Log a Model Swap Before It Breaks Your Stack

A chat reply hangs, a workflow export fails, or a code path returns an error because the model ID is gone. The fix is a log, not a hot patch.

OpenAI is ending support for o1, o1-pro, o3-mini, o4-mini, gpt-3.5-turbo, gpt-4, gpt-4-turbo, gpt-4.1-nano, a gpt-4o snapshot, and gpt-image-1, effective October 23, 2026. A code change can be reverted; a missing owner, an untested route, and an unrecorded price jump cannot.

A model swap is an ops decision

A model ID is a dependency. It sits in code, env files, prompt templates, workflow exports, and internal tools. When one disappears, the failure is organizational as much as technical: nobody knows who should answer, what the fallback is, or whether the new route costs more.

Keep the log where the team already works: a shared sheet, a runbook page, or a ticket. Use columns for current ID, where it lives, replacement ID, pricing delta, fine-tune action, owner, one-token test, test date, and rollback note. Each column answers a question a founder will ask after the first error. The log points to the code, env file, or workflow export; it records the decision and the person responsible.

The checklist fills the log

Work the list below as the one-page log. Each item is a check the row needs. A row that fails any check goes back to the owner before the cutoff.

• Find the current ID in code, env files, or workflow exports.
• Name the exact file or export in the where-it-lives column.
• Assign one owner to the row.
• Enter the replacement ID next to the current ID.
• Write the pricing delta in the log.
• Set the fine-tune action to retrain, drop, or keep.
• Run a one-token request through the real path and record the response.
• Record the test date.
• Write the rollback note: how to undo the swap.

Pricing and fine-tuning decide the cost

OpenAI lists gpt-5.6-terra as the replacement for o4-mini, with standard pricing of $2 input and $12 output per million tokens versus $1.10 input and $4.40 output for o4-mini. If your product sends long answers, the output line is the one that moves the bill.

Some swaps are cheaper. The input token price for gpt-4-0613 drops from $30 to $4 per million tokens, an 87% reduction. Some are much more expensive. Moving from gpt-3.5-turbo to gpt-5.6-terra raises the output token price from $1.50 to $12 per million tokens. Record the delta in the log before you approve the replacement.

GPT-5.6 Sol pricing is promotional and runs at least through November 21, 2026. A row that depends on that rate should be marked temporary, with a note for when the promotion ends.

The ft-gpt-4 and ft-gpt-3.5-turbo fine-tunes retire on October 23, 2026, and their recommended base replacements, gpt-5.6-sol and gpt-5.6-terra, do not support fine-tuning. OpenAI will stop allowing existing customers to create fine-tuning jobs from January 6, 2027. A product that depends on a custom model needs the fine-tune action column filled: retrain, drop, or keep, with an owner named.

The one-token test proves the route

A replacement ID can look correct and still fail in production. The one-token test is the smallest proof that the route works. Send a real request through the actual path: the same env file, the same code path, the same workflow export. Ask for one token, check the response, and record the date. The date goes in the test date column.

Test the route a customer actually hits, not only the happy path. An error means the row is not done. A returned token means the row is ready for pricing review, fine-tune decision, and rollback note.

Keep the log current after the cutoff, in the same sheet or runbook. The next swap starts with an owner, a test date, and a rollback note.