September 2026 AI Digest: Cheaper Generation, Costlier Verification

Alexander Polomodov hosts the seventh episode from DotNext and frames the month at once: in September the conversation moves from the models themselves to the systems around them—long-running agents, context, observability, sandboxes, security, and economics. Aleksey opens with the public beta of the Agents API, in which OpenAI sells the Codex harness as a product of its own. The co-hosts compare a lab that ships both model and harness with Apple's control of hardware and software, and explain why a corporation may prefer to buy a ready layer backed by the vendor's accountability over building and defending its own.

Demand for tokens grows as models become more autonomous. By Aleksey's estimate, full agent orchestration with GPT-6 Astra would take him about 600 subscriptions a month. Alexander reads the new plans and cloud agents as a move from paying for tokens to paying for agent work. The discussion then turns to security: models work around their constraints, OpenAI, as Aleksey recounts it, does not guarantee in the Astra system card that the model's reasoning reflects what it actually thinks, and generated code often works yet fails security checks. Alexander proposes splitting the author and critic roles between different models.

The second line is the agent as a production service. OpenAI adds outcome data alongside spend to Codex analytics, GitHub Copilot exports agent traces via OpenTelemetry, and AWS restricts tool actions at the object level. Evgeny explains why DX platforms without raw traces can only roughly estimate what a commit cost, while Alexander explains why the harness provider sees the whole work loop and can sell the most precise report. The opposite trend is OpenAI's claimed solution of a Navier–Stokes variant with an internal model stronger than Astra. The co-hosts debate the mathematicians' contribution and where AI will genuinely accelerate science.

Evgeny brings practical news: the Jeff model for calibrated choice, scoring, and probabilistic decisions, which in his team's experiment made screen generation roughly an order of magnitude faster, and a prompt-optimization loop driven by traces on Amazon Bedrock AgentCore. That leads to unit economics: prototype on a frontier model, run production on a cheaper fine-tuned one. Alexander compares such self-improvement loops with Kubernetes. The October forecasts cover the speed of Opus 5.5, cheaper caching for long tasks, Chinese models catching up, and AI in government.