This week I spent some time reviewing the cost of running AI in the products we build, and an OpenAI price cut pulled the thread for today’s edition. The rest of the radar ranges from sub-50ms TTS latency to a billing bug in a coding agent.

In brief

  • OpenAI cut GPT-5.6 Sol API pricing by more than 20%, its third cut in a few months.
  • In credit products, every AI call goes directly into the feature’s viability calculation.
  • When a model’s cost falls by 20%, 40% or 80% in a few months, previously unviable use cases can return to the roadmap.
  • The price war among providers turns model selection into a recurring product review.

Lower inference cost changes the roadmap math

OpenAI has just cut the API price of GPT-5.6 Sol, its frontier model, by more than 20%. It is the third cut in a few months.

For people who only use ChatGPT day to day, this may look like another piece of industry news. But for teams that build products and depend on AI to run features, this kind of move changes the math.

In credit product work, we live by unit economics. Every new AI-powered feature carries a cost per call, and that cost goes straight into the viability equation for anything we want to launch.

When the price of a frontier model falls by 20%, 40% or 80% in a few months, something that was unviable in January can become viable in August. Document-analysis automation, receivables triage and intelligent support all become cheaper to run in production.

And the move is not isolated. It is a price war among the major AI providers, also pushed by Chinese models entering the competition aggressively. Teams building products on top of these APIs benefit on both sides: more capability and lower cost.

What to review before bringing a feature back to the roadmap

The reminder is to revisit the roadmap regularly. That feature you dropped six months ago because of cost may be knocking again.

Before reopening the discussion, it is worth looking at three points:

  • cost per call multiplied by expected volume;
  • the quality and latency required by the use case;
  • customer value and the margin left after inference.

That review belongs to AI product management: putting cost, capability and outcome in the same conversation. AI for Product Managers helps structure the evaluation, while the state of AI in product management in 2026 provides context for a stack decision that does not get locked to the model of the moment.

If you want to see the full story, here is the Reuters report.

The rest of the radar

DeepSeek launches a fast vision model (v4-flash-vision-exp) — expands the options for affordable multimodal AI without relying only on OpenAI and Google. Read more

Anthropic opens Claude Academy — a free hub with around 20 courses and certificates, useful for accelerating AI adoption across teams. Read more

Claude Code gets a broad update (MCP, plugins, Remote Control) — more security and stability for teams that depend on the agent in internal workflows. Read more

n8n updates with a focus on agents and MCP — makes it easier to prototype and maintain AI automations without relying only on engineering. Read more

A developer builds a self-hosted “agentic software factory” — a build-versus-buy reference for coding-agent infrastructure. Read more

Autolith runs on a live runtime — a new approach to programming agents could change the UX of assisted development tools. Read more

Claudette removes Claude’s “generic blog” tone — a reminder that an agent’s voice and personality are a differentiation surface. Read more

A week using Codex more than Claude — a qualitative signal of where perceived quality is leaning among competitors. Read more

Nari Labs cuts TTS latency below 50ms — makes real-time voice experiences possible in support products. Read more

That is all for today. Have a great weekend, and see you in the next edition.