Checked 13 October 2026, 12:10 UTC · Public-source review

OpenAI reports a GPT-6.1 Sol Codex/Work performance issue. Broad model decline is not established.

OpenAI lists degraded performance in GPT-6.1 Sol Codex and Work mode as under monitoring after mitigation. Several separate user reports describe capacity errors, narrow request failures, or context metadata changes. These are not one finding, and none establishes a broad decline across popular models.

For a failing Codex task, check status first. Save a local record of the task, expected result, model and effort, client, context, visible route, error or output, correction, final artifact, and allowance. Repeat the same task after the provider marks recovery. Treat any fallback as a separate comparison.

Confirmed service event

The OpenAI status incident names degraded performance in GPT-6.1 Sol Codex and Work mode. At the latest provider update, posted at 09:30 AM on 13 October, OpenAI said mitigation had been applied and recovery was being monitored. Affected components listed include ChatGPT Work and Codex Web, API, CLI, and VS Code. The status page says availability is aggregated across tiers, models, and error types, so an individual account may differ. This confirms a service-performance issue on named surfaces; it does not identify root cause or measure model capability.

Three different user-reported leads

Quality complaint plus capacity errors

A Codex issue opened 10 October describes a Windows Pro account using GPT-6 Astra and GPT-6.1 Sol at max effort. The author alleges worse quality since mid-September in Codex/Work/Dots, says ordinary Chat on the same account remains good, and reports 7 server_overloaded errors among 35 local task_complete events in two threads. The error counts and quality claim are user-provided; logs and matched outputs are not public.

Specific replay can fail before output

A second Codex issue reports small direct-HTTPS tests in which GPT-6.1 Sol intermittently fails on particular saved encrypted reasoning items before producing output, while removing an item or using GPT-6 Sol often completes. This is a narrow, user-supplied reproduction; backend cause and population rate are unknown.

Context metadata changed

A third issue reports the effective context metadata changing from 475,000 to 258,400 between turns while GPT-6.1 Sol and high effort remained selected. It also reports originator-dependent metadata for one account. The report does not measure answer quality or link the metadata issue to today’s status incident.

Task evidence points in mixed directions

A developer reports one 24-task comparison between GPT-6.1 Sol and GPT-6 Sol on a single private Go/TypeScript repository: 9 wins, 10 ties, 5 losses for Sol 6.1 by the authors rubric, with one attempt per task and model. It is a cross-sectional comparison, not a before-and-after test. One repo and one attempt per task cannot establish general performance or a temporal decline.

Two Codex Reddit posts show complaints and a positive reliability report. Both are self-selected. They have no common tasks, denominator, or representative sample.

What the evidence does not show

No public matched task series shows a broad before-and-after capability decline across OpenAI, Anthropic, Google, xAI, or popular models overall.
The status page confirms an incident, not a model-weight or capability change.
Request errors are failures to obtain an answer; do not score them as incorrect completions.
The context metadata report is a runtime/context lead, not proof that completed answers got worse.
Reports from one account, one project, or an online discussion do not establish prevalence.

What to do next

Check the provider status and note whether the affected app or client is listed.
Record the same task, expected result, chosen model and effort, client version, context, tools, visible route, output or error, corrections, final artifact, and allowance locally. Do not share private prompts or logs unless you choose to redact them.
After the incident is marked resolved, rerun that task on the same surface and configuration. Compare output and correction burden.
If you use another model or surface, record it as a separate test. Do not treat it as the same system.

Measurement project to watch

The model-drift public repository describes a frozen, fingerprinted 35-task API suite with deterministic graders and repeated probes. That is a proposed API monitoring design; this review did not run it or verify published results, consumer product coverage, adoption, or paid use.

Sources

Back to current verdict · Read how we check