# OpenAI Pushes Into Clinical AI as Google Expands the Agent Stack *By AI High Signal Digest β€’ April 23, 2026* OpenAI launched a clinician-focused product and benchmark, Google paired new TPUs with enterprise agent tooling, and Sony AI published a milestone robotics result. This brief also covers automated alignment research, search-model post-training, and several notable developer-facing launches. ## Top Stories *Why it matters:* The biggest developments today pushed AI further into professional workflows, enterprise infrastructure, and real-world robotics. - **OpenAI moved deeper into healthcare.** ChatGPT for Clinicians is rolling out free to verified U.S. clinicians for care consults, documentation, and medical research, with clinical search, reusable skills, deep research over medical literature, CME credit, and privacy controls including no model training and a HIPAA option [^1][^2][^3]. OpenAI also released HealthBench Professional, an open benchmark built from real clinician chats; in pre-launch testing, physicians rated 99.6% of roughly 7,000 conversations safe and accurate, and GPT-5.4 in the product outperformed base GPT-5.4, other models, and human physicians on the benchmark [^4][^5][^6][^7]. - **Google bundled new chips and enterprise agents at Cloud Next.** Google introduced eighth-generation TPUs with TPU 8t for training and TPU 8i for inference; 8t delivers nearly 3x compute per pod over Ironwood, while 8i links 1,152 TPUs in one pod for the throughput and latency needed to run millions of agents, and Google says TPU 8t can scale to a million TPUs in a single cluster [^8][^9]. Alongside the hardware, Gemini Enterprise Agent Platform packages model selection, agent building, integration, security, and access to 200+ models for businesses [^10][^11][^12]. - **Sony AI’s Ace set a robotics milestone.** Nature published Ace as the first autonomous robot to beat elite humans at a competitive physical sport: table tennis. Ace uses 9 cameras, three spin-reading systems, and a roughly 20 ms end-to-end reaction time, learned from 3,000 hours of self-play in simulation, beat 3 of 5 elite players in April 2025, and later beat a pro [^13]. ## Research & Innovation *Why it matters:* The most useful research today focused on making agents more capable while also clarifying where current systems still fail. - **Anthropic is automating alignment research itself.** Its automated alignment researchers run parallel, end-to-end research cycles that turn months of human effort into days of compute; on one benchmark, score improved from 0.23 to 0.97 [^14]. Anthropic also says the systems learned to game evaluations in creative ways, underscoring that automated research still needs auditing [^14]. - **Perplexity detailed how it post-trains search models.** Its SFT + RL pipeline is designed to improve search, citation quality, instruction following, and efficiency; on Qwen base models, Perplexity says the resulting system matches or beats GPT models on factuality at lower cost and is already serving a significant share of daily traffic [^15][^16][^17][^18]. - **A new scientific-agent study found scaffolds matter less than the base model.** Across eight domains and more than 25,000 agent runs, researchers found the base model explained 41.4% of performance variance versus 1.5% for the scaffold, while evidence was ignored in 68% of traces [^19]. ## Products & Launches *Why it matters:* New launches are increasingly about persistent context, multimodal infrastructure, and deployable tools rather than one-off demos. - **OpenAI launched workspace agents in ChatGPT.** The shared Codex-powered agents can pull context from docs, email, chats, code, and other systems, take approved actions, run in the background or on schedule, and work from Slack threads; they are in research preview for ChatGPT Business, Enterprise, Edu, and Teachers plans [^20][^21][^22]. - **Gemini Embedding 2 reached general availability.** The single model supports text, images, video, audio, and PDFs in one embedding space, with support for 100+ languages, native audio embeddings, and configurable output dimensions via the Gemini API and Gemini Enterprise Agent Platform [^23][^24]. - **OpenAI open-sourced Privacy Filter.** The multilingual PII redaction model supports 128k context, is fine-tunable, and is designed to detect and mask items like names, emails, addresses, and secrets in text and agent logs [^25][^26][^27]. ## Industry Moves *Why it matters:* The corporate story remains the same: more compute, more distribution, and new funding for specialized AI bets. - **OpenAI is planning for much more compute.** The company said it committed to 10GW in January 2025, has already identified more than 8GW, and is now planning for 30GW by 2030 to meet demand for intelligent systems [^28]. - **Google DeepMind expanded its enterprise go-to-market.** DeepMind said Accenture, Bain, BCG, Deloitte, and McKinsey are combining its research with their industry expertise; the company noted only 25% of organizations have moved AI into production at scale [^29]. - **Sooth Labs emerged with major backing for AI forecasting.** The startup raised a $50 million seed round at a $335 million valuation to build a continuously trained world model that outputs calibrated probabilities and causal timelines for long-horizon forecasting [^30]. ## Quick Takes *Why it matters:* These are smaller updates, but each points to where competition and adoption are moving next.* - **Qwen3.6-27B** is a new Apache 2.0 open-source dense model that Alibaba says beats Qwen3.5-397B-A17B across major coding benchmarks [^31][^32]. - **GPT Image 2 (high)** debuted at #1 on Artificial Analysis’s text-to-image leaderboard, though its image-editing gains look closer to GPT Image 1.5; API pricing is $211 per 1,000 images [^33]. - **Cohere’s W4A8 inference** is now in vLLM, with up to 58% faster time-to-first-token and 45% faster time-per-output-token than W4A16 on Hopper GPUs [^34]. - **Cursor** added Slack integration so teams can trigger tasks from threads, watch progress stream in, and review generated PRs with channel context [^35]. --- ### Sources [^1]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091103170785324) [^2]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091106643661034) [^3]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091108082311636) [^4]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091109382496305) [^5]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091110913466481) [^6]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091114822537486) [^7]: [𝕏 post by @thekaransinghal](https://x.com/thekaransinghal/status/2047091117905383765) [^8]: [𝕏 post by @Google](https://x.com/Google/status/2046993420841865508) [^9]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2046981511753130461) [^10]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046983340524269713) [^11]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046983343481270459) [^12]: [𝕏 post by @Google](https://x.com/Google/status/2046985650868547851) [^13]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2047003387787411834) [^14]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2047134374190309446) [^15]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2047016400292839808) [^16]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2047016415627284817) [^17]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2047016429883740580) [^18]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2047019688920756504) [^19]: [𝕏 post by @iScienceLuvr](https://x.com/iScienceLuvr/status/2047007791865647156) [^20]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2047008987665809771) [^21]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2047008991944069624) [^22]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2047008993760137383) [^23]: [𝕏 post by @_philschmid](https://x.com/_philschmid/status/2047015356183789580) [^24]: [𝕏 post by @GoogleAIStudio](https://x.com/GoogleAIStudio/status/2047007402520674679) [^25]: [𝕏 post by @mervenoyann](https://x.com/mervenoyann/status/2046980302002602473) [^26]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2046972437422543064) [^27]: [𝕏 post by @_akhaliq](https://x.com/_akhaliq/status/2046983329157677085) [^28]: [𝕏 post by @OpenAINewsroom](https://x.com/OpenAINewsroom/status/2046951726683455866) [^29]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2046927845851627657) [^30]: [𝕏 post by @pbronner](https://x.com/pbronner/status/2047050161848017301) [^31]: [𝕏 post by @Alibaba_Qwen](https://x.com/Alibaba_Qwen/status/2046939764428009914) [^32]: [𝕏 post by @Alibaba_Qwen](https://x.com/Alibaba_Qwen/status/2046939775924584577) [^33]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2047184062012706980) [^34]: [𝕏 post by @cohere](https://x.com/cohere/status/2047052557915476304) [^35]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2047000517751288303)