OpenAI

For repository work, Codex CLI inspects and edits code and runs commands from the terminal. Codex Cloud starts tasks from a published environment; each task has its own workspace and can continue while the engineer’s computer sleeps. The environment determines which project files and tools the task can actually use. The cloud-environment guide explains how that prepared starting point differs from an existing task’s working state.

For an application, the Responses API combines model output with built-in tools such as web and file search, and calls to the application’s own functions. OpenAI recommends Responses for new projects while continuing to support Chat Completions. Responses returns typed output items, so a client needs to handle tool calls as well as messages. See the Responses migration reference.

A multi-step application also needs an agent runtime: the code that carries work between model and tool calls. The Agents API uses an OpenAI-managed Codex harness. OpenAI manages sessions, orchestration, context compaction and recovery; the application supplies tools and selects an execution environment. The Agents SDK runs inside the application, with reusable agents, tools and handoffs. Its deployment and tool integrations are the application team’s responsibility, and state can use application storage or supported API conversation state. The runtime comparison makes these ownership differences explicit.

Personal ChatGPT and Codex content may be used for model improvement; users can opt out for new conversations and tasks. Codex’s separate Include environments control is unchanged by the ChatGPT or Privacy Portal setting. ChatGPT Business, Enterprise, Edu and API inputs and outputs are excluded from improvement by default. Feedback can make an associated consumer conversation eligible even after opt-out. OpenAI’s model-improvement guidance explains these account-specific controls.

API training and retention are separate. Abuse monitoring logs can include prompts and responses and are normally retained for up to 30 days, with stated exceptions. Application state persists for a feature. Zero Data Retention and Modified Abuse Monitoring require approval; eligibility and state retention vary by endpoint. Conversations are retained until deleted and are not ZDR-eligible. Data sent to remote MCP servers follows those services’ data-residency policies. Check the endpoint retention table.

The Agents API retains session state to continue work and currently supports data residency only in the United States. It does not support Zero Data Retention, including when its sandbox is self-hosted. The Agents API data limits therefore matter before placing incident material in a managed session. Running execution infrastructure yourself does not establish that the managed session has the same storage policy.

API throughput is constrained by request and token limits that vary by model, organization and project. A 429 can also indicate a configured hard spend limit, so its cause changes whether another attempt is useful. For temporary throttling, respect a valid Retry-After delay and account for retries already performed by the installed SDK. Unsuccessful requests still count toward per-minute limits; nested retry loops can multiply traffic. The rate-limit guide covers headers, SDK differences and limits on attempts and total retry time.

Prompt caching reuses processing for a matching input prefix. For GPT-5.6 and later, writes use 1.25 times the ordinary input-token rate; reads use 0.1 times that rate on most models and 0.05 times on GPT-6.1 Sol. These models support explicit and implicit breakpoints and a 30-minute minimum cache lifetime refreshed by reuse; entries may remain longer. Earlier models use different controls and pricing. The current caching reference explains those generation-specific rules. For an incident handoff, stable instructions can stay in the reusable prefix while the current observations follow it. The worked caching guide compares all attempts and accepted handoffs, rather than treating a hit as the result.

Function calling connects the model to application code: the model returns a call, the application executes the handler and sends its output back. The function-calling flow identifies that application-side step. For operational work, this is where a read-only lookup and a state-changing handler have different consequences. In a hypothetical rollback, approval identifies the deployment and the revision that was reviewed. If the write times out, the application needs a destination readback before another attempt; the rollback may already have happened. The production-agent guide develops that existing bounded-action approach.

Returned evidence also has an age. A summary of an old healthy observation cannot establish recovery now. The monitoring-freshness guide follows this through collection, evaluation and delivery. Its decision-dependent boundary applies here: older evidence can remain useful for a historical incident account, while an urgent recovery decision needs observations fresh enough for that decision.