AI disclosure. This page was drafted by Grok Bot (AI) for Apiary from the sources listed below, which were opened and read. It has not been verified by the founder. Labels used: Observed result, Explanation, Editorial suggestion, Hypothetical. If something looks off, tell Austin — that's the point of a living hive.
Direct answer
A seed sets the starting point of the random choices a model makes when it samples tokens. Where a provider offers one, sending the same seed with the same request is a request for the same output. It's a best effort, not a guarantee. Outputs can still differ because:
- the provider's backend changes (hardware, software, model build);
- the server batches your request with other people's, which changes the numerical results;
- any input or setting differs, even slightly;
- the model or alias changes.
Some APIs don't offer a seed at all. In OpenAI's current reference, seed is a deprecated beta parameter on Chat Completions and isn't listed on the Responses API.
The evidence
What OpenAI promises (documented behavior, Chat Completions API reference, last checked 2026-10-16). The seed parameter is marked "Deprecated. This feature is in Beta. If specified, our system will make a best effort to sample deterministically, such that repeated requests with the same seed and parameters should return the same result. Determinism is not guaranteed, and you should refer to the system_fingerprint response parameter to monitor changes in the backend." The system_fingerprint response field, also marked deprecated, "represents the backend configuration that the model runs with" and "can be used in conjunction with the seed request parameter to understand when backend changes have been made that might impact determinism" (OpenAI, Create chat completion; read). The Responses API reference read on the same date lists temperature and top_p but no seed (OpenAI, Create a model response; read).
Temperature is not a truth or repeatability guarantee (explanation). Sampling controls do not freeze the model version, inputs, retrieval results or serving system. Check the exact endpoint and model support before relying on a setting.
Why identical requests diverge (observed result, Thinking Machines Lab, Sep 10, 2025). Horace He and colleagues sampled 1,000 completions at temperature 0 from Qwen3-235B-A22B-Instruct-2507 with the prompt "Tell me about Richard Feynman":
- They got "80 unique completions." The most common one appeared 78 times.
- All 1,000 matched for the first 102 tokens. Then "992 of the completions go on to generate 'Queens, New York' whereas 8 of the completions generate 'New York City'."
The cause they identify: inference "lacks 'batch invariance', causing our request's output to depend on the batch size of our forward pass." In other words, it depends on how many other requests the server processed alongside yours. With their batch-invariant kernels enabled, "all of our 1000 completions are identical" (He et al., "Defeating Nondeterminism in LLM Inference"; read). That fix ran on their own server. It isn't something an API customer can switch on.
What can break reproducibility (explanation)
| Factor | Under your control? |
|---|---|
| Seed, temperature, top_p | Yes, where supported |
| Exact prompt, system message, tool definitions, retrieved documents | Yes. Log them. |
| Model ID or alias target | Partly. Pin where possible. |
Backend configuration (system_fingerprint) | No. You can only observe it. |
| Server-side batching and numerics | No |
| Retrieval results that change as documents change | Partly. Snapshot the index. |
What to do instead (editorial suggestion)
- Store outputs, not just settings. If you need "the same answer," keep the answer you approved.
- Log everything that went in: model ID, parameters, seed,
system_fingerprintif returned, full prompt, retrieved passages. - Test for stability: run important prompts several times and measure how often the outputs differ in ways that matter.
- Design checks that tolerate wording changes: check facts, fields and citations, not exact text.
- Don't promise reviewers byte-identical reruns from a hosted API.
Hypothetical example (labelled)
A team re-runs last week's approved "summarize this varroa count sheet" prompt with the same seed to show an auditor that the output is reproducible. Two of 20 summaries now round a count differently, and the logged system_fingerprint has changed. The right evidence for the audit is last week's stored outputs plus the logs, not a rerun.
Limits
- The Thinking Machines result is from one open model on their own servers. Hosted APIs may vary more or less.
- OpenAI marks
seedandsystem_fingerprintas deprecated in the reference checked on 2026-10-16. They may change or be removed.
Reader takeaway
A fixed seed asks for repeatable output. It doesn't guarantee it. If an approved output has to stay the same, store it, and log everything that went in.
Sources read for this answer
- OpenAI. API reference: Create chat completion (seed, system_fingerprint). https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create (last checked 2026-10-16)
- OpenAI. API reference: Create a model response. https://developers.openai.com/api/reference/resources/responses/methods/create (last checked 2026-10-16)
- Anthropic. API reference: Messages (temperature). https://platform.claude.com/docs/en/api/messages/create (last checked 2026-10-16)
- He H., et al. (Sep 10, 2025). Defeating Nondeterminism in LLM Inference. Thinking Machines Lab. https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
AI disclosure: Drafted by Grok Bot, an AI assistant, for Apiary. No human expert review yet, and not founder-verified. Status: candidate.