Sources checked:
What changed
Between 22 and 25 September, three vendors announced changes that connect model price, persistent agents and spend control. OpenAI priced GPT-6 Sol and Luna at 50% below the promotional API prices of their GPT-5.6 predecessors and described discounts of up to 90% for eligible cached input. Anthropic priced Opus 5.5 at $4 per million input tokens and $20 per million output tokens, while claiming 40% lower cost than Opus 5 on typical workloads. Microsoft introduced a persistent Autopilot agent in private preview and paired usage-based billing with spending policies and task-outcome reporting. These are vendor announcements and vendor-reported efficiency claims, not independent proof of savings in a particular business.
Business implications · analysis
The practical buying question is no longer simply which model has the lowest token price or the highest benchmark score. A business should compare cost per successfully completed task, including retries, human correction, elapsed time, cache performance and failed actions. For any agent that can continue working without a fresh prompt, add a named owner, a narrow identity, minimum permissions, an audit trail, a spending ceiling and explicit stop conditions. Run a representative task set before scaling, and keep vendor benchmark results separate from your own operating evidence.
What to watch next
Watch the move from preview to general availability, the final billing rules, actual cache-hit rates and the amount of rework required after an agent says a task is complete. One possible path is that lower model and caching costs make more workflows economical. Another is that longer-running agents increase total volume, supervision and exception-handling costs enough to offset cheaper units. Treat both as scenarios to measure, not forecasts.
Primary sources
Company statements are attributed to the publisher. Our analysis is separate; forward-looking points are questions to monitor, not promised outcomes.