Claude broke again on Monday — and Anthropic no longer sells a tier that promises it won't

What actually happened on Monday

Anthropic opened an incident at 05:06 UTC on 24 August reporting elevated errors on requests to Claude Mythos 5, Claude Fable 5, Claude Opus 5 and Claude Opus 4.8. It identified a cause 21 minutes later, reported errors stabilising on Opus 5 and Fable 5 at 07:47, and marked it resolved at 08:30. The closing note gives the honest user-facing window: from 04:50 UTC through 07:36 UTC, users saw elevated errors on requests to Claude models. Roughly two hours and forty-five minutes.

The affected components were the Claude API, Claude Code, Claude Cowork and claude.ai — which is to say all four surfaces through which anyone reaches the models. Anthropic did not publish a root cause, which is normal for a status-page entry and not a criticism.

A second, three-minute login incident followed at 17:12 UTC the same day.

The number that looks damning, and isn’t

Pull Anthropic’s incident feed for August and the shape is unflattering. Twenty-one incidents between the 1st and the 24th, across twelve of those twenty-four days. Cumulative open-to-resolved time of about 33.3 hours, of which roughly 9.2 hours carried a major or critical label. One incident on 5 August ran 429 minutes at “degraded performance.” A critical on 16 August took out the API, Claude Code and the Console for 36 minutes. On 14 August, status.claude.com itself was unreachable for about 175 minutes — the fallback signal failed alongside the thing it was meant to report on.

It reads like a system under strain. But a number without a comparison is a rhetorical device, not evidence, so here is the comparison.

Over the identical 1-21 August window — OpenAI’s public feed runs to the 21st, so that is the fair boundary — OpenAI logged 17 incidents across 8 days totalling about 62.7 hours. Anthropic logged 19 across 11 days totalling about 29.8 hours. OpenAI accumulated more than twice the raw incident-hours.

So the “Anthropic is falling over” framing collapses on contact with the peer data. Frontier inference at scale is unreliable everywhere, and anyone selling you a story about one lab’s uniquely bad month is not counting the other lab.

Two real differences survive the comparison, though, and both matter:

• Severity. Every OpenAI incident in that window is labelled minor or unrated. Anthropic recorded five major and one critical. Anthropic’s outages are shorter and sharper.
• Blast radius. A large share of OpenAI’s entries are confined to ChatGPT surfaces — image generation, custom GPT actions, ads manager login, Sites deploys. Anthropic’s entries overwhelmingly name the Claude API alongside Claude Code and Cowork. If you are a developer rather than a chat user, Anthropic’s month was the more disruptive one even though it was the shorter one.

The part that is actually new

Here is the finding that reorganises this from a news cycle into a procurement question.

Anthropic’s service tiers documentation now opens with a warning: Priority Tier capacity commitments are no longer available for purchase. Organizations with an existing commitment can run it to contract end; everyone else is directed to contact sales.

Priority Tier was the only Anthropic service tier with a published availability number. The documentation still states it plainly: Priority Tier “targets 99.5% uptime with prioritized computational resources.” The other two tiers make no such promise. Standard — the default for every API request — is documented as prioritising requests “alongside all other requests with best-effort availability.” Batch is asynchronous by design.

Note the word “targets,” which was never the same thing as a warranty. But it was a number, in public, that a buyer could point at. That number is now closed to new customers.

Then read the supported-models line, which is where this stops being abstract:

Priority Tier is supported on all available Claude models except Claude Mythos 5, Claude Mythos Preview, Claude Opus 5, and Claude Sonnet 5.

Those exclusions are not obscure legacy SKUs. Mythos 5 is the restricted frontier tier. Opus 5 at $5/$25 is what Anthropic positions for serious agentic work. Sonnet 5 at $2/$10 is the default on Free and Pro and the everyday workhorse for most teams.

So even a customer still holding a Priority Tier commitment has no uptime target on the three models they are most likely to be running. And Monday’s incident hit Mythos 5, Opus 5 and Opus 4.8 — two of those three sitting squarely in the excluded set.

The fair comparison on commitments

This is where the contrast with OpenAI genuinely does cut against Anthropic, and it is worth being precise about why.

OpenAI’s Scale Tier is still on sale, and still publishes a 99.9% uptime SLA with prioritised compute. Whether an SLA is worth much is a separate argument — service credits are a refund of a fraction of your spend, never compensation for your outage, they are typically capped at a low percentage of monthly fees, and they generally require you to file a claim with your own monitoring evidence inside a window. Pay-as-you-go and consumer tiers are excluded. Latency and throughput are excluded. An SLA is a pricing artefact more than a reliability mechanism.

But it is a published artefact. Anthropic’s equivalent is now “contact sales,” and reporting on enterprise agreements suggests negotiated terms in the 99.99% range with credits capped around 5-10% of monthly fees — figures that, being negotiated case by case, no buyer can verify against a peer or hold up in a procurement review.

The asymmetry to take away is not “OpenAI is more reliable.” On the month’s data it wasn’t. It is that one vendor lets you read the commitment before you buy and the other requires a sales conversation to discover it — and that Anthropic moved in the direction of less disclosure at exactly the moment its status page got busy. Those two facts are probably related in the boring way: capacity is scarce, commitments are expensive to honour, and withdrawing the product is cheaper than missing the target.

What to do about it

Assume the endpoint fails, and design for it. This is the whole answer, and it is architectural rather than contractual. A single vendor endpoint is not an availability primitive. Route through a gateway or your own fallback so that a model-family outage degrades output quality instead of halting work — which is precisely the capability that made the routing layer worth $7.5B to Stripe and worth giving away free to Ramp three days before this outage. If Monday cost you three hours of agent throughput, that was a fallback-routing decision you had already made by not making it.

Keep a second family qualified, not just configured. Fallback that has never been tested against your evals is a config file, not a plan. The practical version is a periodic run of your evaluation set against a second family so you know what you actually lose when you fail over — and the honest answer is usually “less than you fear.”

Measure your own error rates. Do not outsource this to a status page. “Degraded performance” spans everything from a 2% error rate to unusable, the label is applied by the vendor at the vendor’s discretion, and on 14 August Anthropic’s status page was itself down for about three hours. Your own client-side error and latency telemetry is the only signal that is available when you most need it.

If reliability is contractual for you, ask before you sign. Get the uptime figure, the measurement methodology, the excluded conditions and the credit mechanism in writing. With Priority Tier withdrawn there is no published number to anchor on, which weakens your position and makes asking early more important, not less. Ask specifically whether the commitment covers the model you intend to run — given the exclusion list, that is not a formality.

Do not restructure your stack over this. Anthropic’s models did not get worse on Monday, the peer comparison does not show an outlier, and switching vendors over a month of status-page entries is how teams acquire migration costs in exchange for a different set of incidents. The multi-harness, multi-model posture that vendors have been building toward all year is the correct response, and it is a response to the industry condition rather than to any one lab.

The verdict

The outage is not the story; outages are the weather. The story is that Anthropic quietly removed the only tier where it put an availability number in public, while shipping three flagship models that tier never covered — and that this became visible during the busiest month its status page has had.

For buyers the practical consequence is narrow and clear. You cannot purchase reliability from Anthropic off a price list any more, so stop trying to. Price the models on capability and cost, where they remain genuinely competitive, and buy your availability architecturally instead: fallback routing, a qualified second family, and your own telemetry. That was already the right design before Monday. Monday just removed the excuse for not having built it.

Questions or corrections? Email Pick Right. Want the full list? See all news.