Example Co.

Summary

Two workloads are 51% of spend: the support agent (26%) and nightly contract extraction (25%). Both have a fix that doesn't change the model.

1. Monthly spend by provider and model

By provider: Anthropic API $13,814 (52.7%) · OpenAI $6,562 (25.0%) · Google Gemini API $2,813 (10.7%) · AWS Bedrock $1,538 (5.9%) · Cursor $893 (3.4%) · Together AI $616 (2.3%).

2. Effective vs list price, per 1M tokens

Effective price is billed cost divided by tokens. Input is blended across uncached input, cache reads and cache writes.

No account has a negotiated discount or commit. Anthropic spend, including Bedrock, runs at $184K a year, under the $250K threshold for a negotiated process. Comps shows what teams in your band get once you contribute your terms.

3. Three savings opportunities

1. Support agent caching. The system prompt starts with the current timestamp and the customer's plan name, so the 11K-token prefix of instructions and 14 tool definitions changes on every request and rarely matches the cache. Move both fields below the cache breakpoint. Effective input price drops from $1.36 to about $0.50 per 1M tokens.
2. Batch for extraction. The job runs from 1am to 5am and nobody reads its output until morning. The Batch API costs 50% less and guarantees completion within 24 hours rather than 4, so the only tradeoff is that a slow batch could land later in the day. Same model, same output.
3. Haiku for triage. Triage picks one of 23 labels and a priority. claude-haiku-5-5 costs $0.10 / $0.50 per 1M tokens against $2.00 / $10.00. Check label accuracy and escalation recall on 50 labeled tickets before switching.

4. Seat vs API: coding agents

All 5 engineers use Claude Code on a Console API key, billed at API rates (mostly claude-opus-5-5).

Seat prices are Claude Team monthly rates ($100 and $20 a seat billed annually). Anthropic doesn't publish plan limits as token counts, so fit is estimated from active days, peak weeks and session length. Trial seats for A and B for 2 weeks before committing.

Cursor

9 Teams Premium seats ($30 each) include Cursor's own models only. Claude usage inside Cursor bills at API list price, $341 in October, plus Cursor's Token Rate of $0.25 per 1M tokens, another $282. The rate applies to cached tokens too, and 93% of those 1,127M tokens were cache reads, so it costs 2.5 times the $0.10 cache-read price itself. Engineers on Claude Team seats can run that work in Claude Code instead.

5. Routing card

"To test" means the recommendation comes from billing data and published prices only. Step 3, the model and effort map, checks it on your own tasks before you switch.

6. Billing hygiene

7. Models worth testing

8. Next steps

1. This week: move the timestamp and plan name below the cache breakpoint in the support agent. About $3,604 a month.
2. This week: switch nightly extraction to the Batch API. About $3,250 a month.
3. This week: pin the call-summary model version, add a limit to extraction-backfill, move evals to their own key, and remove 3 idle Cursor seats.
4. Before Jan 1, 2027: test gemini-3.1-flash-lite on call summaries.
5. Next 2 weeks: trial Team Premium seats for engineers A and B, keeping the API key as overflow.
6. Optional: a model and effort map (step 3) for triage, extraction and coding agents.

Methodology

List prices come from each provider's public pricing page, checked Oct 31, 2026. Effective price is billed cost divided by tokens, including cache reads and writes, batch discounts and regional surcharges, before credits. Savings assume October volumes and current list prices; they are estimates, not guarantees. Recommendations marked "to test" are not yet checked on your tasks. Plan limits are not published as token counts, so seat fit is an estimate.

Price sources: Anthropic · OpenAI · Google · Together AI · Cursor