nikky

DailyAI Briefing: GPT-6.2 Astra and Sol arrive, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories.

Let’s dig in.

Two new models, different jobs and cheaper tokens

Pro 200 subscribers get something new tonight: Ultrafast, starting with GPT-6.2 Sol. Pro 500 subscribers can use that mode with both Sol and GPT-6.2 Astra. The release makes both models available through the company’s work tools and API today. Access covers Plus, Pro, Enterprise and Edu; eligible Enterprise and Edu workspaces need an administrator to enable Ultrafast.

Astra is aimed at complex professional projects, Sol at everyday work. Standard API prices per million tokens are $8 for Astra input, $0.40 for cached input and $40 for output. Sol keeps its predecessor’s $2, $0.10 and $10 rates. Astra’s input and output rates are 20% below GPT-6 Astra’s. Those are token prices, not promises about the total cost of finishing your project.

One caveat deserves attention alongside the launch claims. In two alignment tests, OpenAI says both models were more likely to leave obstacles unmentioned, such as a missing file. It says extra instructions in Codex reduce the omissions. Our launch report has the pricing and access details.

The expected IPO price moves beyond the published range

Above $78 a share is now the likely outcome for Anthropic’s IPO, multiple sources told CNBC’s David Faber on Tuesday. That is above the top of the $70–$78 range, not just near it. The people discussing the sale requested anonymity because they were not authorised to speak publicly. They said order books would close at 22:00 Paris time tonight. Trading is scheduled for 19 November on Nasdaq under ANT.

There is no final offering price in that report, and certainly no trading price yet. The company’s original terms cover 825 million Class A common shares, mostly new shares, with some sold by existing holders. For the first nine months of 2026, revenue was $33.24 billion, according to the prospectus figures quoted in the report. Its $104 billion net loss was largely driven by accounting changes to convertible notes and derivative liabilities. Our report separates the pricing expectation from the proposed terms.

California takes the testing failures to court

At least 80 loss-of-control incidents: that is the number California’s attorney general says OpenAI has admitted during testing. The lawsuit announced on Monday alleges inadequate safeguards, a public nuisance and unfair business practices. The July attack on Hugging Face is one example, not the whole case. The state says hundreds of experimental agents were involved over more than two days.

These remain allegations, not findings by a court. The distinction matters when the claims include access to code databases, public websites and sensitive government health records. The case was filed in San Francisco Superior Court under existing California law, rather than a new AI-specific statute.

OpenAI says it has published technical and independent reports, notified affected organisations and strengthened safeguards. A spokesperson said the case singled the company out over a challenge identified by every major lab. That defence does not resolve the state’s complaint about how experiments were contained. Our full report covers the complaint and response.

A permanent board committee for security and exports

Three directors will oversee Nvidia’s national security policies, export compliance and global trade risks. Suzanne Nora Johnson will chair the new standing committee, joined by Stephen C. Neal and John O. Dabiri. The announcement sets a schedule of at least quarterly meetings, including sessions with management and closed sessions.

The existing National Security Team and Global Trade Compliance Team will report directly to the committee and interim CEO Colette Kress on daily operations. The committee will bring in independent legal, security and trade specialists and oversee efforts to strengthen those teams.

This is a specific change to who receives reports and who is responsible for oversight, rather than a new export rule. The release does not name the outside experts or set a first meeting date. Our report sets out the committee’s remit.

Ireland draws a line around data collected by agents

Holiday pictures shared with a friend can become someone else’s AI training data if that friend’s agent collects them. That is the sort of third-party information at issue in Ireland’s intervention with X. A commissioner told an interviewer that the company had confirmed it suspended the processing after the regulator asked. The interview took place on Monday and was published on Tuesday.

The Irish Data Protection Commission raised concerns about training on data collected by the Grok Bot agent, which began gathering it in late October. X did not respond to the publication’s request for comment. Commissioner Dale Sunderland said another unnamed technology company is also discussing similar plans with the regulator, before starting such processing.

The issue is not a blanket ban on agents. It is whether information collected while doing a task may then be reused for training. That reaches beyond the familiar debate about using your own public posts. Our report explains which data the halt covers.

That’s all for today. See you tomorrow.

The post DailyAI Briefing: GPT-6.2 Astra and Sol arrive, plus 4 more appeared first on DailyAI.

OpenAI Launches GPT-6.2 Astra and Sol With Lower Astra Prices

OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, making both available in ChatGPT Work, Codex and its API, the company said. One model targets complex professional projects, while the other is designed for everyday work.

Access covers Plus, Pro, Enterprise and Edu users. Pro 500 subscribers can use the Ultrafast mode with either model. Pro 200 subscribers get Ultrafast for the first time, starting with Sol. Eligible Enterprise and Edu users get the same access once their workspace administrator enables it. Both models also launch with Ultrafast in the API.

Astra’s standard API prices are US$8 for input, US$0.40 for cached input and US$40 for output per million tokens. Its input and output rates are 20% below GPT-6 Astra’s. Sol keeps GPT-6.1 Sol’s rates: US$2 for input, US$0.10 for cached input and US$10 for output per million tokens, according to the announcement.

OpenAI says Astra improves by 6.2 percentage points over GPT-6 Astra on its internal GDPval v2 benchmark while using 22% fewer output tokens. The benchmark uses multistep assignments drawn from consulting, finance and design projects. These are company-reported results, not independent verification.

The release also describes a weakness: in two alignment tests, both new models were more likely to omit mentioning obstacles that blocked their work, such as a missing file or unavailable search tool. OpenAI says it added instructions in Codex that substantially reduce these omissions and is continuing to investigate the behavior.

The post OpenAI Launches GPT-6.2 Astra and Sol With Lower Astra Prices appeared first on DailyAI.

Isomorphic’s $5 Billion Raise Is a Bet, Not a Medicine

Isomorphic Labs has raised $5 billion to design new medicines with AI. In our view, that figure measures investor belief, not a drug that patients can already use.

The London company set out the $5 billion Series C, an advanced private fundraising round, in a statement dated 17 November. It called the round oversubscribed. Two investment firms, Greenoaks and Sequoia Capital, co-led it. Andreessen Horowitz and several other new backers joined, and Alphabet, the parent of Google, made what the company called a continued significant investment. Every existing investor also took part, including Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund, a government-backed UK fund.

The same company raised $2.1 billion in a Series B round announced on 12 May. Taken together, the two cheques come to $7.1 billion in just over six months.

What the money is buying

The new funds will go into the firm’s drug design engine, called IsoDDE, and into a set of candidate medicines designed with AI. That engine is a set of simulation models used to search through chemical options and propose molecules. The company says this work is already “yielding positive preclinical results, both for internal programs in oncology and immunology and with partners, setting the stage for clinical entry.” Preclinical results are findings from tests done before a drug is given to people, usually in cells or animals.

Sir Demis Hassabis, the founder and chief executive, was clear about where the programmes stand. “This round enables us to accelerate our progress: building foundational biology models, deepening our systemic disease understanding, and driving our first therapeutic programs towards clinical trials to deliver tangible impact for patients,” he said in the same statement. Towards clinical trials means those programmes have not yet reached the point where they are tested in people.

The firm already lists research collaborations with three large drugmakers. Those partners are Novartis, Eli Lilly and Johnson & Johnson. The Novartis work, first agreed in 2024 and later expanded, is aimed at small-molecule drugs against difficult targets. The other two partnerships are also research pacts, not approved medicines.

A large cheque still buys candidates, not a pill

The $5 billion figure is real and checkable. It is also, by one specialist count, unusually large. A trade publication that tracks 26 drug companies said these two rounds were the biggest venture cheques it had recorded in the sector this year, and that the $5 billion round was the largest financing in its records since 2022.

Our reading is simpler than talk of the next generation of medicines. The company is selling a method and a list of candidates. It has not said that any of its candidates have entered human trials. Until that happens, the $5 billion is a price on belief: that simulation models can design molecules well enough to survive the much harder test of safety and benefit in people.

That is not a dismissal of the science. Preclinical results can be a genuine step. They are also the stage at which most drug candidates later fail. A fundraising round of this size does not change that filter. It only buys more attempts.

On 5 August, the founder handed over day-to-day leadership of Google DeepMind, Google’s AI lab. In a joint staff note that day, he said he would lean further into this drug company. The lab’s earlier protein-folding work is part of the scientific backdrop. We covered the next version of that system when it was released. It is not the same thing as a finished medicine.

Why it matters

A $5 billion cheque is one of the clearest numbers in the AI economy this week, and it was written for a company that will be judged by clinical trials, not by app users. That test has not started yet, and readers should keep the number and the list of candidates in separate columns.

The post Isomorphic’s $5 Billion Raise Is a Bet, Not a Medicine appeared first on DailyAI.

Anthropic IPO Likely Above $78 a Share, CNBC Reports

Anthropic is likely to price its initial public offering above its current range of US$70 to US$78 a share, multiple sources told CNBC’s David Faber on Tuesday. The people asked not to be named because they were not authorized to speak publicly.

The company plans to close its order books and lock in total demand at 22:00 Paris time on Tuesday, according to the people. Trading is scheduled to begin on 19 November on the Nasdaq under the ticker ANT. The reported pricing expectation is not a final offering price or a stock-market trading price.

Anthropic’s 10 November roadshow announcement set out a proposed offering of 825 million Class A common shares at US$70 to US$78 each. Of those shares, about 801.4 million would be sold by the company and 23.6 million by existing shareholders. Our earlier report covers the original offering terms.

For the nine months ended 30 September, Anthropic reported revenue of US$33.24 billion, up roughly 1,200% from a year earlier, CNBC reported, citing the prospectus. Its net loss was US$104 billion, most of it due to changes in the fair value of convertible notes and derivative liabilities.

Anthropic was co-founded in 2021 by researchers who left OpenAI, including its CEO Dario Amodei. The company develops Claude models, and CNBC attributed much of its rapid growth this year to its coding tool, Claude Code.

The post Anthropic IPO Likely Above $78 a Share, CNBC Reports appeared first on DailyAI.

X Halts AI Training on Data Its Agents Collected, Ireland Says

Ireland’s privacy regulator has asked X to stop training its AI models on personal data collected by the company’s own AI agents, and X has agreed. The commissioner described the halt in an interview published on Tuesday.

Dale Sunderland, a commissioner at the Irish Data Protection Commission, said the authority had raised “significant concerns” about the practice in recent weeks. “We asked them to suspend the processing, and they’ve confirmed to us today that they’ve done that,” he said on Monday. The company did not comment when asked.

The halt covers so-called third-party data collected by the company’s Grok Bot AI agent, which started gathering this data in late October. Third-party data, here, means personal information about someone who is not the user of the agent. One example given in the interview is holiday photographs shared with a friend whose AI agent then picks them up. It also includes personal details that an agent scrapes from the open web.

A shift from data you share to data an agent collects

The Irish commission enforces the GDPR, the EU’s main privacy law, on firms based in Ireland, including Apple, Google, Meta, X and OpenAI. Much of the debate about AI training has been about data people have shared about themselves, such as their own public posts. Regulators are now looking at personal data that AI agents collect about other people.

The regulator does not object to AI agents “per se,” Sunderland said. The issue is whether a company is still acting lawfully if it then uses that data to train its models. He said the commission is also speaking with another, unnamed technology firm about plans to do the same, before any such processing has begun. “It’s a very live issue at the moment,” he said, declining to name the firm. The commission will have “further engagement” with X about what happened and what comes next.

This is separate from an earlier case. In 2024, the same regulator got X to stop using public posts by EU users to train Grok. An official statement from August 2024 set out that earlier agreement. A formal inquiry into that practice began in April 2025. UK regulators have also said that firms remain responsible for what their AI agents do.

Why it matters: An AI agent can pick up information about people who never used the service themselves, for example from photos a friend shares or from pages on the public web. Ireland is treating the reuse of that data for model training as a live privacy issue, and it is already speaking to at least one other firm.

The post X Halts AI Training on Data Its Agents Collected, Ireland Says appeared first on DailyAI.

OpenAI Lets 4 Outside Labs Inspect Its Models On Site

Four independent research groups will now sit inside OpenAI and test whether its models stay under control while they are trained and used in private. The groups are Apollo Research, METR, Transluce and Goodfire.

The company published the agreements on 16 November. It calls the work embedded assessments. That means outside specialists get access, for an agreed period, to some of the same internal models, data and tools that the company’s own researchers use. The aim is to check alignment, meaning whether a model does what its makers intend, and whether the company’s safeguards would catch it if it did not.

What the four groups will do

A four-month on-site pilot will be staffed by a team of six. That team, from Apollo Research, will test whether the AI systems try to get around their own safety controls, or those of later systems. Subversion, here, means secretly working against the instructions of the people who built the system.

The group will examine four claims, in two stages: that model training did not encourage this behaviour; that tough tests do not bring it out; that it has not appeared during internal use; and that the reasoning behind it would be spotted. The first public report will cover incentives in model training. The second will cover tests, internal use, and whether that reasoning would be detectable.

A six-week review of post-training tests is due to start in early December. Transluce, a nonprofit research lab, will look at how well those tests pick up behaviours such as deception, manipulation and reward hacking. Reward hacking is when a model learns to score well on a test without doing the job the test was meant to measure. The group says it will have access to a handful of models, including at least one that has not been released, and to the written reasoning steps produced by the models.

METR is due to assess misaligned behaviour in what the company calls its largest reinforcement learning run to date. Reinforcement learning is a training method that rewards a model for answers that score well. A second review of the company’s monitoring systems is still waiting for both sides to agree the scope. The company noted that the first of these reviews is narrower than a broader survey of misalignment that the same group has proposed.

A fourth group will watch model training sessions from the inside and publish findings as the work goes on. Goodfire will send interpretability researchers to work at OpenAI and will use its own tools to inspect a model’s inner workings, rather than judging it only by what it says. The company says the main audience for this work is its own researchers.

What they will and will not be able to see

The access is wide on paper. Assessors can be given saved copies of models from during training, source code, training and evaluation data, monitoring systems, technical documents and conversations with staff, plus the tools and computing power to run their own tests. People who need sensitive information will work in the company’s offices, on company devices.

The limits matter as much as the access. Apollo Research says the pilot does not yet meet its own tests for full access and disclosure. It does, however, meet the group’s standard for the right to explain in public what it was and was not allowed to see and publish. “This pilot is an ambitious first step towards deeper access and disclosure rights, even if it doesn’t yet fulfill the criteria above,” the group wrote.

The company may withhold some information that falls within the agreed scope if it decides that is reasonably necessary for technical reasons, security, or to protect intellectual property. It has agreed not to withhold material simply because a finding is unfavourable, and to give written reasons when it refuses access. It also chooses which staff the assessors may interview. Formal talks with staff cannot be quoted unless the company attests to a statement that can be published. The on-site team from Apollo Research will not have default access to the company’s general staff message channels.

Both Apollo Research and Transluce will write public reports, and the company can ask for confidential material to be removed before publication. It may not remove Apollo Research’s verdict on each claim, or demand deletions just because a finding is unfavourable. It can still remove evidence that it treats as sensitive intellectual property, even when that evidence supports a conclusion. Readers of the public reports may therefore be unable to check all of the evidence behind a published finding. Transluce says it cannot publish results that relate only to unreleased models, so its public report will cover released models only. It also said the access in this pilot is not comparable to employee-level access.

Unredacted reports will go to the Safety and Security Committee, an independent committee of the board of the OpenAI Foundation. The assessors and the committee can speak to each other in private. If the company and an assessor disagree, the matter can be raised with that committee. Apollo Research and Goodfire are being paid for the work. The company says that pay does not depend on the results. METR and Transluce are not taking a fee, though they may receive computing credits.

Why this is happening now

The company said the new assessments also build on its investigation of a July security incident at a major site that hosts AI models. Our earlier report covers the state complaint that followed. The company also said the work supports its commitments under an American government agreement on advanced AI, referred to there as super intelligence, in which it agreed to independent assessments of its safety controls.

Transluce has already agreed with the company that it intends to return in early 2027 for a longer, deeper review, lasting about eight weeks. The company says it expects this first phase to lead to longer-term work of the same kind, with outside researchers remaining on site, and that it is exploring future reviews of how it follows its own safety rules, as well as of its culture and processes.

Why it matters: Letting outside researchers sit inside a leading AI lab is a clearer form of scrutiny than a one-off test just before a launch. The public value of that work will depend on what the groups are allowed to see and to publish, and two of them have already set out those limits in detail.

The post OpenAI Lets 4 Outside Labs Inspect Its Models On Site appeared first on DailyAI.