What Vendors Said After Failures: GitHub's August Outage Accounts, Anthropic's Claude Code Postmortem and Statements on September 3's AI Outages

Overview

When an AI coding tool fails, what a developer can learn afterwards depends on who ran the service and what kind of failure it was. This analysis compares three clusters of 2026 cases that the Herald selected for the comparison; it is not a survey. They are Anthropic’s April 23 postmortem on Claude Code quality complaints, GitHub’s written accounts of its August outages, and WIRED’s report on what OpenAI, Anthropic and xAI said after outages on September 3. Anthropic and GitHub published dated explanations, while the September 3 statements that WIRED quotes are shorter and one vendor declined to comment. The Herald’s contributor model is built by Anthropic, whose Claude Code postmortem is one of the three cases.

Background: what each kind of source can show

Three outlets independent of the vendors are cited: The Register, InfoWorld and WIRED. DevClass, whose page lists The Register among its sister sites, is cited for one earlier episode and is not counted separately. Their roles differ by case. On GitHub, The Register and InfoWorld mainly relay GitHub’s own status and root-cause text, so on technical cause they add little beyond GitHub’s accounts; what they show is what was public on a given date. On September 3, WIRED asked the vendors directly and is the only cited outlet for that day. For Anthropic’s three changes, the cited sources contain no outside account: the post is Anthropic’s, and the independent DevClass article predates it.

Case 1: Claude Code quality complaints, March to April

DevClass reported on April 14 that Anthropic had reduced the Claude Code prompt-cache time to live from one hour to five minutes for many requests, and that a user, Sean Swanson, had posted a bug report showing the change. According to the same article, Jarred Sumner, who it says now works for Anthropic, replied that the change made Claude Code cheaper, while Swanson wrote in a revised analysis that he had never hit a quota limit until March. The explanations the article quotes came from Sumner and from Claude Code creator Boris Cherny, who said that prompt cache misses with the 1M-token context window are expensive.

On April 23 Anthropic published a postmortem tracing the reports to three changes: a default reasoning-effort change made on March 4 and reverted on April 7, a bug in clearing older thinking from idle sessions shipped on March 26 and fixed on April 10, and a system-prompt instruction to reduce verbosity added on April 16 and reverted on April 20. The post says all three were resolved as of April 20 (v2.1.116), that neither its internal usage nor its evals initially reproduced the issues, that it was resetting usage limits for all subscribers as of April 23, and that it would run per-model evals for every system-prompt change to Claude Code. It also credits users who used the /feedback command or posted reproducible examples online with allowing the problems to be identified. The post does not mention the cache time-to-live change that DevClass described; that is the Herald’s reading of the page. By the Herald’s own count, 50 days passed between the first of the three changes and the postmortem, and nine days between the DevClass article and the postmortem.

Case 2: GitHub’s August outages

The sources let the sequence be laid out by date.

• August 6. The Register reported the same day that GitHub first reported degraded Actions performance at 1522 UTC, that an update some two hours earlier had mentioned an unnamed cause, and that GitHub had not replied to its request for information by publication. The Herald’s August 7 report likewise recorded no detailed root-cause explanation at that point. The fullest account of this incident in the cited sources is in GitHub’s August availability report, dated September 9, which is 34 days later by the Herald’s count.
• August 17. InfoWorld wrote on August 18 that the disruption was first reported at 1:40 PM UTC and that GitHub reported an error rate of about 20% at the height of the incident. GitHub’s status page, it said, promised a “detailed root cause analysis will be shared as soon as it is available”, and until then InfoWorld called the cause unclear.
• August 19. The Register reported that GitHub had published its account, two days after the outage by the Herald’s count: saturated load balancers in the Central US facility, triggered when an Istio sidecar reached its concurrency limit, with an autoscaling policy that monitored the host service but not the sidecar. The Register, quoting GitHub’s account, names a latent retry bug in VS Code that amplified traffic by approximately 10x; neither the executive’s post nor the August report names VS Code. The Herald did not read GitHub’s root-cause analysis page, which the executive’s post links on GitHub’s status site.
• August 20. GitHub’s chief technology officer posted that the outage lasted 7 hours and 47 minutes, that the two incidents it discusses, on August 6 and August 17, were “capacity failures at their core”, and that GitHub was applying consistent retry limits and retry budgets across service-to-service interactions.
• September 9. The availability report gives each of five August incidents the same four headings, including what went wrong and how GitHub is making incidents less likely. For August 17 it reports that 56.07% of front-door requests to the affected services failed or ran slow at the peak and that about 29K organizations saw at least one failed or slow request.

GitHub’s own figures do not agree with each other. The August 17 outage is “7 hours and 47 minutes” in the executive’s post, “lasting 7 hours and 35 minutes” in the report’s heading, and “approximately 6 hours 44 minutes” under the report’s Duration line. By the Herald’s arithmetic, the first two match a 21:15 UTC end with starts of 13:28 UTC (the time The Register gives) and 13:40 UTC (the time InfoWorld gives); the sources do not explain the third. In all five August incidents the report’s heading and its Duration line differ, by 3 minutes on August 26 and up to roughly an hour and 42 minutes on August 6, as the Herald counts them. The posts also differ in how they describe cause. The executive’s post says neither outage “was caused by a code or configuration change”, The Register calls the August 17 autoscaling policy “misconfigured”, and the report lists correcting autoscaling policies as a repair item. The report’s wording on August 6, a routine deployment whose contents “were not at fault”, explains that case but not the August 17 one, so the Herald reports both descriptions without reconciling them.

GitHub’s August report and its October report, dated November 12, also show how GitHub describes failures that involve an outside provider. For August 27, the August report says 63.3% of Kimi K3 requests failed because of a degradation at the upstream model provider, that GitHub posted a public status update pointing to the provider, and that requests using other models or the Auto setting kept working. For October 28, the October report says about one in 10 requests to several GPT-family models in Copilot failed because the credential Copilot uses with an upstream model provider became invalid, and that GitHub traced the errors to its own credential. It says it recommended choosing another model in the meantime, and that code scanning autofix and code quality, which are built on one specific model, were affected more heavily. Neither entry names the provider, and no independent outlet among the cited sources covers either incident.

Case 3: the September 3 outages

WIRED reported that ChatGPT, Claude and Grok all had outages on the morning of Thursday, September 3. An OpenAI spokesperson attributed the problem to “A routing error starting around 7:43 am PT” that made ChatGPT and Codex unavailable for some users, with a solution implemented by about 8:17 am PT. WIRED says Anthropic declined to comment, while its status alerts began at 6:23 am PT with a “partial outage” involving elevated errors on three named models and were marked resolved by 9:16 am PT. SpaceX, xAI’s parent, said the Grok issues came from “an outage at our Memphis compute center”. WIRED adds that OpenAI and Anthropic did not point to a shared cause and that Cloudflare, Amazon Web Services and Microsoft Azure did not report outages that day. None of the cited sources carries a later vendor account of September 3, and WIRED is the only cited outlet for it.

What differs: the Herald’s reading

• Channel and timing. GitHub moved from status updates to a root-cause account two days after the August 17 outage, an executive post on day three and a monthly report 23 days later (Herald’s count). Anthropic moved from staff replies to a user’s analysis, reported on April 14, to an engineering post nine days later. For September 3 the cited sources show same-day statements to a reporter.
• Content. GitHub’s report gives scope in numbers and lists repair items per incident. Anthropic’s post gives dates, versions, process changes and a usage-limit reset. The September 3 statements WIRED quotes give times and a short cause label from OpenAI and SpaceX; the cited sources show no repair items for that day.
• Provider roles. GitHub’s August 27 and October 28 entries separate a provider’s degradation from GitHub’s own credential, in GitHub’s words; the cited sources show no comparable breakdown for the September 3 outages.

Implications for developers

These points follow from the sources and are not predictions.

• Duration and impact figures need their basis. Within one GitHub report, two measures of the same incident differ, so comparing incident lengths across vendors requires knowing which measure each uses.
• For quality problems rather than outages, Anthropic says user-supplied reports and reproducible examples let it identify the causes, so developers’ own records were part of how the causes were found, on Anthropic’s account. The cache time-to-live explanations quoted by DevClass concerned a change that the postmortem does not list.
• A vendor statement does not settle whether separate failures are related. WIRED’s report says the September 3 reasons “remain murky”, and the sources do not say whether the three outages shared a cause.
• GitHub’s entries say that choosing another model avoided the errors in both provider-related cases, and the October entry notes that features built on a single model had no such option, so the available fallback depended on the feature.
• Limits of this comparison: the cases are the Herald’s selection, the Anthropic case rests on Anthropic’s own page, the September 3 case on one outlet, and the VS Code detail on The Register’s reading of a page the Herald did not open.