Article

Migrate GPT-5 and o3 API apps before December 11

For teams running GPT-5 or o3 API apps, identify affected models, update tool calling and reasoning, and compare replacement token costs.

Editorial illustration for Migrate GPT-5 and o3 API apps before December 11: an arrow connects an old service to its replacement. Not documentary evidence.

OpenAI will shut down the listed GPT-5, GPT-5 mini, nano and pro, o3 and o3-pro API models on December 11, 2026, according to its retirement schedule. Move affected requests to supported models before December 11. Apps that let models call tools may also need an API interface change, while every migration needs a fresh cost check.

The notice dates to June 11, 2026; this is an approaching deadline, not a new announcement. It concerns the specific OpenAI API model IDs below, not every product with GPT-5 in its name or a ChatGPT subscription.

Which model IDs must you replace?

Use the exact IDs sent by your application. The following are OpenAI’s current replacement recommendations, checked on November 21. A dated snapshot is affected too; pinning it does not avoid this shutdown.

Retiring alias

Retiring snapshot

Recommended replacement

gpt-5

gpt-5-2025-08-07

gpt-6.1-sol

gpt-5-mini

gpt-5-mini-2025-08-07

gpt-6.1-terra

gpt-5-nano

gpt-5-nano-2025-08-07

gpt-6-luna

gpt-5-pro

gpt-5-pro-2025-10-06

gpt-6.1-sol with reasoning.mode: "pro" in Responses

o3

o3-2025-04-16

gpt-6.1-sol

o3-pro

o3-pro-2025-06-10

gpt-6.1-sol with reasoning.mode: "pro" in Responses

Inventory deployed settings, actual requests, scheduled jobs, queued batches and fallback routes—not only the default in source code. For each affected feature, record its owner, endpoint, tools and effective reasoning settings. These recommendations are starting points for validation, not guarantees of identical answers.

If a supplier runs the backend, ask it to identify affected models and provide a cutover plan. If you use another cloud provider or gateway, verify its own lifecycle rules. Confirm replacement access, rate limits and regional requirements in the project that will run the app.

Do your tool calls need the Responses API?

A tool call lets a model ask an application to do something, such as retrieve a record. Check this separately from whether the replacement can return plain text:

  • GPT-6.1 Sol and GPT-6.1 Terra support Chat Completions and Responses, but tool calling requires Responses. Requests without tools do not need to leave Chat Completions solely for this restriction.

  • GPT-6 Luna permits function calling in Chat Completions only with reasoning_effort: "none". Use Responses when combining reasoning with tools.

  • For either retiring Pro model, the recommended route is Sol in Responses with reasoning.mode: "pro". Pro mode and reasoning effort are separate controls, as OpenAI’s reasoning guide explains.

If an interface move is needed, follow the Responses migration guide through the whole request-and-result path:

  • Send input instead of messages and handle typed items in response.output. Do not assume every item is a text message.

  • Update function definitions and return each function_call_output with the matching call_id. Preserve required reasoning and tool items when carrying conversation history forward.

  • Move Structured Outputs schemas from response_format to text.format and update streaming handlers for typed Responses events.

  • Preserve your storage policy. Responses stores responses by default; use store: false where required. For stateless reasoning continuity, follow the guide’s encrypted-item flow.

Which reasoning and cache settings change?

Set the effective effort explicitly rather than inheriting a new default. Sol supports low, medium, high, xhigh and max; Terra and Luna also support none. All three default to medium.

OpenAI’s GPT-6 migration guidance recommends starting at low when replacing minimal, then comparing representative tasks. When reasoning is enabled, remove temperature, top_p and top_logprobs. Also remove Chat Completions’ logprobs or message.output_text.logprobs from Responses’ include. Check parameters added by wrapper libraries too.

Where an older integration sets prompt_cache_retention, that guidance says to replace it with prompt_cache_options.ttl: "30m". Review cache billing as well as request syntax.

How much could the token bill change?

At the Standard prices checked on November 21, each million ordinary input/output tokens costs $2/$10 on Sol, $1/$5 on Terra and $0.10/$0.50 on Luna. Tokens are the units the model processes and generates.

Illustrative calculation: hold usage at 10 million ordinary uncached input tokens and 1 million billed output tokens across multiple requests, with no request exceeding 272,000 input tokens. Use Standard processing and standard reasoning mode. Exclude cached reads, cache writes, tools, retries, storage, regional premiums, taxes and migration work.

Migration

Retiring model total

Replacement total

GPT-5 → GPT-6.1 Sol

$22.50

$30.00

GPT-5 mini → GPT-6.1 Terra

$4.50

$15.00

GPT-5 nano → GPT-6 Luna

$0.90

$1.50

o3 → GPT-6.1 Sol

$28.00

$30.00

The formula is 10 × the input rate per million + 1 × the output rate per million. For mini, that is 10 × $0.25 + $2 = $4.50; for Terra, 10 × $1 + $5 = $15. These are fixed-token comparisons, not forecasts of the cost of completing the same task.

Billed output includes reasoning tokens, not just visible answers. Sol’s pro mode bills aggregate model work at Sol’s token rates, so a cheaper rate than an old Pro model does not establish a fixed saving. Measure total usage on acceptable results before setting a production budget.

Caching is enabled by default on supported models. Cache reads and writes have separate rates; a cache-write rate replaces ordinary input pricing for those tokens rather than being added to it. For Sol, Terra and Luna, 30-minute writes cost $2.50, $1.25 and $0.125 per million tokens respectively under Standard short-context pricing. The illustration above includes neither writes nor reads.

Long requests also change the calculation: the replacement model pricing rules double input and cache rates and multiply output rates by 1.5 for the full request when input exceeds 272,000 tokens. The threshold applies to each request, not monthly usage.

Check dependencies with earlier deadlines

Applications using saved prompt objects, hosted Evals or Agent Builder have separate November 30, 2026 deadlines in OpenAI’s feature-retirement schedule. Deal with those dependencies first where applicable. The guides to moving prompt objects into code and replacing Agent Builder workflows cover those distinct tasks.

Validate the replacement and finish the switch

Before increasing production traffic, use representative cases with clear acceptance criteria. A successful text response is not enough for a workflow that must also retrieve data or perform an action.

  1. Check answer quality, extraction schemas, tool arguments and the resulting actions. Include multi-turn conversations, streaming and tool failures.

  2. Check incomplete outputs and error handling. The reasoning token limit includes reasoning as well as visible output; a request can exhaust max_output_tokens before producing an answer.

  3. Compare latency and total billed usage, including reasoning, cache activity and retries. Keep the replacement’s cost and response time within the feature’s requirements.

  4. Switch traffic in stages, then inspect live and queued requests for every retiring ID. Keep a validated, supported fallback; a retired model cannot be your post-deadline rollback.

Assign an owner and finish before December 11. The published schedule does not specify a shutdown hour or timezone, so do not plan around availability throughout that day.