OpenAI lists the API model snapshot gpt-4o-2024-05-13 for shutdown on October 23, 2026. Its API deprecations table names gpt-5.6-sol as the substitute. If an application still sends the exact old identifier, its operator needs to find that route, test a replacement and remove the old calls before the deadline. Check the live table again before making the change; model availability and documentation can change.
This is an API snapshot deadline. OpenAI retired GPT-4o in the ChatGPT app on February 13, 2026, under a separate announcement. The October API row names gpt-4o-2024-05-13 specifically. It does not say every GPT-4o API identifier shuts down that day.
The guide below uses OpenAI's published deprecation, model and API documentation, checked October 11, 2026. BIG CHANGE has not run a customer's integration, made a test API call or measured the substitute's behavior. The search and rollout plan is an operator procedure to apply in the reader's own environment.
Step 1: Find every place the snapshot is selected
You need access to the application repositories and the settings that reach each deployed environment, including deployment manifests, environment variables, secret references and job definitions. Start with a literal search for the full model ID in source and text configuration. For a repository with ripgrep installed, run:
rg -n -F 'gpt-4o-2024-05-13' .The expected output is a list of files and line numbers containing that string, or no matches in the searched tree. This command searches the current tree; it cannot see values injected at deployment, a remote configuration service or a runtime default. Record each matching call path, its owning service, deployed environment and the task it performs. Then inspect the actual model value in each deployment configuration and any approved runtime request metadata or logs. Search the same literal value in infrastructure repositories and scheduled jobs. Treat a zero-match repository search as a prompt to check the other sources, not proof that production is clear.
OpenAI's Responses API reference and Chat Completions API reference both document a model request field. Find where your application sets that field, including wrappers that choose a model from an environment variable. If you have permission to use the API, OpenAI's List models endpoint can show identifiers available to your account. Availability there is a useful access check, but it does not establish that your prompts, output parsing or tools work with the substitute.
Keep the inventory in your normal change record. It should name the old identifier, each environment and owner, the replacement to test, the acceptance checks and the rollback decision. Do not paste API keys, customer prompts or sensitive responses into that record. OpenAI's quickstart documents creating an API key and supplying it through OPENAI_API_KEY; use your organization's approved secret handling and access controls.
Step 2: Test the listed substitute in staging
OpenAI's deprecations table lists gpt-5.6-sol as a substitute, not a promise that changing a string preserves behavior. The model page lists its supported endpoints and features. Compare those with each inventoried call before editing its model setting. The API references also caution that parameter support can differ by model. Keep endpoint or parameter changes explicit in the change review.
In a staging environment with authorized API access, set model to gpt-5.6-sol for one call path at a time. Keep a rollback setting for the last working production configuration while the old snapshot remains available. Use representative, permitted inputs from that path: for example, the application's existing evaluation cases and failure cases, with sensitive data removed or approved for this use. The input is the same application task under the proposed model setting; the expected output is an API response your application can parse and evaluate. Do not infer equivalence from a successful HTTP response alone.
Set acceptance checks before running the comparison. Check the fields your code consumes, any required JSON schema or tool-call behavior, task-specific correctness judged against known answers or human review, error handling, latency and token use at the volume your service expects. OpenAI's Responses reference documents a returned model, status and usage with input and output token counts; Chat Completions documents its own returned model and token usage. Capture only the request and response metadata your privacy policy permits. If a case fails, inspect the request parameter, response shape and application parser, then revise and rerun that case. A substitute that cannot pass the application's checks needs an alternate reviewed plan before the deadline.
Step 3: Roll out and prove the old route is gone
After staging meets the checks, update the approved configuration and deploy through the service's normal review and release process. Have a person approve the change and the production rollback threshold, especially where generated output affects customers or consequential decisions. During rollout, compare errors, task checks, latency and usage against the agreed baseline. If the release fails its criteria, use the prepared rollback while it is still viable, then repair and retest. The old snapshot is scheduled to become inaccessible on October 23, so it is not a durable rollback target beyond that date.
Finally, inspect actual production request metadata for each inventoried call path over a window that covers its normal traffic and scheduled jobs. Confirm the outgoing or returned model identifier is gpt-5.6-sol and count any remaining gpt-4o-2024-05-13 requests. A passing migration has zero observed old-snapshot calls across the covered paths, with the replacement meeting the agreed task checks. Record the observation window and any paths that did not run; no traffic is inconclusive for those paths. If an old call remains, trace it to its deployment or configuration owner, update that route and repeat its staging and production checks.
Cost and limits
OpenAI's GPT-5.6 Sol model page currently lists standard short-context text prices of $4 per million input tokens and $20 per million output tokens, with separate cached-input and long-context rates. The page says this promotional pricing is available at least through November 21, 2026. Actual cost depends on the selected processing mode, context size, input and output tokens, caching and any paid tools. Check the live pricing page and your response usage before budgeting. This guide makes no comparison with the retiring snapshot's price and no savings estimate.
The model page lists supported endpoints and features, but access and rate limits depend on the account and usage tier. It does not establish compatibility with a particular application's prompt, response parser or quality threshold. The only way to settle those questions for a deployment is to run its own representative checks. Do not expose credentials or production data to a test environment without authorization.
The big change
OpenAI's April 22 deprecation notice gives operators a fixed date for one pinned API snapshot: gpt-4o-2024-05-13 is listed to shut down October 23, 2026, with gpt-5.6-sol named as substitute. The practical work is to locate the exact identifier wherever a deployed request can acquire it, check the replacement against real tasks, and observe production long enough to catch rarely used paths. The documentation supplies the deadline and API interfaces; each operator's staging and production evidence must establish whether its migration works.
Sources & further reading
- OpenAI API deprecations lists the April 22, 2026 legacy-snapshot notice and the October 23 row for
gpt-4o-2024-05-13. It namesgpt-5.6-solas substitute. Recheck this living page before release. - GPT-5.6 Sol model page lists supported endpoints, features, limits and current model-specific prices. Those listings are not an application compatibility test.
- Responses, Chat Completions and List models document model selection, returned metadata and account-visible model identifiers. Read the reference for the endpoint your application actually uses.
- OpenAI API pricing provides current rates and processing-mode details; pricing may change after this article's October 11 check.
- OpenAI's ChatGPT retirement notice describes the separate February 2026 ChatGPT change. It is context for distinguishing the app from the API snapshot.
