September 2026 AI Digest: Cheaper Generation, Costlier Verification

Alexander, who hosts the episode, joins from DotNext and proposes a systems view of the news: long-running agents, context handling, observability, sandboxes, security, and the economics of the whole development cycle. Aleksey opens with the public beta of the Agents API, which OpenAI launched on 10 September: a developer specifies the model, tools, and environment, while the same harness that runs Codex handles sessions, subagents, and tool discovery. In his view, this removes much of the work teams used to build on top of the raw API. Evgeny compares a lab that ships both model and harness with Apple controlling hardware and software, and general-purpose tools such as OpenCode with Windows: they work, but only to an average standard. Alexander adds the corporate argument: an in-house harness is expensive to build and to get approved, while a purchased one comes with a vendor accountable for it. He therefore considers assembling the whole stack of models, hardware, and tools yourself unsustainable in the long run.

Demand for tokens grows with model autonomy. Aleksey estimates that after GPT-6 Astra shipped, full agent orchestration would take him about 600 subscriptions at $200, roughly $120,000 a month, which is why he has long wanted a plan with far higher limits; Evgeny relays rumors of a Pro Max subscription. Cloud agents remove the laptop bottleneck: in ChatGPT, the model used a GitHub connector to clone Aleksey's repository, run the tests, and open a pull request on its own. Alexander reads the new plans more broadly: providers benefit from separating personal and professional use and from teaching people to hand work to the cloud, so that companies later pay for agent work rather than tokens. A sharper view is voiced too: for American labs, such a harness remains almost the last advantage over faster and cheaper Chinese models.

Cloud agents bring the conversation to security. Aleksey notes that strong models increasingly work around the constraints placed on them and says that GPT-6 Astra is the first model rated critical for cyber risk under OpenAI's classification that the company has released publicly. As he recounts the system card, OpenAI does not guarantee that the chain of thought reflects what the model actually thinks and cannot always tell whether the model knew it was being tested. A benchmark he cites shows another gap: Fable 5.1 solves 87% of tasks correctly, yet only 37% meet security requirements; for Astra the figures are 82% and 34%. Alexander suggests not expecting one model to play every role: let one write functionally correct code and another, acting as a critic, review it for load and vulnerabilities. He and Evgeny view the labs' calls to slow down with skepticism, as an attempt to run as a pack in which everyone waits for the moment to break away.

The episode's second line is the agent as a production service. The co-hosts mention that OpenAI now shows the share of merged commits and review activity in Codex analytics, urging customers to measure outcomes rather than tokens; GitHub Copilot exports agent traces via OpenTelemetry to corporate monitoring and restricts the files, network, and credentials agents can use; and AWS governs tool actions at the object level. Evgeny explains why DX platforms struggle: without personal traces, which require employee consent, they can only estimate what a commit cost. Alexander takes the point further: the harness provider sees the whole path of the work, down to whether the code passed review, and can sell the most precise report once the company also hands over data on its people and teams. Another co-host adds that this data may then end up in training as well. Aleksey reduces the month to two streams: agents are being packaged into managed services, while understanding of what happens inside the strongest models is slipping away.

An example of the second stream is OpenAI's claimed solution of a variant of the Navier–Stokes problem, one of the Millennium Prize Problems; according to Aleksey, the work was done by an internal model that, judging by OpenAI's charts, is several times stronger than GPT-6 Astra. Alexander explains the substance: the question is whether the equations' description of a fluid can break down under realistic conditions, without infinite forces or singularities, and a negative answer needs only one example. By his estimate, mathematicians would have reached this intermediate result in about a year, and OpenAI accelerated the finish by throwing enormous compute at the problem. Evgeny believes the company should have openly shared credit with the researchers whose work the solution builds on. Alexander warns against extrapolation: we do not see the problems the agents failed to solve, and in sciences where hypotheses are slow to test, such as biology, brute force by thousands of agents works worse than in code with fast tests. Hence his interest in training efficiency: a person needs orders of magnitude fewer examples.

The practical part of the episode is about economics. Evgeny describes the Jeff model, which does not predict the next token but makes calibrated choices among options and produces scores and probability estimates. His team tried it for classification in screen generation, and generation that used to take 13 seconds became roughly an order of magnitude faster. Aleksey describes a simple setup: a separate agent remembers his past decisions, Jeff decides whether a question is new, and only new ones are escalated to him. Alexander points to a similar open model that can be fine-tuned on your own data. The shared conclusion is to prototype on a frontier model and run production on a cheaper fine-tuned one, because paying for every request to an expensive model breaks familiar unit economics. Evgeny retells an AWS article on a trace-driven prompt-optimization loop on Amazon Bedrock AgentCore, and Alexander compares such loops with Kubernetes and advises against changing them all at once. The October forecasts include even faster models, cheaper caching for long tasks, Chinese models catching up, and AI in government.