dependencies, and underperform simple hyperparameter-tuning baselines for vLLM and SGLang [^10]. They also found weak strategy diversity, with most agents defaulting to vLLM rather than exploring alternatives [^10]. ## Products & Launches *Why it matters: new launches are pushing AI from standalone chat into reusable workflows for coding, science, and open deployment.* - **Cohere open-sourced Command A+**, its fastest and most powerful model yet, under Apache 2.0 [^11][^12]. Cohere says it supports **48 languages**, multimodal input, and can run on as few as **two H100s** [^11][^13]. - **OpenAI brought Codex to mobile.** Users can work with Codex from the ChatGPT mobile app, answer questions on the go, and continue the same thread later from a computer [^14][^15]. - **Google DeepMind launched Science Skills for Antigravity**, integrating insights from more than 30 life-science sources including UniProt and the AlphaFold Database [^16]. In a test on a rare disease caused by AK2 mutations, the team said the tool sped up structural analysis and surfaced novel insights into the condition’s mechanisms [^17]. ## Industry Moves *Why it matters: capital and partnerships are concentrating around the data, search, and vertical workflows that agents need to be useful in production.* - **Exa raised $250M at a $2.2B valuation** to keep building search for agents [^18]. The company says it now serves **5,000+ companies** and **500,000+ developers**, and makes agents smarter and cheaper by returning **90% less text** with little to no tradeoff in RAG quality [^18]. - **Genesis and Incyte expanded their AI drug-discovery partnership** with **$120M upfront**, recurring research funding, and potential milestone and royalty payments [^19]. Incyte’s proprietary experimental data will help train the next generation of Genesis foundation models, starting with at least five new collaboration targets [^19]. - **RADAR became a new physical-AI unicorn** with a **$170M Series B** [^20]. The company says its retail system delivers **99% item-level inventory accuracy** in real time and is already deployed in more than **1,400 stores** [^20]. ## Policy & Regulation *Why it matters: governments are moving closer to pre-release oversight of frontier systems.* - The White House briefed OpenAI, Anthropic, and Reflection AI on a planned executive order that would create a **voluntary** framework asking labs to share frontier models with government agencies up to **90 days before public release** [^21]. ## Quick Takes *Why it matters: a few smaller updates added signal on agents, search efficiency, and generative media.* - More than **50% of Devin sessions** are now triggered by agents rather than humans [^22][^23]. - **Perplexity** said its query-aware context compression cuts context tokens by up to **70%** while improving answer quality [^24][^25]. - **Stable Audio 3** launched with three open-source variants plus a closed “large” model [^26]. - **MiniMax Speech 2.8 Turbo** arrived on Together AI with **600+ voices** and sub-250ms latency across **40+ languages** [^27][^28][^27]. --- ### Sources [^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2057176201782075690) [^2]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2057176203166171317) [^3]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2057183672348921947) [^4]: [𝕏 post by @polynoamial](https://x.com/polynoamial/status/2057178198228586824) [^5]: [𝕏 post by @alexwei_](https://x.com/alexwei_/status/2057182874995155041) [^6]: [𝕏 post by @eliebakouch](https://x.com/eliebakouch/status/2057222864332320999) [^7]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2057205565776908297) [^8]: [𝕏 post by @DbrxMosaicAI](https://x.com/DbrxMosaicAI/status/2056818063215878618) [^9]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2057071550352781771) [^10]: [𝕏 post by @maksym_andr](https://x.com/maksym_andr/status/2057106398228439148) [^11]: [𝕏 post by @cohere](https://x.com/cohere/status/2057121297491009564) [^12]: [𝕏 post by @cohere](https://x.com/cohere/status/2057122131410813016) [^13]: [𝕏 post by @cohere](https://x.com/cohere/status/2057121964406284734) [^14]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2057142816497906045) [^15]: [𝕏 post by @OpenAIDevs](https://x.com/OpenAIDevs/status/2057142817944969367) [^16]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057256257153884161) [^17]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057256259037131226) [^18]: [𝕏 post by @jeffzwang](https://x.com/jeffzwang/status/2057176652862644239) [^19]: [𝕏 post by @genesismolai](https://x.com/genesismolai/status/2057134349611672047) [^20]: [𝕏 post by @SpencerHewett](https://x.com/SpencerHewett/status/2056731580157047078) [^21]: [𝕏 post by @steph_palazzolo](https://x.com/steph_palazzolo/status/2057179490900377620) [^22]: [𝕏 post by @imjaredz](https://x.com/imjaredz/status/2057226010467463531) [^23]: [𝕏 post by @shayanshafii](https://x.com/shayanshafii/status/2057214878960828769) [^24]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2057151032569045325) [^25]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2057151002105753950) [^26]: [𝕏 post by @multimodalart](https://x.com/multimodalart/status/2057154251558682798) [^27]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2057149932172112260) [^28]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2057149934508351651)