# Anthropic's Reported Surge, NVIDIA's 4-Bit Breakthrough, and ChatGPT's Finance Push *By AI High Signal Digest • May 16, 2026* Reports this week pointed to a sharp jump in Anthropic's scale, NVIDIA showed frontier-style training in 4-bit precision, and OpenAI pushed ChatGPT closer to a personal agent with connected finance data. Also in view: new reasoning research, developer tools, and major funding and infrastructure signals. ## Top Stories *Why it matters: the biggest signals today were commercial scale, cheaper frontier training, and assistants moving closer to acting on personal context.* - **Anthropic's reported economics jumped again.** A Financial Times-linked post pegged Anthropic at a **$900B valuation**, up from **$350B** in February, and said ARR rose from **$9B** at the end of 2025 to **$45B** by the end of May. Separate interview notes this week also said Anthropic has **9 of the Fortune 10** as customers and **$100B** in combined compute commitments. Together, those figures point to how quickly enterprise AI spending is concentrating around a few frontier labs. [^1][^2][^3] - **NVIDIA pushed 4-bit training from an efficiency trick toward a frontier-scale method.** NVIDIA said it trained a **12B** parameter LLM in **NVFP4** on **10T tokens** with near-zero intelligence loss, matching 8-bit baselines on MMLU, GSM8K, and coding benchmarks. The company also said NVFP4 delivers **2x-3x faster arithmetic**, **50% lower memory use**, and has already been used to pretrain **120B** and roughly **500B** Nemotron models. [^4][^5][^4] - **ChatGPT moved deeper into personal data.** OpenAI launched a personal finance experience for U.S. Pro users that lets them securely connect financial accounts, view spending, and ask GPT-5.5 questions grounded in transaction data. A follow-up post said the feature uses Plaid, cannot move money or see full account numbers, and is part of the broader push toward ChatGPT as a personal agent for home and work. [^6][^7][^8] ## Research & Innovation *Why it matters: today's most interesting research updates were about stronger reasoning, better training data, and model reliability.* - **A new reasoning model reached Olympiad-level results.** A **30B-A3B** model was released with gold-medal-level performance on **IPhO** and on **IMO/USAMO** evaluations through test-time self-verification and refinement, alongside what its authors called a simple unified scaling recipe for proof search. [^9] - **FrontierSmith targets the open-ended coding data bottleneck.** The system mutates closed-ended coding tasks into runnable optimization environments for long-horizon agents, and its authors said FrontierSmith-trained models outperformed models trained on human-curated open-ended data on **FrontierCS** and **ALE-bench**. [^10] - **A new fine-tuning result exposed a safety failure mode.** Researchers found that models fine-tuned on documents discussing implausible claims - even when those documents explicitly say the claims are false - can end up believing the claims anyway, raising doubts about how robust some current control methods are. [^11][^12] ## Products & Launches *Why it matters: new launches were less about flashy chat and more about making agents useful inside real workflows.* - **Cohere launched Compass** for search and retrieval over unstructured data, including handwritten or typed scans and other difficult documents, using a visual parsing model plus an advanced embedding stack. [^13] - **Notion expanded its developer platform for agents.** New additions include agent tools, webhook triggers, an External Agents API, and a Notion Agents SDK, with Notion saying the long-term aim is for users' agents to build workflows for them. [^14] - **VS Code added AI-generated risk badges for terminal commands.** Commands are now labeled as safe, caution, or review carefully before execution, with an experimental setting to enable the feature. [^15] ## Industry Moves *Why it matters: capital, revenue, and infrastructure scale are now moving almost as fast as the models themselves.* - **Cognition's Devin is showing unusually fast business traction.** Posts this week said Devin reached a **$445M** revenue run rate in its first 18 months, with usage doubling every eight weeks, customers including the **US Army, Goldman Sachs, and Mercedes-Benz**, and a new raise at around a **$25B** valuation. Cognition also said AngelList completed a troubled **14,000-dashboard** migration **5.2x faster** than projected using Devin. [^16][^17][^18] - **Recursive_SI launched with a $650M raise.** The company said more than a third of its team is based in the UK and described its work as contributing to UKSovereignAI goals with UK government support. [^19][^20] - **The AI buildout is becoming a capital-markets story.** One analysis this week said hyperscaler capex is set to cross **$600B** this year, while Big Tech is spending roughly **$400B/year** on AI infrastructure against about **$100B** in AI revenue, highlighting the financing strain behind the current buildout. [^21] ## Quick Takes *Why it matters: these smaller updates still help map where the ecosystem is heading next.* - xAI said its **Grok V9 1.5T** run is complete and looking strong even before supplemental training with Cursor data. [^22] - Anthropic reset users' **5-hour and weekly Claude limits**. [^23] - **DALL-E 3** will retire from **Bing Image Creator** in the coming weeks; Microsoft says it is building a dedicated replacement. [^24] - **vLLM v0.21.0** added DeepSeek V4 support, speculative decoding that respects reasoning budgets, and NVFP4/MXFP4 quantization, alongside breaking changes including a **C++20** requirement. [^25][^26][^25] --- ### Sources [^1]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2055222574561321017) [^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2055222524774846576) [^3]: [𝕏 post by @patrick_oshag](https://x.com/patrick_oshag/status/2054742528302215386) [^4]: [𝕏 post by @HowToAI_](https://x.com/HowToAI_/status/2055130148614050264) [^5]: [𝕏 post by @ctnzr](https://x.com/ctnzr/status/2055393135971492034) [^6]: [𝕏 post by @ChatGPTapp](https://x.com/ChatGPTapp/status/2055317612687675545) [^7]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2055320528198521041) [^8]: [𝕏 post by @gdb](https://x.com/gdb/status/2055335361921130861) [^9]: [𝕏 post by @stingning](https://x.com/stingning/status/2055123219506725201) [^10]: [𝕏 post by @MangQiuyang](https://x.com/MangQiuyang/status/2055340003321164157) [^11]: [𝕏 post by @OwainEvans_UK](https://x.com/OwainEvans_UK/status/2055318932857459009) [^12]: [𝕏 post by @RyanPGreenblatt](https://x.com/RyanPGreenblatt/status/2055326806589522066) [^13]: [𝕏 post by @cohere](https://x.com/cohere/status/2055343638360752351) [^14]: [𝕏 post by @NotionHQ](https://x.com/NotionHQ/status/2054625030970220834) [^15]: [𝕏 post by @code](https://x.com/code/status/2055408023506469337) [^16]: [𝕏 post by @colossusmag](https://x.com/colossusmag/status/2053801052571312414) [^17]: [𝕏 post by @cognition](https://x.com/cognition/status/2055360353089827159) [^18]: [𝕏 post by @cognition](https://x.com/cognition/status/2055360350862610690) [^19]: [𝕏 post by @KanishkaNarayan](https://x.com/KanishkaNarayan/status/2055268272623243358) [^20]: [𝕏 post by @_rockt](https://x.com/_rockt/status/2055301478168871391) [^21]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2055293526125232332) [^22]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2055296105949401342) [^23]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2055347539923308703) [^24]: [𝕏 post by @JordiRib1](https://x.com/JordiRib1/status/2055387069803815020) [^25]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2055466395937526042) [^26]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2055466391898354074)