OpenAI says Rust now handles 95% of production requests to Habitat, the online storage layer behind ChatGPT, Codex, the API, and internal services. The migration is an unusually large real-world change in software architecture: a two-engineer team rewrote a Python service that had reached more than 20 million requests per second, while the broader platform now processes more than 70 million requests per second and stores over 500 petabytes.
What changed
Habitat began in 2023 as a Python client library in front of Azure Cosmos DB. As more products adopted it, storage behavior such as routing, authorization, encryption, schema lookup, request shaping, and connection pooling became difficult to update safely across dozens of independently deployed services. OpenAI therefore moved that logic behind a standalone service, creating one control point for deployment, observability, access policy, audit logging, data residency, and storage routing.
The standalone service initially remained in Python. OpenAI says it scaled by limiting per-process concurrency, running many worker processes, monitoring asyncio scheduling delay, moving connection management toward Istio and Envoy, and using HTTP/2 multiplexing to reduce downstream connection load. It also changed connection reuse from last-in-first-out to first-in-first-out after identifying a feedback loop that concentrated traffic on already slow workers.
In the second quarter of 2026, two engineers working with Codex and GPT-5.5 rewrote the service in Rust. OpenAI says the Rust implementation is now serving 95% of production requests and that the remaining Python traffic will be retired in the coming weeks.
The practical consequence
Centralizing the storage layer lets OpenAI change routing and security controls once instead of waiting for every product service to ship a compatible client. That reduces rollout coordination and gives the company a single enforcement point between application code and storage systems. Habitat deliberately exposes a constrained NoSQL object-and-edge API rather than arbitrary SQL, making request cost more predictable and isolating complex analytics through change-data-capture feeds to separate Rockset instances.
The Rust rewrite is principally a capacity and latency change rather than a new user-facing feature. OpenAI reports that it uses six times less CPU and fifteen times less memory than the Python version, with lower average and tail latency. If those measurements hold across the full workload, the migration reduces the infrastructure needed for the same request volume and creates headroom for further growth.
Limits and evidence
All scale and efficiency figures come from OpenAI. The company did not publish raw telemetry, hardware configurations, traffic mixes, test duration, absolute CPU or memory use, or before-and-after latency distributions, and the results have not been independently reproduced. The sixfold and fifteenfold figures therefore describe this service and its production workload, not a general benchmark of Rust against Python.
The migration is not complete, and the article is the first of a planned two-part series. It does not yet explain the database-layer work behind 500 petabytes and 70 million requests per second, nor does it quantify the migration’s cost, incident rate, correctness risk, or effect on end-user latency. OpenAI also says Habitat’s limited query model shifts complex joins and graph traversal work to clients or secondary systems, so its scalability comes with an explicit flexibility tradeoff. DrComps did not audit the code or reproduce the measurements.