OpenAI and Anthropic are switching off Claude Sonnet 4.5 on November 30, GPT-5 and o3 on December 11, and GPT-4o and GPT-4.1 on January 14. Paste the prompts your product depends on. Last Baseline runs the retiring model and its replacement side by side, flags every answer that changed shape, and saves the old model's answers to a file you keep.
A retired model returns an error. From that day nobody can find out what it would have said about your prompts, so there is nothing left to compare a replacement against. The comparison is only possible while both models still answer.
Use the inputs your product actually sends: a system prompt and a handful of real user messages. Or import a JSONL or CSV export of your logs.
Your browser calls the retiring model and the replacement with your own key. You see both answers with every changed word highlighted, plus latency and token change.
Download the old model's answers as a checksummed archive. Load it later to test any future model against it, long after the original is gone.
Most migration bugs are not about quality. The new model is usually better. They are about shape: the code around the model expected one thing and now gets another.
JSON that no longer parses, keys that went missing or appeared, answers newly wrapped in a code fence, markdown where plain text used to be.
Refusals the old model did not give, one word labels that flipped, answers cut off at the token limit, replies that doubled or halved in length.
Latency and output tokens for each model on your own prompts, so the bill and the response time do not surprise you after the switch.
Read from the vendors' own deprecation pages on November 13, 2026. Sources: OpenAI deprecations and Anthropic model deprecations.
| Model and API IDs | Switched off | Replacement the vendor lists | |
|---|---|---|---|
| Claude Sonnet 4.5 claude-sonnet-4-5-20250929 | November 30, 2026 | claude-sonnet-5-5 | Test it |
| GPT-5 gpt-5, gpt-5-2025-08-07 | December 11, 2026 | gpt-6.1-sol | Test it |
| GPT-5 mini gpt-5-mini, gpt-5-mini-2025-08-07 | December 11, 2026 | gpt-6.1-terra | Test it |
| GPT-5 nano gpt-5-nano, gpt-5-nano-2025-08-07 | December 11, 2026 | gpt-6-luna | Test it |
| GPT-5 pro gpt-5-pro, gpt-5-pro-2025-10-06 | December 11, 2026 | gpt-6.1-sol | Test it |
| o3 o3, o3-2025-04-16 | December 11, 2026 | gpt-6.1-sol | Test it |
| o3-pro o3-pro, o3-pro-2025-06-10 | December 11, 2026 | gpt-6.1-sol | Test it |
| GPT-4o gpt-4o-2024-08-06, gpt-4o-2024-11-20 | January 14, 2027 | gpt-6-luna | Test it |
| GPT-4o mini gpt-4o-mini-2024-07-18 | January 14, 2027 | gpt-6-luna | Test it |
| GPT-4.1 gpt-4.1-2025-04-14 | January 14, 2027 | gpt-6-luna | Test it |
| GPT-4.1 mini gpt-4.1-mini-2025-04-14 | January 14, 2027 | gpt-6-luna | Test it |
Pasting an API key into a website should make you nervous. So the tool is built so that we could not collect your key or your prompts even if we wanted to.
The tool is a static page. Model calls go from your browser straight to OpenAI and Anthropic.
The site's content security policy only allows connections to api.openai.com, api.anthropic.com and this site. No analytics, no third party scripts.
Open the network tab while it runs. The whole app is one unminified file at /assets/app.js.
Try it free on three prompts. Pay once when you want to run your whole prompt set and keep the baseline.
You pay your model provider for the API calls you run, on your own key. 14 day refund on either plan, no questions asked. Secure checkout by Stripe.
No. The tool is a static page. Your browser calls OpenAI and Anthropic directly with your key. This site's security policy only lets the page connect to api.openai.com, api.anthropic.com and our own license check, which receives nothing but a Stripe checkout id after a purchase. You can confirm all of this in your browser's network tab, and the whole app is one readable JavaScript file.
Because the comparison needs the old model to answer. Once a model is retired its API returns an error and nobody can find out what it would have said about your prompts. If you save a baseline archive now, you can still test any replacement against it next month or next year.
Answers where JSON stops parsing or loses keys, answers that gain or lose a code fence or markdown, refusals, cut off answers, one word labels that changed, large length shifts and heavy rewording. With a pass, a model of your choice also grades each pair as equivalent, improved or regressed and says why.
One run shows you where to look. Structural breaks such as invalid JSON or a new code fence are usually consistent. For wording differences, treat a flag as a lead to check and rerun the cases that matter.
The tool itself is free for three prompts per run. The API calls are billed by your provider at their normal rates, once for the retiring model and once for the replacement.
Any two text models your OpenAI or Anthropic key can reach, including across providers. The retiring models and the replacements each vendor lists are prefilled, and you can type any other model ID.
Yes. Write to clawdcomputer@gmail.com within 14 days of purchase and we refund it in full.