Token Cost Radar

1How much cheaper does a smaller model need to be?

Reuters reports that OpenAI cut the price of its GPT-5.6 Luna model by 80% and its mid-tier Terra model by 20% as businesses scrutinize rising AI spending. OpenAI attributed the reductions to efficiency gains, while competition from less expensive Chinese models continues to pressure the market.

2EY says invisible routing reduced token consumption by up to 60%

Business Insider reports that EY's internal router automatically assigns requests to suitable models based on capability and cost. EY also gives departments token budgets and requires approval for overages, combining routing with attributable governance.

3Atlassian puts employee AI spending into monthly wallets

The Guardian reports that Atlassian introduced monthly AI allowances ranging from $500 to $2,000 per employee, with additional funds available by request. The system gives the company a clearer view of variable usage while avoiding a blanket ban on expensive experimentation.

4Open weights remain a hedge, not a free lunch

Reuters Breakingviews says organizations are considering open-weight models for lower prices, customization, and data sovereignty, while warning that hosting, chips, energy, security, and engineering can replace the API bill with an infrastructure bill.

CHILL-Harness: Counterfactual Harness Learning for Efficient Reasoning in Long-Horizon Agents

This arXiv paper treats agent orchestration as an adaptive causal-learning problem. Its harness authorizes workflow changes only when evidence suggests they will improve execution, and the authors report reduced token consumption and execution time while preserving or improving task success across information-seeking, software-engineering, and terminal tasks.

Why it matters: Fixed agent policies can purchase the same planning, verification, and context for every task. Adaptive orchestration turns those costs into decisions rather than rituals.

Keeping the Cache Warm Pays

This arXiv paper measures whether replaying a stable prompt prefix during long tool calls can preserve provider caches. It reports that keepalives cut the post-pause request cost by up to 12.5 times in tested conditions, while finding that the economical interval depends heavily on each provider's cache-retention behavior.

Why it matters: Agent pauses can quietly convert discounted cached input back into full-price input. Cache lifetime is therefore part of workflow economics, not merely an API implementation detail.

Tool Attention Is All You Need

This arXiv paper proposes dynamic tool gating and lazy schema loading to reduce the MCP or tools tax. In a simulated 120-tool benchmark calibrated to public deployment audits, it reports reducing per-turn tool-schema tokens from 47,300 to 2,400, while clearly labeling end-to-end cost and performance figures as projections.

Why it matters: Tool-surface bloat is a recurring charge paid before the agent uses a tool. Loading full schemas only for relevant capabilities converts that fixed tax into selective spending.

Token Reduction Is Not Cost Reduction

This arXiv study analyzes 2,848 paired provider-billed coding-agent runs and finds that reducing visible tool-output tokens does not necessarily lower total cost. One tested compression arm removed 38% of estimated tool-output tokens while costing 6.8% more, partly because caching, retries, and task completion altered the economics.

Why it matters: Tokenminimizing requires a success-adjusted cost measure. The smallest prompt can still be the expensive one when it damages cache reuse or removes evidence needed to finish the task.