Does OpenAI or Anthropic owe you an SLA credit when the API goes down?

Your LLM provider is now a line item the size of a database cluster, and it goes down like one. After a bad afternoon, an ops or FinOps lead asks the same question they'd ask of AWS: is any of that money coming back?

The short answer

Almost certainly not — and for two different reasons. OpenAI does state a 99.9% uptime SLA, but only on the Scale Tier and Fast mode product pages, and only for Enterprise customers; it publishes no uptime definition, no measurement period, no credit percentages and no claim window, so there is no public schedule from which a credit can be computed. Anthropic publishes no availability SLA at all — its Commercial Terms of Service state no uptime figure, define no downtime and establish no service credit, and expressly disclaim that use of the Services will be uninterrupted. If you are on pay-as-you-go with either vendor, you are owed nothing for the outage. If you are on an OpenAI enterprise agreement, the remedy exists but its terms are in your contract, not on the web.

That is a sharp break from the infrastructure these services now compete with for budget. AWS, Google Cloud, Azure, Cloudflare and DigitalOcean all publish the whole tariff — percentage, formula, tiers, deadline.

Does OpenAI publish an uptime SLA?

Partly, and the distinction is the whole article. OpenAI's Scale Tier page says Scale Tier traffic "offers a 99.9% uptime SLA and prioritized compute", and lists a 99.9% Uptime SLA plus a per-model latency SLA (p50 request latency, measured per five minutes) for every model in the table. The Fast mode page carries the same two columns — with a footnote reading, in full, that they are "applicable to Enterprise customers only". Scale Tier itself is described as available to Enterprise customers.

What is *not* published is everything you would need to file:

The one thing OpenAI does say about the remedy is how the two SLAs interact. Its Scale Tier FAQ states that if both the latency and uptime SLA are violated, you are credited with the greater of the two amounts for that calendar month — and the Fast mode FAQ adds that Fast mode SLAs are treated the same as Scale Tier SLAs, with service credits offered if OpenAI fails to meet them "for customers on enterprise agreements". Greater of two amounts that are never stated.

Two smaller details are worth carrying into a contract review. Fast mode traffic downgraded to Standard is, in OpenAI's words, "not eligible for Priority processing Service Level Objectives" — objectives, in a document that elsewhere says SLA. And the Reserved Tier FAQ says the SLA "for the service tier you use for your requests, such as Fast mode or Standard, will apply", while no uptime SLA is published for Standard anywhere. We reproduce both as printed rather than resolving them; if you hold a Reserved Tier order form, put the question to your account team in writing.

Careful: "Service Credits" in OpenAI's terms are not an outage remedy

This is the most common way a false claim starts. Search OpenAI's Services Agreement for "service credit" and you land on a clause reading that a customer "may need to prepay for Services by purchasing Service Credits". Those are prepaid purchase credits — money you buy in order to spend it on the API. They are not compensation for downtime, and that section is not a claim route.

The same agreement warrants conformity, not availability: OpenAI warrants the Services "will conform in all material respects with the Documentation", and subject to that, provides them "AS IS". An outage is not on its face a failure to conform with documentation, so a status-page screenshot is not evidence for that warranty. There is also a thirty-day channel for disputing an invoice — a billing mechanism on its own clock that bars nothing about availability. Do not file an outage claim against it.

Does Anthropic pay SLA credits?

No, and its terms are more explicit about it than mere silence would be. Anthropic's Commercial Terms of Service (effective 17 June 2025) contain no occurrence of "service level", no uptime percentage, no definition of availability and no service credit; anthropic.com/legal/sla returns a 404. The warranty disclaimer provides the Services "AS IS" and "AS AVAILABLE", and states that Anthropic "does not warrant, and disclaims that, the Services or Outputs are accurate, complete or error-free or that their use will be uninterrupted".

Anthropic's platform documentation describes the default Standard tier as having "best-effort availability". Its Priority Tier is a capacity commitment — prioritisation during peak load, not availability — and is no longer sold; existing commitments run to their contract end date.

So a heavy incident record on Anthropic's status feed is not a remedy. It is a reliability signal for a renewal negotiation, and that is all. The same is true of OpenAI's.

How to find out what you are actually owed for an AI API outage

1. Establish which tier served the traffic. On OpenAI, Standard is the default; Fast mode and Scale Tier are opt-in and Enterprise-gated. On Anthropic, check the service_tier value returned in the response usage object.
2. If you are pay-as-you-go on either vendor, stop here — there is no published remedy, and filing anyway spends goodwill you may want for the renewal instead.
3. If you are on an enterprise agreement, open the order form and the MSA, not the marketing page. Search for "Service Level", "Uptime" and "Service Credit". The percentages, the measurement method and the deadline live there.
4. Confirm how downtime is measured before you calculate anything — vendor-side and customer-side clocks routinely disagree. See how SLA uptime is measured.
5. Keep your own evidence regardless: request-level error rates and timestamps matched against the vendor's incident timeline. A status page proves an event happened, not that your traffic was affected for a given number of minutes.
6. Diary the deadline the day the outage ends — missed windows are the most common reason claims die, and an unpublished window is still a window.
7. Check what the rest of your stack owes you for the same day: which SaaS vendors actually pay SLA credits.

Frequently asked questions

Does OpenAI have an SLA?

For Enterprise customers on Scale Tier or Fast mode, yes — OpenAI's product pages state a 99.9% uptime SLA and a per-model latency SLA, and the Fast mode footnote limits both to Enterprise customers. For pay-as-you-go Standard traffic, no uptime SLA is published. OpenAI publishes no SLA document containing a downtime definition, credit schedule or claim window.

Am I owed a credit if the OpenAI API was down for two hours?

Only if you hold an enterprise agreement that grants one, and then only on the terms in that agreement. There is no public credit schedule to compute against, so any figure quoted to you from a public source is inferred rather than read. On pay-as-you-go, nothing is owed.

Does Anthropic pay SLA credits for Claude API downtime?

No. Anthropic's Commercial Terms of Service publish no availability commitment and no service credit, and disclaim that use of the Services will be uninterrupted. If you believe you hold an availability commitment, it can only be in a separately negotiated agreement.

Is OpenAI's "Service Credit" the same as an SLA credit?

No — and conflating them is the trap. The Service Credits named in OpenAI's Services Agreement are prepaid credits a customer buys to spend on the API. An SLA credit is compensation for a breached uptime commitment. See SLA credit vs service credit for the general distinction.

Which AI-adjacent vendors do publish a claimable SLA?

The cloud platforms underneath. AWS, Google Cloud, Azure and Cloudflare publish commitments, tier tables and deadlines you can compute a number from — the AWS SLA page includes a working credit calculator. Browse the full set in the vendor directory.

Methodology & caveats

The OpenAI figures — the 99.9% uptime SLA, the per-model latency SLA and its p50/5-minute basis, the Enterprise-only footnote, the greater-of-two-amounts rule and the Reserved Tier answer — are transcribed from OpenAI's Scale Tier, Fast mode and Reserved Tier pages, read 22 September 2026. The Services Agreement clauses (prepaid Service Credits, the conformity warranty, the "AS IS" provision, the thirty-day invoice-dispute channel) were read from that agreement on the same date. The Anthropic findings are from its Commercial Terms of Service (effective 17 June 2025) and its published platform documentation, read 22 September 2026; anthropic.com/legal/sla returned 404.

One inconsistency to name rather than hide: Ontracko's OpenAI vendor profile records "no published uptime SLA", and that refers to OpenAI's contract documents, which contain no availability term. The 99.9% figures above sit on product and pricing pages. We do not score or calculate against them, and neither should you — a percentage without a downtime definition, a measurement period, a credit schedule and a deadline is a marketing claim, not a filable term. The Anthropic profile shows no figure at all, for the same reason. Vendors revise these pages; re-read the source and your own order form before acting. Terms are defined in the SLA glossary, and the incident records behind our scoring are at reliability rankings.

*Your AI vendors may owe you nothing — the rest of your stack probably does. Ontracko watches every vendor's status feed, timestamps the incidents, computes the credit where a published tariff exists, and reminds you before the window closes. Monitor your vendors free — 8% only on recovered credits. See which vendors pay or browse the vendor directory.*