Saturday, 12 September 2026

OpenAI has released ChatGPT for Financial Services, a tailored version of its enterprise product that pairs its most advanced model, GPT-6 Astra, with built-in access to financial data providers [1][2]. The company says the product was built with "design partners" Morgan Stanley and Evercore and is aimed at the research, modelling and pitchbook work that has traditionally fallen to entry-level investment bankers [1][2]. OpenAI's vice president of product, Nick Turley, told CNBC the tool takes on tasks handled by "analysts and associates" [2].

The company says data from Daloopa, PitchBook, LSEG News and Crunchbase is indexed and hosted by OpenAI, so answers can carry granular citations back to specific tables and passages [1]. Firms centrally manage access, and administrators can publish Excel, Word and PowerPoint templates so valuation models, research notes and pitchbooks follow their in-house formats [1]. Sample workflows named by the company include LBO modelling, buyer screening, earnings analysis and pitchbook preparation [1].

For a small team that already lives on ChatGPT, the change is not the model but the plumbing around it: connected data, firm-managed access and templated outputs baked into an enterprise plan. The question for buyers is which data providers you rely on today, and whether OpenAI's indexed set matches [1][2].

Skild AI has launched S1, a robotic foundation model that its chief executive says lets a robot learn a new multistep task from a single video demonstration [1][2]. The company says S1 reached about 66% success per step on new long-horizon tasks it had never seen, compared with roughly 9% for language-prompted comparison systems on the same benchmark, and can execute jobs up to 10 minutes long with dozens of manipulation steps [1][2]. In one demonstration, Skild says it turned an 11-minute video of a plant-potting routine into an autonomous run without any task-specific retraining [1].

Alongside the model, Skild disclosed that its software now runs on hundreds of robots at more than 60 customers, and that annualised revenue has reached about $100 million, roughly 10 months after its first commercial deployment [1][2]. Skild's chief executive Deepak Pathak says "learning by experience, and not preprogramming, is the step change that has happened in robotics" [1]. The company says Foxconn is using the model on the factory floor to help assemble Nvidia Blackwell AI systems, with dual-arm manipulators installing a busbar and fastening 16 screws in one workflow [1][2].

For anyone watching robotics costs, the useful signal is not the demonstration but the revenue run rate: a foundation-model business for physical work now has customers paying at meaningful scale, on top of a stack that leans on Nvidia's simulation and inference software [1][2].

Twenty-five winners of the Fields Medal have signed a statement, published on the mathematician Terence Tao's blog on 11 September, that calls the race by AI labs to solve famous math problems a "severe misalignment" with what mathematicians do [1][2]. The declaration, titled "A Severe Misalignment of AI in Mathematics," argues that hurried announcements of AI solutions bypass the verification, documentation and attribution that let a proof become part of the field [2]. Signatories include Terence Tao, Peter Scholze, Maryna Viazovska, June Huh, James Maynard, Maxim Kontsevich and Manjul Bhargava, alongside 2026 medalist Yu Deng [2].

According to TechCrunch, OpenAI has since withdrawn its sponsorship of a CalTech math event, and the mathematicians accuse the company of pressuring NYU professor Tristan Buckmaster over the credit for an important result [1]. The statement frames the concern as the "mass production" of answers that skips the human chain of understanding, warning that "without mathematicians willing to develop and integrate AI-generated ideas into the canon, those ideas would never become fully alive" [2].

For anyone using AI to write, code or reason with, the row matters less as a mathematics dispute than as a template. It is the first time a discipline's most senior figures have publicly said the pace and manner of AI announcements is doing damage to the record. Expect similar objections from other technical fields where credit and verification are load-bearing [1][2].

The Chinese AI company Moonshot AI has told investors it is targeting about $2 billion in annualised revenue by the end of 2026, roughly double the $1 billion run rate it reported for August [1][2]. Moonshot's numbers climbed sharply this year: about $200 million in April, $300 million in June, then a jump to $1 billion in August, according to reporting from Crypto Briefing and TechCrunch [1][2]. TechCrunch attributes the acceleration to Moonshot's open-weight Kimi K3 model, released this summer, which shows around 300 billion tokens generated per day on the OpenRouter aggregator [1].

Crypto Briefing puts Kimi K3 at 2.8 trillion parameters and a 1 million-token context window, and dates its launch to 16 July 2026 [2]. The same report says Moonshot has raised money at successive valuations of $20 billion in May and about $35 billion in July, and has filed confidentially for a Hong Kong IPO of about $3 billion [2].

For context, TechCrunch reports that OpenAI and Anthropic are running at roughly $40 billion and $65 billion in annualised revenue [1]. Moonshot is much smaller, and its open-weight approach carries thinner margins than closed-model rivals, so the $2 billion target is a growth claim, not a profitability one — a useful sign that Chinese open-weight labs are converting free-to-run models into real revenue at speed [1][2].

Anthropic researcher Jacob Coxon has resigned, telling colleagues on Slack and the public on X that Anthropic and OpenAI "are racing straight to self-improving superintelligence and gambling with our lives" [1][2]. Coxon spent three years working on AI model training research at both Anthropic and OpenAI, according to Fortune, and wrote that "neither company is acting responsibly" [1]. He said OpenAI staff "have not deeply internalized the civilizational stakes" and that Anthropic staff understand the risks but are "locked in a race to get there first" [1].

The story reached a wider audience because Anthropic's alignment science lead, Evan Hubinger, publicly agreed with the tone of the warning [1][2]. Fortune quotes Hubinger writing on X that he and colleagues "really do earnestly believe AI could kill all humans" and estimating the chance is "greater than 10 percent within the next decade" [1]. TechCrunch's Equity podcast frames the moment as unusually loaded because Anthropic filed confidentially for an IPO in June and is preparing for a listing this autumn [1][2].

For a working operator, the immediate reading is not the philosophy but the signal: it is the second time in a year a senior safety hire at a frontier lab has publicly said the pace of development scares them. That is a fair reason to keep production controls — approvals, logs, kill switches — in place around any agentic system you actually deploy [1][2].