Model spend behaves differently from cloud spend. It scales with tokens, not machines, and the biggest levers are in how requests are written. This checklist is built from what Anthropic, OpenAI and Google publish on their pricing and caching pages, read on 18 October 2026.
Visibility
- Log the usage block of every call. Anthropic returns cache creation and cache read counts. OpenAI returns
cached_tokens. - Attribute spend. Use separate keys or projects per team and per product, so a bill can be explained.
- Track cache hit rate. OpenAI's guide suggests cached tokens divided by total input tokens. A rate near zero on a long prompt is a bug.
- Track tokens per task, not just per call. Agents make many calls per task.
Control
- Set a budget and an alert per key. Alert at a fraction, not at the limit.
- Set maximum output tokens and maximum turns. Runaway loops are the usual surprise.
- Cap context. Start a fresh session instead of carrying a long one.
- Pick the smallest model that passes your test. Published input prices range from $0.10 to $10 per million tokens on Anthropic's page alone.
Unit price levers
- Cache stable prefixes. Anthropic lists reads at 0.1x input, and 0.05x or 0.025x on some newer models. See what caching saves.
- Batch work nobody waits for. Anthropic, OpenAI and Google each list it at about half price. See batch discounts compared.
- Watch price brackets. Haiku 5.5 on Anthropic's page has a higher price over 100,000 tokens, Gemini 3.1 Pro Preview over 200k, and OpenAI lists long context rates over 272K tokens.
Surcharges and change
- Know the surcharges. Anthropic lists a 1.1x multiplier for US only inference on Claude 4.6 and later. OpenAI lists a 10 percent uplift for regional and FedRAMP endpoints. Priority or Fast modes cost more: Google lists Priority at roughly 1.8x standard for most models.
- Calendar the price changes. Google's page lists higher prices for Gemini 4 Argon from 1 December 2026 and for Gemini 3.8 Flash and 3.6 Flash from 1 January 2027. Anthropic's page notes Sonnet 5 introductory pricing became the standard price. Recheck pages monthly.
- Compare plan, API and enterprise yearly. See our decision guide.
- Price the self hosting option honestly. Include idle time and staff. See the GPU estimate.
Review rhythm
Weekly: top ten spenders by key and cache hit rate. Monthly: model choices against price pages. Quarterly: plan against API against self hosting, using your logged tokens.
Connecting Safeguard to an agent over MCP does not add tool call charges because MCP tool calls are unlimited, and you can bring your own model, so the line items above are about the model provider.
Quick check
Why track cache hit rate?
A low rate on long prompts means you pay full input price for text that could have been read at a fraction of it.
Sources
- Claude API pricing, fetched 18 October 2026
- OpenAI prompt caching guide, fetched 18 October 2026
- OpenAI API pricing, fetched 18 October 2026
- Gemini API pricing, fetched 18 October 2026
