Today’s radar revolved around money: who charges how much per token and why that is changing so quickly. The day’s highlight follows that thread through an Anthropic decision, while the rest of the curation brought plenty of agent launches to complete the picture.

I have been following how AI companies set prices because it directly affects my work. I work on the product side of credit, and over the past few months almost every architecture decision has come back to the question: “how much will this cost per model call?”

In brief

  • Anthropic had announced a 50% increase in Claude Sonnet 5 pricing, effective September 1: from $2 per million input tokens and $10 per million output tokens to $3 and $15.
  • The increase was canceled. The promotional price became the standard price, permanently.
  • The most likely explanation is competitive pressure from models such as DeepSeek and GLM, which deliver comparable quality at a fraction of the cost.
  • For anyone building on language models, AI token pricing is now a strategic variable—and architecture needs to make model switching easy.

The price increase that never happened

This week brought a case that illustrates the game well. Anthropic had announced a 50% increase in Claude Sonnet 5 pricing, effective September 1. The promotional price of $2 per million input tokens and $10 per million output tokens would rise to $3 and $15.

The increase was canceled. The promotional price became the standard price, permanently.

The most likely reason is competitive pressure. Models such as DeepSeek and GLM have been delivering comparable quality at a fraction of the cost, leaving little room to maneuver for anyone building on an API and depending on predictable spending.

For anyone who wants to understand the full case, here is the news report.

AI pricing has become a strategic variable

This stands out because, from a product perspective, AI pricing is no longer just an operating-cost line. It has become a strategic variable, somewhat like interest rates or credit spreads: whoever moves first sets the market’s pace for a while.

For teams working in AI product management, that means tracking inference cost alongside quality, latency and availability. The price of each call can change the priority of an entire feature.

Architecture needs to follow the token war

For anyone building a product on language models, the message is direct. It is less valuable to stay attached to today’s supplier and more valuable to design the architecture for easy model switching, because the competition over token cost is unlikely to cool down soon.

This is a product management decision, but it is also about infrastructure and resilience. In workflows that use AI agents, model choice can change cost per task, product margin and the viability of a use case.

The rest of the radar

GPT-5.6, OpenAI’s new model — A new benchmark for intelligence and cost changes the competitive benchmark for anyone building on the OpenAI API. Read more

Salesforce and Anthropic launch “Claudeforce” — Shows the distribution model winning in enterprise AI: a deep partnership between model and CRM, not just a standalone API. Read more

OpenAI retires the DALL·E GPT from ChatGPT — A classic example of sunsetting a legacy feature in favor of product consolidation, with a clear sunset deadline. Read more

Google launches “Ask Gemini” in Google Chat — A UX reference for anyone designing an AI agent as a central “command line” inside a workflow already used every day. Read more

AWS turns Web Search into a managed Bedrock tool — Reduces the engineering cost of giving agents real-time web access, a ready-made option for teams weighing buy versus build. Read more

n8n strengthens agent workflows with MCP — A common tool for prototyping AI automation without heavy coding, with updates that speed up discovery and internal MVPs. Read more

DeepSeek open-sources “Harness” — An open-source, MIT-licensed agent framework that lowers the barrier for teams building coding agents without depending on a closed platform. Read more

PLG swaps the “human user” for an “agent” as the activation unit — Redefines activation, onboarding and pricing metrics, a central decision for any PM working on SaaS or AI products in 2026. Read more

Simon Willison explains “ChatGPT Work” — A useful reference for communicating distinct modes of use (chat for answers, agent for execution) within the same product. Read more

That is all for today. The week is starting with plenty of news to digest.