The short answer
Per-token price tells you very little. Check these in order: eligibility, rate limits, SLA, real cost, contract and invoicing, data privacy. Then load-test with production-shaped traffic and get the terms in writing.
1. Eligibility comes before price
Being able to call the API does not mean your company qualifies. Anthropic's Supported Regions Policy takes effect November 12, 2026. It covers the API and Claude sold through cloud platforms and authorized resellers, and lists three unsupported situations:
- Personnel located in an unlisted region.
- An entity incorporated or headquartered in an unlisted region.
- An entity majority-owned or controlled by individuals or entities from those regions.
Cases 2 and 3 apply even if your people sit in a supported region. Anthropic calls the new page a clarification of existing enforcement. I could not verify the older wording for anything except the majority-ownership rule, so read both versions.
If you buy through a third party, ask in writing whether its upstream channel is authorized. A different entry point does not make an unsupported entity eligible.
2. Rate limits: not every 429 deserves a retry
Anthropic limits each model by RPM (requests per minute), ITPM (input tokens per minute) and OTPM (output tokens per minute). Exceeding any one usually returns a 429. The limiter is a token bucket that refills continuously, so a short burst can trip it.
| Situation | What you see | What to do |
|---|---|---|
| RPM / ITPM / OTPM limit | 429 with retry-after | Back off and retry |
| Account monthly spend cap hit | 429, usually no retry-after; error.details.error_code is enforced_spend_limit_reached | Retrying is pointless. Wait for the 1st of next month (UTC) or request an increase |
| Org/Workspace spend limit you set | Usually 400 (invalid_request_error) | Raise or remove the limit |
Concurrency is not a number you can read off a plan. In-flight requests ≈ arrival rate × average duration. 10 requests/s at 8 s each is about 80 in flight (arithmetic, not measured). For most Claude models, cached input tokens read from the cache do not count toward ITPM, so prompt caching raises effective throughput.
3. SLA: only what is in the contract counts
Standard tier is best-effort. Priority Tier is open only to organizations that already have a capacity commitment, and it is no longer sold to new customers. A new buyer's SLA therefore mostly comes from a cloud platform or vendor contract.
Ask for:
- What is measured. Gateway success rate, end-to-end availability and perceived latency are three different numbers.
- The time window, exclusions, notification deadline and compensation formula.
- Whether failover to a backup channel for the same model can double-bill or cut off a stream.
4. Cost, contract and privacy
- Cost: monthly cost ≈ input + output + cache + platform and payment fees + ops overhead. Sample production logs, add retries and growth, and compare vendors on the same exchange rate and fee basis.
- Contract: the contracting company, payee and invoicing entity should match. Also check invoice type, tax, payment terms, refunds, outage compensation and data deletion after termination.
- Privacy: Anthropic's commercial API inputs and outputs are generally deleted within 30 days. Exceptions include services where you control retention, separate agreements such as zero data retention, policy enforcement and legal requirements. Content flagged as a violation can be kept up to 2 years.
For every hop (gateway, upstream channel), ask what it sees, how long it keeps it and where it is processed. Ask for the DPA and the sub-processor list. A gateway that says it does not persist prompts has not told you that the upstream retains nothing.
5. Pre-signing checklist
Trial with a limited budget and non-sensitive data. Have engineering, finance and legal sign off.
- [ ] Entity, ownership, user locations, upstream authorization statement
- [ ] Model IDs, Messages protocol, streaming, tool calls, caching
- [ ] Peak and burst test; record concurrency, RPM, ITPM, OTPM, P95/P99
- [ ] Count 429s, 5xxs and dropped streams; verify backoff, failover, idempotency
- [ ] Compare client logs to the invoice (tokens, cache, overage)
- [ ] SLA exceptions, compensation, invoicing entity, DPA, log retention
6. One vendor's published figures (affiliated)
Per LMU AI, since August 20, 2026 its China site (api.lmuai.com) serves only China-origin models, and Claude is on its international site for overseas users. Its enterprise docs say:
| Item | Vendor-stated | Caveat |
|---|---|---|
| Concurrency | "from 40" (standard) and "from 80" (flagship) | Plan labels, not measured throughput |
| Availability | 99.5% target, monthly report with success rate and P95 latency | Self-computed. Counts only its own request success; excludes user network issues, bad parameters, upstream refusals and outages under 60 s. Compensation is negotiated separately. 99.5% over 30 days allows about 3.6 hours of downtime (720 h × 0.5%, arithmetic) |
| Invoicing | Purchase contracts and VAT invoices | Confirm entity and taxes in writing |
| Data | Content not persisted; metadata and error logs kept 90 days; account data stored in mainland China | Error logs may contain truncated request fragments. Content still reaches the model provider through an overseas channel. The DPA goes through sales, and an unsigned DPA is no safeguard |
Sources: enterprise plans, reliability, data and privacy, service scope FAQ (vendor statements, not third-party monitoring).
Teams whose eligibility is confirmed can look at the LMU AI international site (our tracking link). We have not tested these figures with a real key. Hold any vendor, including us, to the same checklist.
What did your legal or finance team ask that is not on this list? And if you bought Claude through a cloud platform, what did the SLA end up looking like?
