How to report the AI model version in a research paper: ChatGPT, Claude, Gemini

On 2 November 2026, Nature Machine Intelligence published a short Comment, "AI needs meaningful versioning"¹. Its authors are Vasilios Mavroudis and Dan Ristea of King's College London and the Alan Turing Institute, and Lauritz Morlock and Carlos Mougan of the European AI Office at the European Commission; the paper notes that its views are not necessarily the Commission's and are given "in a personal capacity". The standfirst states the problem in two sentences: "Artificial intelligence (AI) providers frequently update their models and system integrations without corresponding changes to their public identifiers. This undermines regulatory oversight and scientific reproducibility."

For anyone writing a paper, that lands close to home. "We used ChatGPT" names a product, not a model. The model behind a product name can change from one week to the next, and even an exact API model name may be an alias that moves, or a pinned snapshot that is later switched off. A reader who wants to understand the work, or repeat it, needs more than the name on the app.

The short answer: report the provider; whether you used a chat app or the API; the model identifier exactly as the API returned it (in an app, the label it showed and your plan); the dates of use; the settings, such as reasoning effort; the tools that were on; the system prompt or custom instructions; and keep every prompt and output as supplementary material. When the result must be repeatable, use a pinned model ID, or open weights pinned to a full commit hash.

Below: what the commentary argues, as far as its public parts show; why a product name is not a version; how OpenAI, Anthropic and Google name and change their models; where to find the exact identifier; what to record and why; what journals and reporting guidelines already ask; a template for the methods section; and what to do when a model is retired. Every source was checked on 5 November 2026.

What does "AI needs meaningful versioning" argue?#

The full text is behind Nature's paywall. We read what is public: the standfirst, the figure titles, the reference list and the author notes. The standfirst ends with the aim: "We argue for meaningful AI versioning that meets the needs of regulators and researchers."¹ The two figure titles show the worked example, "Timeline of the GPT-4o sycophancy episode in ChatGPT" and "The GPT-4o sycophancy episode under a possible commitment-based versioning scheme". What that scheme would require is in the paywalled text, so we do not describe it.

The reference list points to the evidence the authors draw on. One item is "Silent Updates" by Sophia Abraham and Ben Bucknall, posted on 12 August 2026 and accepted at the AIES-26 conference (arXiv²). It describes providers changing a deployed model through fine-tuning, classifier updates, system prompt revisions, retrieval changes and routing changes, made "without public disclosure, a version increment, or re-evaluation". In a sample of nine first-party API providers and seven third-party inference hosts, it found that "no provider in our sample published information allowing an external party to verify that the artifact being served is the same one referred to in this documentation."

Another is Lingjiao Chen, Matei Zaharia and James Zou's study of GPT-3.5 and GPT-4, which the commentary cites in Harvard Data Science Review (2024); we read the arXiv version (arXiv³). They compared the March 2023 and June 2023 versions available through OpenAI's API. GPT-4 told prime from composite numbers with 84% accuracy in March and 51% in June. Their conclusion: the behaviour of the "same" LLM service "can change substantially in a relatively short amount of time".

Why isn't "ChatGPT" or "Claude" a model version?#

Because a chat app is a product that changes underneath its name. The commentary's own example shows how. On 24 and 25 April 2025, OpenAI rolled out an update to GPT-4o in ChatGPT that made the model "noticeably more sycophantic", and began rolling it back on 28 April. Afterwards it wrote: "Because we expected this to be a fairly subtle update, we didn't proactively announce it." The same post says GPT-4o had received five major updates in ChatGPT in the year since its launch there (OpenAI⁴). Throughout, the app called it GPT-4o.

This keeps happening. OpenAI's Model Release Notes for ChatGPT record an update on 29 October 2026 to "GPT-6 in Chat", for shorter answers and to "reduce overly flattering language". It applied to Free and Go users and to people not signed in. On Plus and Pro it applied only to the model ChatGPT switches to after GPT-6 message limits are reached, and "The main GPT-6 model on Plus and Pro is unchanged." (OpenAI Help Center⁵) Two people who both wrote "GPT-6 in ChatGPT, October 2026" may have used different models. The same notes describe stand-ins the picker does not show: from 18 March 2026, GPT-5.4 mini answered for GPT-5.4 Thinking when paid users reached rate limits, and "GPT-5.4 mini will not appear as a selectable model in the model picker".

Chat apps also wrap the model in a system prompt. Anthropic publishes the system prompts of claude.ai and its mobile apps, by model and date. It says the prompt "is periodically updated to improve Claude’s responses" and that "These system prompt updates do not apply to the Claude API." (Anthropic, System prompts⁶) The page for Claude Opus 5.5 has two entries already, dated 22 September and 27 October 2026 (Claude Opus 5.5 system prompts⁷). Add your own custom instructions, memory, web search and uploaded files, and the model in an app sees much more than your prompt.

Are API model names fixed?#

Some are. Each provider has names pinned to one model and names that move, and each retires names on a published schedule.

OpenAI. Each model page lists snapshots, which "let you lock in a specific version of the model so that performance and behavior remain consistent." For older models a plain name sits over a dated snapshot: the default snapshot of gpt-5.2 is gpt-5.2-2025-12-11 (GPT-5.2⁸). Some names are built to move. chat-latest "points to the latest Instant model currently used in ChatGPT", and "The underlying model snapshot will be regularly updated." (Chat Latest⁹) For some current models, such as gpt-6.1-sol, the page lists only the plain name as the snapshot (GPT-6.1 Sol¹⁰); there, the name and the dates of use are all you can give.

Anthropic. "Each Claude model ID identifies a pinned version of the model." From the Claude 4.6 generation on, IDs carry no date (claude-opus-4-6, claude-sonnet-5, claude-opus-5-5), and "Anthropic does not update the weights or configuration of an existing model ID." Earlier models carry a date, as in claude-sonnet-4-5-20250929, and their short aliases, such as claude-sonnet-4-5, point to the most recent dated snapshot. One caveat applies to all of them: "Model weights are fixed for a given ID, but the serving infrastructure around the model can change over time", which occasionally produces "minor differences in observable behavior" (Anthropic, Model IDs and versioning¹¹).

Google. Gemini model names come in four kinds (Gemini API, Models¹²). A stable name points to a specific model, and "Stable models usually don't change." Preview models can be deprecated with at least two weeks' notice. A latest alias, such as gemini-flash-latest, "will get hot-swapped with every new release of a specific model variation", after two weeks' notice by email. Experimental models "are not stable". Google's changelog shows the swaps: on 21 January 2026, gemini-flash-latest switched to gemini-3-flash-preview (Gemini API, Release notes¹³).

Even a stable name can be pointed elsewhere when it is retired. Google's release notes of 8 October 2026 say that requests to gemini-3.5-flash "are automatically routed to gemini-3.6-flash".¹³ From 1 December 2026, requests to gemini-2.5-pro will go to gemini-3.8-flash, and requests to gemini-2.5-flash to gemini-3.5-flash-lite (Gemini API, Deprecations¹⁴). Code that still names the old model gets its answers from a different one.

The notice periods come from each provider's deprecation page: OpenAI¹⁵, Anthropic¹⁶ and Google's.¹⁴

How do I check which ChatGPT, Claude or Gemini version I used?#

Through an API, read it from the response. Each provider returns the model with every response. In OpenAI's Responses API the field is model, the "Model ID used to generate the response" (OpenAI API reference¹⁷). Anthropic's Messages API returns model in each message (Claude API reference¹⁸). Gemini's generateContent returns modelVersion, "The model version used to generate the response." (Gemini API reference¹⁹) Log that field from every response, next to the name you sent: they are two facts, and the one that came back is the one to report.

OpenAI's Chat Completions API also returns system_fingerprint, which "represents the backend configuration that the model runs with". Its optional seed makes a best effort at repeatable output, but "Determinism is not guaranteed", and OpenAI points to the fingerprint for spotting backend changes (Chat Completions reference²⁰). Where it exists, record it.

In a chat app, you will not see a snapshot ID. Record what you can: the app (web, desktop or mobile), your plan, the label in the model picker, any mode such as Thinking, the dates, and a screenshot of the picker. OpenAI's release notes for 3 October 2025 say "ChatGPT will continue to tell users which model is active when asked",⁵ but a model's account of itself can be wrong, so compare any answer with the picker. Given the fallbacks described above, write what the app showed, not what you assume ran. Our guide to citing ChatGPT, Claude and Gemini covers share links and exports.

For open-weight models, the version is a repository and a commit. On Hugging Face, "By default, the latest version from the main branch is downloaded", so a repository name alone can mean different files on different days. Pass a revision to pin it, and use the full hash: "it must be the full-length hash instead of a 7-character commit hash." (Hugging Face Hub docs²¹) For example, OpenAI's open-weight gpt-oss-20b is published as openai/gpt-oss-20b, and on 5 November 2026 the repository's current commit was 6cee5e81ee83917806bbde320786a8fb61efebee (Hugging Face API²²). Add the file format and quantisation you ran (a 4-bit file is not the same set of numbers as the original), and the runtime and its version.

For agents, the model is only part of it. Record the agent tool and its version, every model it called, its permissions and tools, and keep its logs. An agent's own write-up is not a full record: see whether AI research agents hide negative results.

What should you record about the model and its settings?#

Settings matter as much as the name, and they have changed. Temperature is no longer the main dial on several current APIs. In Anthropic's API version 2026-10-21, "The temperature, top_p, and top_k request parameters are removed", and a request that sends any of them returns an error (Anthropic, API versions²³). Google lists "removed sampling parameters" among the breaking changes when moving to Gemini 3.8 Flash, with thinking_level replacing thinking_budget.¹³ The dial now is reasoning effort. OpenAI's GPT-6.1 Sol accepts reasoning.effort from low to max, with medium as the default,¹⁰ and Anthropic's defaults differ by model: medium for Claude Opus 5.5, high for Claude Sonnet 5.5 (Anthropic, Models overview²⁴). If you did not set the effort, say so and give the default.

Chen, Zaharia and Zou show the habit that helps: they state that they set the temperature to 0.1 and "left the system prompt as default".³ Two settings, two lines, and a reader knows what was held fixed.

What do journals and reporting guidelines ask for?#

Most major policies now ask for the version, in so many words. Where to put the statement is covered in our guide to writing an AI disclosure statement, and the version in an APA, MLA or Chicago reference in our guide to citing ChatGPT, Claude and Gemini.

TRIPOD-LLM, a guideline for LLM studies in healthcare, also names the alias problem directly. Its text says studies "should document the model version date and whether the model was frozen or remained dynamic during the data collection phase."³¹ CHART, for studies of chatbot health advice, says it is published jointly in nine journals, among them BMJ Medicine, The Lancet and NEJM AI.³²

A template for your methods section#

The paragraph below is an invented example: the study, numbers and dates are made up to show the form. Adapt it to your own use.

Invented example. We used Claude Opus 5.5 (Anthropic) through the Claude API to label 1,200 abstracts as reporting a randomised trial or not. Every request named the model ID claude-opus-5-5, and every response returned the same ID in its model field. Requests used API version 2026-10-21, effort "high" (the model's default is "medium"), at most 4,096 output tokens and no tools; this API version accepts no sampling parameters. All requests were sent between 6 and 10 October 2026. The system prompt and prompt template are in Supplementary File 2, and every request and response, with timestamps, is in Supplementary File 3. Two authors checked a random sample of 200 labels against the full texts (Supplementary File 4). To test whether the labels depended on one provider, we repeated the task with the open-weight gpt-oss-20b (details in Supplementary Table 1).

The supplementary table holds the details a reader would need to try again:

If AI also helped with the writing, that usually belongs in the acknowledgements or a separate AI declaration, not the methods; the disclosure guide shows where each publisher wants it.

What if the model I used is retired?#

Then requests to it fail, and the analysis cannot be rerun through the provider. OpenAI shut down gpt-4-0314, one of the original GPT-4 versions, on 26 March 2026, and gpt-4-0613 and gpt-4o-2024-05-13 on 23 October 2026.¹⁵ Anthropic retired claude-3-7-sonnet-20250219 on 19 February 2026 and will retire claude-sonnet-4-5-20250929 on 30 November 2026.¹⁶ Google shut down gemini-2.0-flash on 1 June 2026.¹⁴

Anthropic's deprecations page names the cost plainly: "Researchers lose access to models for ongoing and comparative studies." It says Anthropic has "committed to long-term preservation of model weights" and "hopes to make past models publicly available again".¹⁶ Preserved weights do not let a reader rerun your analysis today. Only what you kept does that.

So plan for retirement from the start:

Can anyone check which model actually answered?#

Not today, for the models most people use. Anthropic's Frontier Red Team put it plainly on 2 November 2026: "A user querying a model today cannot check that the responses came from the specific model that a developer claims is responding" (Anthropic³⁵). Its FlashProve prototype has the provider publish a commitment, "a short fingerprint to pin down one exact choice of weights while the weights themselves remain hidden", and then prove that given replies came from those weights. It is a research prototype, measured on one open-weight model, Llama 4 Maverick, and not a service; we explain what it does and does not prove in our FlashProve post.

The Nature Machine Intelligence commentary also speaks of a "commitment-based versioning scheme" in a figure title.¹ Whether it means the same kind of commitment cannot be told from the public parts. Until providers offer something of that kind, the model named in a paper is the authors' report, and the logs behind it are the best evidence a reader can get.

Where does Papex fit?#

A Papex record says how AI was involved, as the authors state it, including the models and tools they name; those names are self-reported, and Papex does not confirm which model ran. The person accountable signs one exact version, and a salted fingerprint of that version is timestamped with OpenTimestamps and anchored in Bitcoin, so the signed text, including any model IDs and dates in it, stays as it was signed; anyone can check the signer, version and date at /verify. A correction becomes a new version with its own date, and the earlier one stays. From the signed version, Papex can write the AI-use statement in a journal's own form, listed on /docs/journals. On /ai-claims, the model behind each AI-assisted claim on a named open problem is given as its source states it.

A record does not show that the model named is the model that ran, that its output is right, or that the work has been peer reviewed.

Sources#

1. Vasilios Mavroudis, Dan Ristea, Lauritz Morlock and Carlos Mougan, AI needs meaningful versioning (Nature Machine Intelligence, 2 November 2026), checked 5 November 2026 (standfirst, figure titles, references and author notes; the full text is paywalled and was not read) ↩ a b c
2. Sophia Abraham and Ben Bucknall, Silent Updates: Measuring and Closing the Post-Deployment Disclosure Gap (arXiv, 12 August 2026), checked 5 November 2026 (abstract) ↩
3. Lingjiao Chen, Matei Zaharia and James Zou, How is ChatGPT's behavior changing over time? (arXiv, version 3, 31 October 2023), checked 5 November 2026 ↩ a b
4. Expanding on what we missed with sycophancy (OpenAI, 2 May 2025), checked 5 November 2026 (read via the Internet Archive copy of 5 January 2026; the live site blocked automated access) ↩
5. Model Release Notes (OpenAI Help Center), checked 5 November 2026 (read via the Internet Archive copy of 30 October 2026; the live site blocked automated access) ↩ a b
6. System prompts (Anthropic, Claude Platform docs), checked 5 November 2026 ↩
7. Claude Opus 5.5 system prompts (Anthropic, Claude Platform docs), checked 5 November 2026 ↩
8. GPT-5.2 model page (OpenAI API docs), checked 5 November 2026 ↩
9. Chat Latest model page (OpenAI API docs), checked 5 November 2026 ↩
10. GPT-6.1 Sol model page (OpenAI API docs), checked 5 November 2026 ↩ a b
11. Model IDs and versioning (Anthropic, Claude Platform docs), checked 5 November 2026 ↩
12. Models (Gemini API docs, Google), checked 5 November 2026 ↩
13. Release notes (Gemini API docs, Google), checked 5 November 2026 ↩ a b c
14. Gemini deprecations (Gemini API docs, Google), checked 5 November 2026 ↩ a b c
15. Deprecations (OpenAI API docs), checked 5 November 2026 ↩ a b
16. Model deprecations (Anthropic, Claude Platform docs), checked 5 November 2026 ↩ a b c
17. Create a model response (OpenAI API reference, Responses), checked 5 November 2026 ↩
18. Create message (Claude API reference, Messages), checked 5 November 2026 ↩
19. Generating content (Gemini API reference), checked 5 November 2026 ↩
20. Create chat completion (OpenAI API reference, Chat Completions), checked 5 November 2026 ↩
21. Download files from the Hub (Hugging Face Hub documentation), checked 5 November 2026 ↩
22. openai/gpt-oss-20b (Hugging Face model API), checked 5 November 2026 ↩
23. Versions (Anthropic, Claude API docs), checked 5 November 2026 ↩
24. Models overview (Anthropic, Claude Platform docs), checked 5 November 2026 ↩
25. ICMJE Recommendations: Preparing a manuscript for submission, checked 5 November 2026 ↩
26. Generative AI policies for journals (Elsevier, updated June 2026), checked 5 November 2026 ↩
27. AI guidelines for researchers (Wiley), checked 5 November 2026 ↩
28. Using AI in your research and manuscript preparations (Taylor & Francis Author Services), checked 5 November 2026 ↩
29. Science journals' guidelines for AI usage (AAAS, PDF), checked 5 November 2026 (read via the Internet Archive copy of 27 October 2026; the live site blocked automated access) ↩
30. AI use in research practice (Springer Nature editorial policies), checked 5 November 2026 ↩
31. Jack Gallifant et al., The TRIPOD-LLM reporting guideline for studies using large language models (Nature Medicine, 8 January 2025), checked 5 November 2026 (checklist in Table 2) ↩ a b
32. The CHART Collaborative, Reporting guideline for chatbot health advice studies: the CHART statement (JAMA Network Open, 1 August 2025), checked 5 November 2026 (read in the published version, CC BY 4.0, held by LSE Research Online) ↩ a b
33. NeurIPS paper checklist guidelines, checked 5 November 2026 ↩
34. Responsible NLP research checklist (ACL Rolling Review), checked 5 November 2026 ↩
35. Progress on provable inference for large language models (Anthropic, 2 November 2026), checked 5 November 2026 ↩