HAI Daily

GPT-6.2 Astra and Sol, Nvidia's Nemotron for Science and a $5 Billion Bet on AI Drug Design

OpenAI releases GPT-6.2 Astra and Sol with a 20% price cut on Astra, Nvidia opens a chemistry dataset and post-training recipes for science agents, Isomorphic Labs raises $5 billion, and Nvidia reports $115.8 billion in quarterly revenue under its interim CEO.

By H.AI Tools Team4 stories
Products in this issue: ChatGPT, Claude, GitHub Copilot, Elicit, Consensus, Google's AI Co-Scientist, Excelly-AI, Perplexity
In this issue

The short version

  • GPT-6.2 Astra costs 20% less than GPT-6 Astra, at $8 per million input tokens and $40 per million output tokens. Sol keeps its $2 and $10 pricing and comes close to Astra on coding at a fraction of the cost.
  • Nvidia's Nemotron for Science combines an open vision-language model, about 50,000 chemistry questions in LitmusBench v0.2 and post-training recipes for scientific agents.
  • Isomorphic Labs raised $5 billion in a Series C co-led by Greenoaks and Sequoia, to move its first AI-designed drugs toward clinical trials.
  • Nvidia's revenue doubled to $115.8 billion, and its board said it will not bring back Jensen Huang as chief executive.

November 18 brings a new pair of OpenAI models built around cost per task, an open toolkit from Nvidia for training science agents, the largest funding round yet for Isomorphic Labs and Nvidia's first quarterly report under an interim chief executive. Each section explains what happened and what you can do with it.

01

GPT-6.2 Astra and Sol: judged by cost per task

What OpenAI released

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on November 17. Astra is the higher-end model for complex work, and Sol is the cheaper model for everyday tasks. Both are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol.

Astra's API price fell 20% to $8 per million input tokens and $40 per million output tokens. Sol keeps its predecessor's price of $2 and $10. On Agents' Last Exam, a test of long professional workflows, Astra reportedly scored 78.3% at about $2.89 per task, against about 73% and $60 per task for Claude Fable 5.2, roughly 5% of the cost. OpenAI says Sol came within half a percentage point of Astra on a coding benchmark at about a fifth of the cost per task. These are launch-day figures, mostly from OpenAI, and independent scores published the same day were mixed. OpenAI also says Astra is the first model it has released that it rates High for AI self-improvement capability, and both models ship with tighter safeguards on higher-risk cybersecurity help.

Choosing between Astra, Sol and what you use now

The useful number is no longer the price per token but the cost of finishing a task, because a stronger model that uses fewer steps can be cheaper overall. Test on your own work rather than the benchmark. Take ten real tasks, such as a report from three documents or a bug fix with tests, run each on Sol, Astra and your current model, and record the result, the time taken and the cost shown in your usage dashboard.

Use Sol by default and send tasks to Astra only where Sol fails. For coding, compare the same tasks in GitHub Copilot or Codex, and keep reviewing generated code before merging. If you also use Claude, rerun the comparison rather than trusting a headline that compares settings you will not use. Our ChatGPT vs Claude comparison covers the plans, and ChatGPT has the new models now.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Nemotron for Science: open tools for chemistry agents

What Nvidia announced

Nvidia introduced Nemotron for Science, a set of open tools for building AI agents that work on scientific problems. It combines the Nemotron 3.5 Super VL vision-language model, post-training recipes and LitmusBench v0.2, an open dataset of about 50,000 chemistry questions for training and testing language models. The questions draw on molecular property data and experimental drug discovery records. The model and dataset are on Hugging Face.

Early users report results on specialist tasks. Terray Therapeutics says a post-trained Nemotron recovered about 76% of annotated reactions on a standard test, and Therna reports nearly doubling accuracy in predicting protein expression. Those figures come from the companies, and no outside group had published LitmusBench scores at launch.

Making AI useful in a research team

General assistants are good at reading and summarising papers but weak at specialist judgement, which is the gap post-training tries to close. Most teams do not need to train a model. Start by using Elicit or Consensus to find and summarise the literature, and ask for the source of every claim. Check a sample of those sources yourself each week.

If your group has its own data and computing time, an open benchmark such as LitmusBench is a cheap way to check whether a model understands your field before you rely on it. Run a candidate model on the questions closest to your work and compare it with your current tool. Keep experimental plans under human review, and record which model proposed what. Our AI academic research workflow shows a full example, and tools such as Google's AI co-scientist are listed in the Healthcare AI category.

03

Isomorphic Labs raises $5 billion for AI drug design

What was announced

Isomorphic Labs, the Alphabet company founded by DeepMind chief Demis Hassabis, announced a $5 billion Series C on November 17. Greenoaks and Sequoia Capital co-led the round, with Andreessen Horowitz among the new investors and Alphabet continuing to invest. Thrive Capital, GV, CapitalG, Temasek and the UK Sovereign AI Fund also took part. It comes about six months after a $2.1 billion Series B. The company did not disclose a valuation.

Isomorphic says the money will expand its drug design platform and pipeline and prepare its first candidates for clinical trials. It has not yet disclosed any programme in clinical testing.

Reading AI health claims carefully

Large funding rounds show investor belief, not results. An AI-designed drug still has to pass the same clinical trials as any other, and most candidates fail. When you read about AI in medicine, check what stage a claim refers to: a computer prediction, a laboratory result, or evidence from patients.

For personal health questions, use an assistant to prepare, not to decide. Ask it to explain a diagnosis or list questions for your doctor, and check any treatment information against official health services. Never upload medical records to a tool without reading its data policy first. The Healthcare AI category lists tools with a medical focus.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

04

Nvidia: revenue doubles under an interim CEO

What Nvidia reported

Nvidia reported third-quarter fiscal 2027 revenue of $115.8 billion on November 17, up 103% from a year earlier, with $107.6 billion from data centres. Net income was $65.3 billion. On the call, Nvidia said its new Vera Rubin platform brought in nearly $20 billion in its first quarter of production shipments. It guided fourth-quarter revenue to about $130 billion and assumes no data-centre computing revenue from China.

This was the first report since Jensen Huang resigned, covered in our November 13 issue. Interim CEO Colette Kress led the call, and board chair Harvey Jones said the board does not intend to reinstate Huang and will consider internal and external candidates for a permanent chief executive.

What it means for AI tool users

Nvidia's numbers show demand for AI computing still rising faster than supply. For users, that points to more capacity across cloud services next year, while the newest models may stay rate-limited at first. Keep a fallback model tested for each important workflow so a limit on one service does not stop your work.

If you want to read the release yourself, ask an assistant for a table of revenue by segment for this quarter and the same quarter last year, with the source sentence for each figure, and then check each number. Excelly-AI can turn the table into a chart, and Perplexity can find coverage to cross-check. This is a reading exercise, not investment advice; the Finance AI category lists more research tools.

What to try next

Put one change into practice

Run ten of your real tasks on Sol and Astra and compare the cost of finishing each. If you do research, test an assistant on questions from your own field before trusting it. Read AI health news by asking what stage the evidence is at.

Explore Chatbots, Code Generation and Healthcare AI to build your shortlist.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper