OpenAI reports 1,700 ChatGPT investigations as automated repairs lag diagnosis
Editorial Team
••📖6 min read
OpenAI says its internal Performance Factory tracked more than 1,700 investigations in just over three months. About 400 cases reached proposed fixes, but far fewer shipped repairs used Codex’s code unchanged, and engineers retain approval.
OpenAI says its Performance Factory tracked more than 1,700 investigations and identified causes in over 600 cases in just over three months.
About 400 cases reached proposed fixes; OpenAI’s team says around 15 shipped fixes used Codex’s exact pull requests.
Engineers still approve deployments, and the published results do not quantify overall speed gains or operating savings.
OpenAI, the company behind the ChatGPT conversational AI service and Codex software engineering agent, is using its own coding technology to investigate performance problems and propose repairs. Its internal Performance Factory coordinates work from an alert through a post-deployment check, but reliable code changes remain harder to automate than diagnosis.
Company-reported investigation and repair counts over just over three months. The shipped-fix categories may overlap and must not be added. Chart: Magica, using OpenAI’s figures reported by VentureBeat. Source: VentureBeat.
More investigations than shipped repairs
In its November 19 engineering account, OpenAI says the factory’s first application, the Regression Workflow, tracked more than 1,700 investigations and identified root causes in over 600 cases in just over three months. The company reports an average of 17 automated investigations a day. These figures measure investigative activity, rather than how many slowdowns were repaired or how much faster ChatGPT became.
The company says performance engineering needs to keep up with faster code development using agents. Its response goes beyond asking a coding assistant to explain a slowdown: the factory keeps a record of the problem, assigns successive steps and resumes work after reviews and deployments.
OpenAI’s performance team supplied a more revealing breakdown in comments accompanying the implementation report. Of the more than 600 investigations with an identified cause, around 400 reached a proposed fix. Around 50 shipped fixes reflected or matched agent suggestions, while around 15 shipped fixes used the exact pull requests authored by Codex.
Epic says it is using AI to rewrite parts of Unreal Engine and plans to let developers connect their preferred models to UE6. The strategy combines claimed maintenance gains with a broader overhaul of programming, distribution and portable game assets, while rival Unity already offers outside-model connections.
Editorial Team
Those last two categories should not be added together: the account does not establish whether they overlap. Nor do the approximate counts establish a final success rate for every investigation. They do show why investigation volume alone is an incomplete measure of automated repair.
Codex had the coding tools; the factory adds coordination
Codex was introduced in a May 16, 2025 research preview as a cloud-based software engineering agent able to write features, fix bugs and propose code changes for review. Parallel tasks, repository access and human validation were already part of that design. Performance Factory applies those capabilities to recurring operational work with records and handoffs that persist beyond one agent session.
OpenAI’s performance engineers had previously built more than 30 tools, including CPU and allocation profilers. They first guided Codex through investigations, then recorded their procedures in a reusable skill called perf-forensics. Linking alerts to cloud sessions automated diagnosis, but engineers still had to follow up, implement repairs and check recovery.
The factory has four components: a task service with configurable workflows, a dispatch loop assigning ready work, an internal API running Codex in the cloud, and event subscriptions that resume paused tasks. Engineers specify the task states, allowed transitions and points requiring human involvement.
Agents work inside OpenAI’s network with access to repositories, metrics, logs, traces, CPU profiles and memory snapshots. A service identity lets the company limit permissions and trace actions. The deployment therefore draws on accumulated engineering tools and internal production evidence, as well as the coding agent.
OpenAI’s team told the implementation report that agents save findings, metrics and reproducible queries in task records instead of passing along large volumes of raw profiling data. A later agent can rerun a query to check the evidence. Subscriptions let tasks wait for a code review, failed check or deployment without keeping an agent running throughout the wait.
OpenAI’s Performance Factory connects task records, cloud Codex sessions and event subscriptions that resume waiting work. Diagram: OpenAI. Source: OpenAI.
Finding a plausible cause is not enough
A slowdown can have different explanations. OpenAI says its investigator checks whether comparable requests became slower or whether the service simply received more requests of an already slow kind. It then examines deployments, configuration changes and code against the timing of the affected traffic.
Even so, early agents sometimes settled on the first plausible explanation. OpenAI added a fresh Codex session to challenge diagnoses and test alternatives. In one investigation, a reviewer reopened a case because the service blamed for image-download failures had recovered while the failures continued.
In a manual review of 100 root-cause suggestions, OpenAI graded 70% as accurate; the rest needed refinement or had not fully established causality. This is a company assessment of that sample, not a demonstrated accuracy rate for all production investigations.
OpenAI’s team told the implementation report that adversarial review currently returns around 25% of diagnoses for further investigation. A separate summary describes 25% as the share receiving review. The measures differ; the team’s direct comments support the return-for-investigation figure.
Repairs introduce another difficulty. The team says Codex works best when the cause is clear and the remedy contained—for example, replacing repeated serialization of the same conversation with a single serialization whose result can be reused. Changes requiring architectural decisions are harder. A cache might reduce latency while introducing stale data.
OpenAI says proposed changes can be unnecessarily complex or introduce new problems, so fixes also receive adversarial review. An engineer then reviews the pull request and decides whether to merge it. Engineers can edit the proposal or write their own repair using the diagnosis.
After deployment, an agent checks performance against its earlier baseline. No improvement reopens the investigation. Partial recovery resolves the original task but creates another for the remaining slowdown, so a task marked resolved does not necessarily mean all of the performance loss has disappeared.
OpenAI’s example regression report combines company-reported startup measurements, a root-cause explanation and a draft fix for engineer review. Screenshot: OpenAI. Source: OpenAI.
What would justify more deployment authority
OpenAI says agents can generate more findings than engineers have time to read. The workflow now groups related incidents, ranks estimated impact and supplies concise reports with charts. The team also told the implementation report that automated investigations had freed time for longer-term performance work, though it supplied no quantified time saving.
OpenAI’s team says it has completed a workflow evaluation suite to improve agent-generated fixes. The company also backtests investigations against past incidents and assesses summaries separately. Historical testing has limits: metrics and traces can expire, while evidence from after recovery can reveal the answer. OpenAI restricts access to saved notes, incident writeups and previous test results, and checks transcripts for such leakage. A daily agent also looks for broken tools and missing permissions that can block investigations.
OpenAI is working on synthetic performance tests so agents can measure and revise proposed fixes before seeking review. It also wants to detect slowdowns that do not trigger alerts and run recurring optimizations, such as finding autoscaling settings that keep unnecessary service replicas running. Shipping some low-risk fixes without human approval remains an exploration.
The next decision is how much deployment authority to delegate. The published accounts do not quantify aggregate latency improvements, compute savings, agent operating costs or engineering hours saved against a comparable manual baseline. Evidence of reliable repairs and net savings would help establish whether the factory’s greater investigative capacity warrants that next step.
At Ignite, Nvidia and Microsoft are expanding local coding AI, cloud desktop streaming and Azure infrastructure. The announcements build on an existing Windows AI strategy, with AMD alternatives, staggered availability and no comparable deployment-cost results in the announcements.
Microsoft and Nvidia are expanding Azure’s Vera Rubin virtual-machine preview, with systems already running in its data centers. Microsoft’s separate confidential-AI showcase covers four GPUs, while the announcements leave pricing, regions and general availability unspecified.
Meta is reconsidering its 2030 climate deadline after net emissions reached 11.8 million tonnes in 2025. Hardware and infrastructure purchases dominate its reported footprint, while new gas generation adds a longer-term challenge.
Three House committee chairs want the Remote Access Security Act in the annual defense bill as scrutiny of Nvidia intensifies. The available Senate text would authorize controls on specified foreign users’ remote access, with computing limits left to Commerce.
Google Quantum AI reports tests of quantum contextuality spanning up to 105 qubits on Willow processors. The experiments expose hardware errors and beat restricted classical strategies, while the largest task relies on an unproven classical benchmark.
A bipartisan House disclosure shows requests to sell 3.4 million AI chips to China and approvals for more than 879,000. The oversight question is whether Commerce verified the buyers, supply protections and access restrictions behind those authorizations.
Activists protested Anthropic’s stock-market debut outside Nasdaq, with at least one arrest and conflicting accounts of the total. Their campaign targets retirement-fund exposure to AI shares, alongside concerns about jobs, data centers and public accountability.
Anthropic says cash, growing revenue and multiyear payment schedules support its vast compute commitments. Its filing shows why flexibility matters: short-term capacity costs more, while major cloud contracts require payment regardless of usage.
OpenAI is rolling out live transcription for dictated ChatGPT prompts across free and paid plans. Users can see words as they speak and review the text before sending, but the company has not supplied performance measurements or a replacement recording maximum.
Boltz has raised a Bessemer-led $110 million Series A to expand molecular-design tools sold by usage. Its open-source research brought broad adoption; the commercial test is whether paid computation delivers repeatable gains in laboratory discovery.
Donald Trump said he and Xi Jinping supported minimal AI guardrails but had not discussed definitions. Their Shenzhen meeting leaves unresolved how bilateral safety cooperation would coexist with U.S. opposition to global regulation.
Gemini can now click and type across Mac apps and websites in a background window, Google says. The paid rollout expands its desktop automation, but app permissions expose active accounts and file changes that backups cannot always undo.
The House AI Safety and Innovation Caucus promises legislative advocacy and rapid incident briefings. Its members already have bipartisan safety proposals in play, making the choice of safeguards its next test.
Binance Wallet says it will return all subscription commitments, distribute roughly 2,957 Anthropic tokenized shares free and add a $1 million token giveaway after demand far exceeded its allocation.
Early Anthropic investor Dustin Moskovitz is urging frontier AI developers to slow down. He and Cari Tuna moved their stake to a nonprofit in early 2025, directing potential returns toward philanthropy while leaving questions about the holding's value and governance rights.
A national study finds higher odds of hosting data centers in majority-Black, Hispanic and Asian neighborhoods. Facility types, zoning rules and unequal access to public hearings complicate who bears the costs and receives the benefits.
A debut-day estimate values the Anthropic shares FTX sold in 2024 at $49.7 billion. The comparison exposes the scale of subsequent paper gains, but contrasts cash raised for bankruptcy repayments with a hypothetical stock-market valuation.
Anthropic raised $81.3 billion in a record IPO and opened 34% above its offer price before retreating. Rapid revenue growth and adjusted operating profits sit alongside vast computing commitments and limited investor control.
The Kimi developer reportedly abandoned planned Nvidia B300 capacity for 2027 after new US restrictions. The deal exposes financing risks for overseas AI clouds, but neither a model delay nor a $150 million cash loss is established.