Model retirements / GPT-5.1, GPT-5.3-Codex and GPT-5.4 nano

OpenAI deprecation · 132 days left

GPT-5.1, GPT-5.3-Codex and GPT-5.4 nano retire from the OpenAI API on 1 April 2027.

OpenAI removes gpt-5.1, gpt-5.3-codex and gpt-5.4-nano on 1 April 2027. It names gpt-6-sol for the first two and gpt-6-luna for gpt-5.4-nano. Source: OpenAI: deprecations, checked 9 Nov 2026.

Retirement date
1 April 2027
Days left
132
Official replacements
GPT-6 Sol and GPT-6 Luna

Paste your prompt. See how the replacement answers it.

We run it on GPT-5.1, GPT-6 Sol, Gemini 4 Hydrogen and Qwen4 Max: the model you're on, its official replacement and up to two cheaper current models. Same test cases for each, every answer checked, with cost and speed side by side.

$10 of free test credit · No credit card needed · We keep a pasted prompt for 24 hours (privacy)

A real example · Experiment 03

What you give us, and what you get back.

A test we ran and published, with GPT-6 Luna among the models. Yours has the same columns, with the model you're on now beside its replacement.

  1. 1You paste a prompt

    Summarise this meeting transcript. Return JSON with the keys summary, decisions and action_items, where each action item has task, owner and due.

    Decisions and action items from a transcript, without turning ideas into decisions or inventing owners.

  2. 2We test it

    • 18 test cases the prompt had not seen, each run twice
    • 8 models answer every case
    • 6 checks mark each answer right or wrong

    We also rewrote the prompt once. The figures are for the rewritten prompt.

  3. 3You get quality, cost and speed for each model

    ModelAnswers that passCost per 1,000Time per answer
    Best qualityGPT-6.1 Sol medium thinking97% 35 of 36$2.544.6 s
    Fastest worth usingMercury 3 medium thinking92% 33 of 36$0.482.9 s
    Best valueGPT-6 Luna medium thinking89% 32 of 36$0.184.5 s

    Best value: GPT-6 Luna. 89% of answers passed at $0.18 per 1,000, against 97% at $2.54 for GPT-6.1 Sol. Their likely ranges overlap, so this test alone does not show a clear lead.

Tested 4 November 2026. Cost is per 1,000 answers; time is the typical time for one answer. See every answer and the method →

What's retiring, and what replaces it

RetiringOfficial replacementOur data
gpt-5.1
GPT-5.1
gpt-6-sol
GPT-6 Sol
GPT-5.1 · GPT-6 Sol · the two compared
gpt-5.3-codex
GPT-5.3-Codex
gpt-6-sol
GPT-6 Sol
GPT-5.3-Codex · GPT-6 Sol · the two compared
gpt-5.4-nano
GPT-5.4 nano
gpt-6-luna
GPT-6 Luna
GPT-5.4 nano · GPT-6 Luna · the two compared

Dated snapshots go too: gpt-5.1-2025-11-13 and gpt-5.4-nano-2026-03-17.

Dates and replacements are OpenAI's, from https://developers.openai.com/api/docs/deprecations. Plan to have moved before the date; check the source for any change.

Why a swap can quietly change your outputs

A replacement may score higher on benchmarks, but that doesn't mean it does your task the same way. Usually nothing errors. The answers change, and your users may notice before you do.

Format
The new model may wrap JSON in a code fence, add Markdown, change headings or drop a field your code reads.
Tone
Warmer, blunter, more formal or more chatty. It reads as a different writer.
Refusals
It may refuse or hedge on inputs the old model handled, or answer ones it used to decline.
Length
Answers often get longer or shorter. That can break a limit on screen, in an SMS or in a feed.
Instructions
Edge-case rules in your prompt (never mention X, always end with Y) are where models differ most.
Cost and speed
Price per token, tokens per answer and thinking time all change, so the cost of each answer can move even at a similar list price.

Re-test before you switch

  1. Find every place your code, config and environment variables name the old model ID.
  2. Collect 10 to 20 real inputs, including the awkward ones that went wrong before.
  3. Write down what a good answer does: its format, length, what it must say and must not.
  4. Run the old and the new model on the same prompt and the same inputs, side by side.
  5. Read the answers that changed, not just the pass rate.
  6. Compare cost and time per answer, then adjust the prompt if you need to and test again.
  7. Switch before the date, and keep the cases as a regression test for the next change.

We can run both sides of these swaps for you: GPT-5.1 against GPT-6 Sol; GPT-5.3-Codex against GPT-6 Sol; GPT-5.4 nano against GPT-6 Luna.

All model retirements and tool shutdowns