A Monday with a single theme pulling the rest along: cost. Model pricing, open models inside the enterprise and even inference speed all land in the same conversation.

Every time someone asks me how much an AI feature will cost, I know the answer I gave last month is already stale. The AI agent cost management guide helps keep that math current.

I read Max Woolf’s survey of the price cuts from OpenAI and Anthropic between June and September. GPT-5.6 Luna dropped about 80% in July, and the GPT-6.1 family launched roughly 75% cheaper than GPT-5.6. Anthropic also lowered its cache price and launched Opus 5.5 at a lower cost than Opus 5.

The detail that caught my attention most is in the cache. In agentic coding use, cache reads account for 96% to 99% of token spend. That is why cutting the cache price by 75% brings the bill down almost proportionally.

Worth remembering that the explanations for these cuts, such as competition from open models and efficiency gains, are the author’s hypotheses. He himself says there is no way to know the real margins. Vercel’s data also shows that cost per token fell but total spend rose, because usage grew.

What changes for product builders

On the product side, I would take two lessons. The first is to stop planning around a fixed price per token. What matters is cost per completed task, and in it the cache design weighs more than it seems. The AI agent cost and task-budget template is a starting point. The second is that a low price doesn’t mean a smaller budget. When it gets cheap, the team uses more, and the roadmap needs to anticipate that effect.

I’m optimistic about this. The lower the cost of each call, the more use cases fit inside a product, and those who measure well benefit first. To measure well, the guide on how to evaluate AI agents shows how to structure the evaluation.

For those curious to see the full price timeline, here is the reading.

The rest of the radar

  • IBM CEO on on-prem open models — for common enterprise tasks, open models would cost 10% to 25% of the frontier token with about 90% of the effectiveness, according to the CEO, and the bottleneck becomes operating the infrastructure. The Register.
  • The state of context engineering — an agent’s quality depends on context as much as on the model, and improving it is usually the cheapest path. dbreunig.
  • JetBrains Thinkrail — an open IDE with parallel agents, each on its own branch, and specs as the source of truth; a reference for testing a spec-first flow, still at version 0.1.2. GitHub.
  • Cerebras and trillion-parameter models — the company talks about 1,000 to 2,000+ tokens per second, which would open real-time experiences with large models, but with no price disclosed. Cerebras.

That’s it for today. More tomorrow.