Google Launches a Decisions API That Scores 1,000 Questions in One Pass as AlphaGenome 2 Reads DNA Cell by Cell and OpenAI Warns Models Are Getting Harder to Monitor
Google launched an API that answers yes/no, multiple-choice and rubric questions without generating text, scoring up to 1,000 of them in one pass, and DeepMind released AlphaGenome 2, which predicts how DNA variants behave in individual cell types. OpenAI published research showing its newest models are getting harder to monitor through their reasoning. Plus: Runway opens the weights of its Praxis-1 robot model, ElevenLabs adds NVIDIA Nemotron to its voice agents, Stripe completes its OpenRouter acquisition, Gemini arrives on Apple CarPlay, Alibaba previews Qwen4-Max and New York opens an AI testing center.
Google Launches the Gemini Decisions API and the Gemini Decision 1 Model
Google introduced a new kind of endpoint on Wednesday: one that does not write answers at all. Google’s announcement of the Gemini Decisions API describes a service for “fast, closed-set decisions” powered by a new model, Gemini Decision 1 (gemini-decision-1), available through the Gemini API in Google AI Studio. Developers send one shared input, which can mix text, images, audio, video and PDFs up to 256K tokens, together with up to 1,000 independent questions. The model scores the candidate options in a single forward pass instead of generating text token by token.
“The Decisions API gives you purpose-built primitives for bounded choices like tool selection, policy checks, model routing, and context grading.”
Three question types are supported. A yes/no type estimates the probability that a condition holds, a choice type picks from 2 to 255 options and returns normalized probabilities for each, and a score type grades input against an ordered rubric. Every answer carries a confidence score from 0 to 1, so code can send unclear cases to a fallback or a human reviewer. Google quotes 150 to 250 milliseconds of end-to-end latency, staying nearly flat from 1 to 100 questions per request, and prices the model at $0.10 per million input tokens with zero output tokens billed. The shared input is billed once per request no matter how many questions are asked. Google says that on its shared benchmark suite the setup is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API.
The Decisions API documentation in Google AI Studio adds that questions cannot see each other, that calls require version 3.2 or later of the google-genai SDK on the v1beta endpoint, and that the recommended pattern is to ask narrow, single-purpose questions and combine the answers in ordinary code. Google says it already uses the API inside its own products, including a real-time router in Gemini Notebook that it credits with up to 40% lower latency on everyday queries, and a safety layer for MCP tools in Search. A no-cost tier is available for prototyping, integrations are rolling out across OpenRouter, LangChain, Vercel and others, and support for Gemini in Google Cloud is coming soon.
This is one of the most practical launches of the month for anyone running agents in production. Much of an agent’s cost goes into small judgment calls such as which tool to use, whether a message breaks policy, or which model should handle a request. Teams should benchmark their existing classifier and router prompts against this API, starting with routing and guardrail checks.
DeepMind’s AlphaGenome 2 Predicts How DNA Variants Act in Specific Cell Types
Google DeepMind also announced AlphaGenome 2, the successor to the DNA sequence model it released last year. DeepMind’s AlphaGenome 2 announcement says the first model predicted variant effects across broad tissue averages, while the new one is a conditional model that predicts DNA effects for one tissue, cell type or cell state at a time, taking that condition as input alongside a million-letter DNA sequence. It was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles drawn from 68 tissues, according to DeepMind, and it adds DNA methylation to the readouts it predicts.
The headline result comes from TenK10K, a single-cell map of 28 primary immune cell types. There, DeepMind says, AlphaGenome 2 identified nearly twice as many causal variants as the original model fine-tuned on single-cell data (30% versus 16%) and nearly twice as many directional gene effects (43% versus 23%). The model also reproduced, computationally, a recent finding that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. Scientists can upload single-cell RNA-seq data from a few dozen cells to condition the model on states it has not seen, without retraining.
“What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells”
That quote comes from Professor Sarah Teichmann, co-founder of the Human Cell Atlas consortium, in DeepMind’s post. Nature’s news report on AlphaGenome 2 notes the model was trained on 637 human and 311 mouse cell types. AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API and in private preview for commercial research, and DeepMind says it is intended for research use only, not clinical or diagnostic purposes.
For biotech and pharma teams, the important change is that cell-level predictions now come from an API rather than a custom research build. The commercial preview is the gate to watch. Genetics groups should test it first on loci they have already validated in the lab, so they can see where its confident calls hold up before using it to pick targets.
OpenAI Research Finds Models Are Getting Harder to Monitor Through Their Chain of Thought
OpenAI’s alignment team published a study on how training choices affect chain-of-thought monitorability, meaning how well safety teams can catch misbehavior by reading a model’s reasoning. OpenAI’s research post on chain-of-thought monitorability across the model lifecycle, dated November 10, says that with GPT-6 Astra the company observed degraded monitorability in non-adversarial settings, a greater ability to evade chain-of-thought monitors under adversarial testing, and significantly greater “no-CoT” capability, meaning reasoning done in a single forward pass without visible reasoning tokens.
The team reports four findings. No-CoT capability rises smoothly as base-model perplexity falls, so better pretraining alone makes models stronger at hidden reasoning. Some architectures do better than their perplexity predicts on multi-hop planning and serial computation tasks. CoT controllability, a model’s ability to reshape its own reasoning when told to, increased during reinforcement learning even with no process supervision beyond length limits and final-answer grading, and rose slightly more when tool supervision was added. And controllability can pass from a teacher model to a student through supervised fine-tuning on the teacher’s outputs.
“These results are sobering: as we scale toward ever-more intelligent systems, models are becoming harder to monitor through their CoT.”
OpenAI calls both measures imperfect proxies, says it found no direct reward pressure on reasoning during Astra’s training, and says it has tracked monitor-evasion capability in its safety cases since GPT-6 Astra. It expects further pretraining progress to require new approaches to monitoring.
Many enterprise oversight plans quietly assume that reading a model’s reasoning will reveal when an agent is going wrong. OpenAI is now saying publicly that this safeguard is weakening as models improve. Companies deploying agents should log and review actions, permissions and outcomes directly, and treat reasoning traces as helpful context rather than the main control.
Runway Opens the Weights of Praxis-1, Its 27B Robot Action Model
Runway made its robotics model broadly available. Runway’s post announcing that Praxis-1 is available says the world action model, introduced in September, now ships with open weights under the Apache 2.0 license, including inference and fine-tuning code and commercial use. Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions together, built on the same large-scale video pretraining as Runway’s general world models.
Runway says Praxis-1 is the top-performing world action model on NVIDIA’s RoboLab-120 leaderboard, reaching 46.2% success across 1,200 episodes, compared with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. Fine-tuned on Google’s RT-1 robot, which was not in its training data, it reached 80.0% success on SimplerEnv against 52.4% for the original RT-1 policy. The model is on GitHub and Hugging Face with LeRobot support, runs on NVIDIA hardware from H100 and B200 down to Jetson Thor, and supports LoRA fine-tuning. Launch partners Noble Machines, Standard Bots and Ultra have been running it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.
“Our research has shown that web video matches teleoperated robot video as pretraining data for policy performance, which shifts the bottleneck on a capable policy to how much general video a model can learn from.”
A permissive license on a strong robot policy lowers the barrier for hardware startups and factory automation teams that cannot afford to train their own. The benchmark figures are Runway’s own and come from simulation, so buyers should run Praxis-1 on their own cells and grippers before trusting the leaderboard. LoRA support makes that pilot inexpensive to try.
ElevenLabs Adds NVIDIA Nemotron to ElevenAgents and Expands Its Compute
On Wednesday in New York, ElevenLabs widened its partnership with NVIDIA. ElevenLabs’ announcement of its expanded NVIDIA partnership says NVIDIA Nemotron 3.5 Super VL is now available in ElevenAgents, its platform for conversational voice agents. On ElevenLabs’ own benchmark of realistic customer-support tasks, the company says, Nemotron achieved leading end-to-end task-completion accuracy at 70% lower cost than other leading open models, and it is among the three fastest models on the platform when reasoning is turned off.
ElevenLabs is also adding 15MW of compute based on NVIDIA Blackwell Ultra systems and the Vera Rubin platform, is powering a sovereign deployment of a voice agent for Poland’s national e-Health platform, and is exploring on-device voice AI for robotics with NVIDIA edge hardware. The two companies have worked together since 2022, and NVIDIA became a strategic investor in ElevenLabs in 2025.
Voice agents are moving from pilots to call-center volume, and model cost per call is starting to drive vendor choice. Open models that can be post-trained on a company’s own transcripts, and run inside government infrastructure, matter most to regulated buyers. Note that the 70% saving comes from ElevenLabs’ own benchmark, so test it on your own call flows.
Stripe Completes Its Acquisition of OpenRouter
Stripe closed the deal it announced in August. Stripe’s official post confirming OpenRouter is now part of Stripe, published November 10, says the combined company will help businesses explore new models and optimize token routing and usage. In OpenRouter’s August post on joining Stripe, the company said it processes more than 10 trillion tokens a day from more than 400 models for over 10 million developers and companies, and promised that the product, name and roadmap would stay the same. Neither company has disclosed a price, and we are not repeating unconfirmed figures.
Payments and model routing are now under one roof, which fits a market where inference is becoming a large line item for software companies. Customers should watch whether OpenRouter keeps its promise of neutral routing across labs and providers, and whether Stripe bundles token billing with its agent payment products. For now, nothing changes in existing integrations.
Quick Hits: Gemini on CarPlay, Qwen4-Max Preview, New York’s AI Testing Center and Anthropic’s IPO Calendar
The Gemini app now runs on Apple CarPlay. Google’s post on bringing Gemini to Apple CarPlay says iPhone users can hold hands-free Gemini Live conversations from the car display, ask for a daily brief from Gmail and Calendar, and pick up chats later on their phone. It is rolling out globally to all Gemini users on iOS in every supported Gemini Live language.
Alibaba opened a short preview of its next flagship. Alibaba Cloud’s Model Studio page for Qwen4-Max-Preview describes a multimodal model for coding, office work and agentic tasks with a 1M-token context window and up to 393,216 output tokens. It is available only through Token Plan until November 25 (Beijing time), after which the preview model will be retired.
New York launched a state-run AI safety research center. Governor Hochul’s announcement of the New York Center for AI Responsibility and Research says the center, based at Binghamton University and led by interim co-directors Greg Morrisett and Jeannette Wing, will bring together more than 80 researchers using Empire AI computing to independently test frontier systems. The state also opened a RAISE Act portal, where large frontier developers must register by January 1, 2027.
Finally, Anthropic’s official IPO information page now spells out the calendar after we covered the roadshow launch on Tuesday: subject to regulatory clearance, the share price is expected to be set on November 18 (EST), and shares are expected to begin trading on Nasdaq under the ticker “ANT” on November 19 (EST).
New York’s registration deadline is the item with real compliance weight. Any company that develops frontier models and operates in the state should confirm whether it counts as a large frontier developer and prepare its safety framework now. The 72-hour incident reporting rule will need an internal escalation path that most teams do not yet have.
What We Left Out
Social posts and secondary reports attributing a revenue run rate, a long-term revenue target and a three-day AMD chip port to Anthropic’s roadshow could not be matched to the prospectus or Anthropic’s IPO page, so we did not use them. A reported price for the Stripe and OpenRouter deal is unconfirmed. Reports of an OpenAI $10 billion revenue quarter, DeepSeek pretraining on Huawei chips, remarks attributed to Sam Altman in a private meeting, and a teased Codex usage reset were not confirmed against primary sources. ChatGPT’s voice mode retirement was covered in yesterday’s edition.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →