Planning your AI budget for next fiscal year: a checklist

Budgets for the next fiscal year are being drafted now, against a price environment that changed repeatedly this summer. Token prices fell on some tiers, held on others, and in at least one case are scheduled to rise. A forecast that multiplies today's price by projected volume will be wrong in both directions. This checklist is designed to make the number more robust.

Where prices stand

The checklist

1. Budget by tier. List the use cases that need a flagship model, those that suit a mid tier, and those a small, cheap model handles. AI Weekly's description of the GPT-5.6 lineup (Sol for the hardest problems, Terra for high-volume business tasks, Luna for summarization, drafting and routine automation) is a useful template ^[6].

2. Treat introductory prices as expiring. Gemini 3.7 Flash's rate is described as introductory ^[4], and DeepSeek has said it will raise API prices ^[5]. Model the post-promotion price as a scenario.

3. Also model price cuts, but do not bank them. Luna fell 80% three weeks after launch ^[1]. Cuts are a bonus, not a basis for approving spend.

4. Add an effort-level assumption. Opus 5 lets customers tune effort, with named levels "high," "xhigh" and "max" plus a lowest-effort setting, trading tokens for intelligence ^[2]. Decide the default per workload and size the budget on it.

5. Check the tokenizer. Anthropic's documentation says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text ^[3]. If you compare vendors on price per token, compare on a sample of your own text.

6. Use caching and batching where the workload allows. Anthropic's documentation lists a cache hit at 10% of the standard input price for most models and a 50% discount on input and output with the Batch API ^[3]. Both apply only to suitable workloads: repeated context, and work that can wait.

7. Budget premium modes deliberately. Opus 5's fast mode runs at around 2.5 times the default speed at twice the base price ^[2].

8. Include non-token costs. Evaluation sets, integration work, security review, human review of outputs and monitoring typically do not appear in a pricing table. Allocate them explicitly.

9. Schedule a quarterly re-forecast. Set a trigger, for example a vendor price change or a volume change of a chosen percentage, that reopens the number.

A simple structure for the spreadsheet

For each use case, estimate monthly requests, average input and output tokens, the model tier, the effort level, the share that is cacheable and the share that is batchable. Then compute low, expected and high cases using the price scenarios above. The spread between low and high is often a more useful planning figure than the midpoint.

Implication for businesses: present finance with a range and a re-forecast date, not a single price-based number. Tie each line to a use case and a tier so cuts or increases can be applied in one place.