The Product Desk

The Information reports that Microsoft cut per-person AI budgets from as much as $100,000 a month to about $10,000. It also cut projected internal Claude spend by more than a third. Meta's Claude Code users fell from about 60,000 to about 30,000, partly because of layoffs. Your enterprise customers will likely copy this playbook: cap the heaviest users and route work to cheaper models. Microsoft's total payments to Anthropic stayed near earlier highs, so this is rationing, not retreat.

AINews reports that OpenAI's Decisions API charges from $0.10 per million input tokens with no output charge. Perplexity's open-weights decision model costs $0.02 per million. Google's Nano Banana 2.1 cuts the cost of an image to $0.034, down from $0.134. Every classifier, moderation and image call in your product now has a cheaper benchmark to beat. Reports disagree on open weights: Bloomberg says Reflection released a model Monday, while The Information says Beam's weights arrive later in October.

Over the coming weeks, OpenAI is adding its invisible textGrain watermark to eligible ChatGPT and Codex text for EU users, tied to the EU AI Act. Simplifying AI and Techpresso both report this. In the API, the watermark is off by default worldwide. If your features call OpenAI for EU users, the transparency decision is yours to make. Detection is also gated and fragile: Techpresso reports that swapping 10% of words cut detection from about 92% to 66%.

The Information reports that OpenAI told some partners it will no longer accept ChatGPT ads for image and audio products that compete with its own features, and Adobe is reportedly affected. ServiceNow's new Flow is expected to compete with third-party chat layers built on ServiceNow. Microsoft also moved its engineers from Claude Code to its own Copilot CLI. Any platform that distributes your product and also builds features can cut off your access once you overlap.