Three Major Models Recover After Simultaneous Failures
OpenAI, Anthropic, and xAI's official status pages recorded a spike in inference layer errors on the same day, which were subsequently marked as resolved. All three pages currently show system availability, but this does not equate to zero errors for users at any given moment.
OpenAI's historical page recorded the September 3 incident as "Increased error rates for ChatGPT and Codex," with a note stating the issue has been resolved; some Codex remote control users need to re-pair their mobile devices. Additionally, there was a "High error rate in ChatGPT Work mode," which was marked as fully restored by early morning. During the incident, tracking components covered conversations, logins, Work, desktop Codex, compliance interfaces, search, file uploads, voice, GPTs, image generation, deep research, Agent, Atlas, Sites, and connectors. The homepage now states "Fully operational, no issues affecting the system found." The rolling availability is approximately API 99.94%, ChatGPT 99.63%, and Codex and FedRAMP nearly 100%, as aggregated metrics.
Anthropic marked the increased errors across multiple models as resolved, with the impact ending at 9:16 PM Pacific Time, 4:16 AM UTC. The complete list includes Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6; it affected claude.ai, the console, interfaces, Claude Code, and Cowork. The process involved investigation, isolating Mythos/Fable 5.1 and Opus 5, reverting the others to baseline, leaving only Opus 4.8 and Opus 5, deploying fixes and monitoring, before announcing resolution. On the same day, Sonnet 5 also experienced a brief spike in errors which has since been resolved. All components are now operational.
xAI labeled the incident as a "model interruption," with the webpage recording that investigation began around 1:30 PM, and traffic was restored by 5:07 PM, lasting approximately 3 hours and 37 minutes; the duration was similar across iOS, Android, web, X's built-in Grok, and East Coast interfaces. The status page currently shows no ongoing incidents, with Grok clients, Build, office plugins, East and West Coast interfaces, the console, documentation, and the official website all marked as available. Real-time monitoring shows East Coast inference at approximately 99.57%, West Coast cross-region inference at about 99.24% to 99.28%, with non-inference items mostly at 100%. Logins and the European interface were already available in user-side snapshots.
Who is buying and who is selling: The incident was driven by simultaneous overloads in the inference clusters, not by the delisting of any specific product. Buyers are companies that bundle conversation, programming agents, and office plugins into the same inference pool; sellers are those who write error rates into status pages and use mitigations to exchange for recovery windows on the cloud side. Beneficiaries are the interface calls still operational in the European region, while those under pressure are teams that have their workflows tied to a single vendor and a single region. Funds are flowing from subscription limits to retries, model changes, and re-pairing devices.
Source: Public Information
ABAB AI Insight
The simultaneous increase in error rates for the three companies indicates that the bottleneck lies in the shared inference and upstream cloud, not in the chat window text. OpenAI's classification of ChatGPT and Codex under the same incident title acknowledges that consumer conversations and programming agents are competing for the same capacity. Anthropic routed traffic based on model versions, with older Opus versions returning to baseline last. xAI's use of "model interruption" cut off both East and West Coast clients, leaving the European interface as a control, exposing that partition isolation already exists, with default traffic still pressuring the U.S. region.
The capital path is to write intelligence as an online service billed by tokens. Money is invested in training and next-generation models, but runtime relies on a few regional GPU pools. During failures, the entire product matrix turns red because Code, office plugins, and web interfaces were not designed as independently degradable product lines. Once enterprises connect agent workflows to remote control, a single interruption requires re-pairing, with switching costs exceeding those of changing a chat label.
Comparative cases include Anthropic's complete model shutdown in June 2026 due to export controls, and OpenAI's status history, which almost daily reports "increased error rates." The industry position is that the Frontier model has been integrated into office and code agents, expanding faster than multi-active disaster recovery. Control lies in the wording of the status page: investigation, mitigation, monitoring, resolved.
Structural changes involve technological substitution combined with capital concentration. The mechanism is that a few companies sell general inference as a utility; when the utility fails, all downstream "intelligent functions" simultaneously lose power; those who first partition and first model-fuse can write the incident from a full-site outage to a localized error rate.
ABAB News · Cognitive Laws
The more general the model, the wider the workflow covered by a single power outage.
Writing mitigations on the status page does not equate to the client being usable.
The end of the partition that remains operational is the true backup.