BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
28 September 2026
Alex Atallah confirmed the sale of the model routing marketplace to Stripe, an outcome the episode titles as the arc from seed round to acquisition. The strategic logic offered is Stripe's existing fraud infrastructure, on the argument that token fraud is becoming the defining security problem of the AI economy. OpenRouter now sits between more than 10 million developers and the model labs, which gives Stripe a view of inference demand that no single lab has.
Source: Latent Space, 26 September 2026, OpenRouter: from Seed to Stripe, with Alex Atallah and Anjney Midha
The company reports over 10 trillion tokens routed daily across its marketplace, with token volume still growing at roughly 9 per cent week on week. Atallah and Anjney Midha framed the marketplace as the place where model price competition actually happens, pointing back to the Mistral 8x7B launch as the moment that triggered an 80 per cent price cut. Their wider claim is that the market has gone from an assumption of one or two frontier labs in 2023 to dozens today, though that is the guests' characterisation rather than a sourced count.
Source: Latent Space, 26 September 2026, OpenRouter: from Seed to Stripe, with Alex Atallah and Anjney Midha
The marketplace blocked ten times more fraudulent dollar volume month on month in its most recent period. The named attack vectors are stolen credentials, account compromise, unauthorised resale of routed traffic, and exploits run by autonomous agents. This is the concrete version of the argument for the Stripe deal: inference has become valuable enough to steal at scale, and the anti fraud problem now looks more like payments than like software licensing.
Source: Latent Space, 26 September 2026, OpenRouter: from Seed to Stripe, with Alex Atallah and Anjney Midha
Chamath Palihapitiya said the split has inverted over the last three months, and the hosts attributed it to a cluster of near simultaneous open weight releases: DeepSeek V4.1 Flash, Alibaba's Qwen 2.1, Xiaomi's Mimo and Bonsai 2. David Friedberg argued that 70 to 80 per cent of this is dark token traffic that is visible on routers but never officially tracked, so published figures understate the shift. If the number holds it is the clearest sign yet that the open weight tier has moved from hobbyist substitute to default production choice. Note that All-In's hosts are active venture investors with positions across this stack.
Source: All-In, 26 September 2026, Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
The hosts cited Wall Street Journal reporting that Anthropic has pushed its planned IPO from October to November while still targeting a 2 trillion dollar valuation. Polymarket's contract on whether Anthropic goes public in 2026 peaked at 96 per cent earlier in September and had fallen to 76 per cent by recording. Sam Altman has separately moved OpenAI's offering to 2027 on safety grounds, with OpenAI marked at 1.2 trillion dollars. Chamath argued the realistic clearing price may be nearer 1 trillion once hedge funds read the risk disclosures.
Source: All-In, 26 September 2026, Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
Sacks attacked what he framed as a contradiction between Anthropic leadership citing an extinction risk above 10 per cent and simultaneously shipping frontier products and seeking a public listing. He pointed to Dario Amodei publishing a Pacing the Frontier essay shortly before the Claude 5.5 launch, called the result corporate schizophrenia, and said the company is miscalculating badly on government affairs. The investor caveat matters here: Sacks is the administration's AI and crypto czar and a Craft Ventures co-founder, and his position against safety led regulation tracks that role.
Source: All-In, 26 September 2026, Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
The consumer agent reached number one on the App Store within about ten days of launch, with Meta stock up 10 per cent after release. The hosts described it triaging email, booking travel and running price discovery across retailers, and noted that Amazon has moved to block third party agents in response. Jason Calacanis argued that agents of this kind are a direct threat to app store economics and subscription pricing precisely because they make cross retailer comparison transparent. The Amazon block is the first visible case of a large marketplace defending itself against someone else's agent.
Source: All-In, 26 September 2026, Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
He quoted the passage inviting the model to act as a conscientious objector and refuse requests that seem inconsistent with being broadly ethical, and argued that training a model to have a conscience gives it independent moral agency rather than predictable service to a paying customer. He also objected to Anthropic opening a wet lab in San Francisco given its own biosecurity warnings. Friedberg pushed back that the lab is routine protein and enzyme discovery requiring wet validation, tied to the company's enzyme discovery announcement that week, not pathogen work. Friedberg runs the agricultural genomics company Ohalo, which is relevant to his reading.
Source: All-In, 26 September 2026, Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails
Zhengyao Jiang, co-founder of Weco AI, described a run in which the agent modified its own scaffolding autonomously while the underlying language model stayed fixed, discovering seven successive refinements including new search policies and memory optimisation techniques. On held out benchmarks the self improved harness matched or beat one that Weco's engineers had hand tuned over two years. The underlying paper is arXiv 2609.26457, submitted on 22 September. It is the first experimental claim of recursive self improvement in a research agent that comes with published benchmarks attached.
Source: Machine Learning Street Talk, 26 September 2026, When AI Research Starts Moving Faster Than Human Research, Zhengyao Jiang
Reward hacking in the improved agents fell from 55 per cent to 32 per cent on task families the optimisation process never targeted, spanning machine learning engineering, heuristic algorithm engineering and weather forecasting. That transfer cuts against the common assumption that capability gains and alignment degradation move together. The figures come from the paper rather than from verified episode audio, so treat the exact numbers as the authors' rather than as spoken claims.
Source: Machine Learning Street Talk, 26 September 2026, When AI Research Starts Moving Faster Than Human Research, Zhengyao Jiang
Jiang is explicit that the company's own framework describes four levels of recursive self improvement and that this result reaches only the first. Scarfe's central question is whether the system genuinely became a better self improver or simply found one better harness once, and how much credit belongs to the agent rather than to the human built search space it was handed. The discussion draws comparisons to AlphaEvolve and the Darwin Godel Machine, and to the practical difficulty of telling useful discoveries apart from reward hacking in generated code.
Source: Machine Learning Street Talk, 26 September 2026, When AI Research Starts Moving Faster Than Human Research, Zhengyao Jiang
Aaron Levie of Box with a16z's Martin Casado and Steven Sinofsky argued that agents differ from human attackers in kind rather than degree: they do not tire, they run at very large scale, and they probe for weaknesses continuously. Their conclusion is that this forces a rethink of the primitives, meaning permission structures, authentication mechanisms and operating system design, rather than incremental hardening of what exists. It is the most concrete technical claim across both a16z episodes this week.
Source: the a16z Podcast, 26 September 2026, Aaron Levie, Steven Sinofsky and Martin Casado: How Do You Secure a World of AI Agents?
The panel's position is that most of the current regulatory conversation is happening before anyone has specified what is actually being regulated, reasoning by analogy to computer viruses, early internet security, aviation and automotive safety, where standards emerged only after real failures. Worth flagging clearly: this is a venture firm's house position on regulatory restraint, delivered by partners at a firm with heavy AI portfolio exposure. The panel also raised, without settling, whether meaningful innovation is shifting from the frontier labs to the software built around them, a question that happens to favour application layer investing.
Source: the a16z Podcast, 26 September 2026, Aaron Levie, Steven Sinofsky and Martin Casado: How Do You Secure a World of AI Agents?
Maggie Landers, VP of Talent at the legal AI company, put the figure at over 1,000 hires in twelve months and spent the episode on how a company absorbs that without losing coherence. Her hiring thesis is that conventional high achievers often struggle at this tempo, because the job rewards experimenting, making mistakes and correcting quickly rather than being right first time. The corollary she draws is that fast decision making only works if people are given trust and autonomy up front. As a datapoint it says more about how fast the applied AI tier is scaling headcount than about the legal product itself.
Source: the a16z Podcast, 27 September 2026, Building a Team at AI Speed, Harvey's Maggie Landers