AI spend and model access come under tighter control

September 29, 2026

Microsoft introduced Copilot Credits, a usage-based billing layer for advanced Copilot agents and other AI workloads on top of its existing per-user subscriptions, and paired it with FinOps tooling to track that spend. This shifts Copilot further from a flat software licence toward a governed utility: buyers now have to watch which AI work is worth metered consumption, not just which staff get access. Because Microsoft sits inside ordinary office software, that change is likely to spread a new habit beyond one vendor - enterprise AI stops being something approved once per seat and becomes something budgeted continuously by workload.

OpenAI decided not to release Astra 6.1. That appears to make safety gating less theoretical than many model launches suggest: a frontier provider is showing that a stronger model can still lose access to the market if its behaviour cannot be kept inside bounds. For companies building on outside models, the effect is that availability itself becomes conditional - roadmaps, product choices and procurement cannot assume every announced generation will arrive on schedule.

Mistral opened a Munich hub for industrial and physics AI with teams working directly with German enterprises and described sovereign open-weight deployments on customers' own infrastructure under EU law. Open-weight means the model weights are distributed so customers can run and adapt the system themselves, not only call it through the vendor's cloud. We read this as another step away from AI being consumed only through remote APIs: local and jurisdiction-bound deployment is becoming a normal commercial offer, especially where industrial data and legal control matter.

Anthropic launched Claude Sonnet 5.5. Anthropic also changed its enterprise discounts so customers lose discounted rates after consuming their committed token volume. OpenAI and the Lenfest Institute expanded their newsroom AI programme with $5 million in cash and up to $5 million in credits and engineering support. Meta launched a Meta Enterprise Platform for business AI tools and hired MongoDB's CEO to lead it.