Hero card for a leak report titled 'Aura — what the leak actually shows', with an UNVERIFIED badge, the subtitle 'A claim about Gemini 3.1 Pro distillation posted on 18 November 2026', three chips reading 'Source: X/@chetaslua, 18 November 2026', 'Evidence: one screenshot of a Wayback capture' and 'Status: program name unconfirmed', and two cards reading 'The claim: Gemini 3.1 Pro was handed out as a distillation teacher, then pulled after complaints' and 'The check: the doc URL returns 404 today; the documented teacher is now GLM 5.2'. The OrcaRouter logo is composited in the bottom-right corner.
Guides & Insights

Gemini 3.1 Pro Distillation: What the Aura Leak Actually Shows

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On 18 November 2026 an account posting as @chetaslua put up a three-line claim: that Google had been handing out Gemini 3.1 Pro as a distillation teacher, that Anthropic and OpenAI complained, and that Google pulled the program — with Gemini 4 Carbon offered in the same breath as the best model in the world. The only evidence attached is a screenshot of a Google documentation page, inside an Internet Archive frame, listing Gemini 2.5 Flash as the student and gemini-3.1-pro as the teacher, during early access. Nothing in the post names a Google product called Aura, and no Google page I can find uses that name for a model or a program.

So this is a leak piece, and the honest version of it is narrower than the post. Gemini 3.1 Pro is not new — Google announced it on 19 February 2026. Gemini 4 Carbon is ten days old, generally available since 10 November. What is new, and what is checkable this week, is the claim that a Gemini-teacher distillation service existed and was withdrawn. Below is what the artifact shows, what I could confirm against Google's live documentation, and what remains a single person's assertion.

What the post actually claims, line by line

Three separate assertions are bundled into 256 characters, and they do not have the same standing.

• That Google ran a distillation program with Gemini 3.1 Pro as the teacher. This is the part with an artifact behind it. The screenshot shows a documentation page titled for distillation, with a "Supported models" section reading "The following models are supported for Distillation during early access", naming a teacher model of gemini-3.1-pro and a student model of gemini-2.5-flash. The Internet Archive frame puts that capture somewhere between 18 May 2026 and 9 August 2026.

• That Anthropic and OpenAI complained, and Google withdrew it because of that. No document, filing, statement or second source is offered for the motive or the sequence. Anthropic and OpenAI have published nothing on it that I can find. Treat the cause as unsupported.

• That "Aura" is the name of this thing. Unverified. Google's Gemini 4 line is Carbon, Neon, Hydrogen and Argon; there is no Aura in the model list, no Aura endpoint, and no Aura in Google's release notes through 19 November 2026. XREAL sells a pair of Android XR glasses called AURA, which is a different product entirely and a reminder of how easily a name like this attaches itself to the wrong noun.

The quoted post in the same thread adds a fourth claim, and it is the one that frames all the others: "we never distilled other models. Distilling other models creates a poor imitation." That is a statement of principle, not of fact about the service, and the account posting it does not state an employer anywhere in the thread. I cannot verify that it belongs to anyone at Google, and this piece does not assume it does.

The screenshot is of Google's own documentation

The artifact is worth taking seriously because it is not a screenshot of a chat, a slide or a leaked deck. It is a capture of a docs.cloud.google.com page, in the Gemini Enterprise Agent Platform documentation tree, under models/tuning/distillation. Chrome and the archive toolbar are visible around it. Whatever else is true, a page with that content was live on Google's documentation site at some point between mid-May and early August 2026.

Google's own release notes corroborate that a service existed. The Gemini Enterprise Agent Platform release notes still carry a 8 May 2026 entry headed Gemini Distillation Service Early Access: "We're introduction Gemini Distillation Service in Early Access. For information about requesting access, see Gemini Distillation Service." The sentence is ungrammatical in Google's own published notes and has been left that way, which is a small sign of how much attention the entry got. It links to exactly the page path in the screenshot.

A four-row claims audit card titled 'The Aura claim — what survives', with rows reading 'Teacher: gemini-3.1-pro — shown in a Wayback capture of a Google docs page', 'Student: gemini-2.5-flash — same capture; the model is now deprecated', 'Withdrawal: pulled after rivals complained — no source given', 'Program name: Aura — not found on any Google page', above a footer reading 'Artifact: one X post, 18 November 2026; checks against Google docs run 20 November 2026'.

Today that page returns 404

Here is the one thing I could add that the post does not say, and it is a fact rather than an inference: as of 20 November 2026, the URL that Google's own release note links to does not exist. docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning/distillation returns HTTP 404. The legacy Vertex AI path that used to hold the same document 301-redirects to that dead URL, so the redirect chain now terminates in nothing.

A screenshot of Google's Gemini Enterprise Agent Platform documentation site, captured 20 November 2026, showing the browser at the distillation documentation URL with the page body reading 'Sorry, we couldn't find that page.' beneath the Google Cloud Documentation header.

The 404 is not a site-wide breakage, which is what makes it worth reporting. Sibling pages in the same navigation tree answer normally on the same day: /models/tuning, supervised-tuning, reinforcement-tuning, preference-tuning and open-model-tuning all return 200. So the distillation page was removed on its own, not orphaned by a migration. The release note that announced it, meanwhile, is still up — Google unpublished the documentation and kept the announcement, which is the ordinary shape of a withdrawn early-access feature.

What that does not tell you is when the page went. I could not get a capture list out of the Wayback Machine this week — the CDX endpoint answered 429 to every request — so the last archived capture I can point at is the one inside the post's own screenshot, dated to a 18 May–9 August 2026 window. The page could have come down in August, in October, or last Tuesday. Anyone presenting a removal date for this is guessing.

The documented distillation teacher is now GLM 5.2

This is the check that turns a rumour into a picture, and it does not need the archive at all. Google still documents a distillation tuning feature — it just documents a different one. The page that replaced the missing URL in the navigation, Supervised and distillation fine-tuning for open models, was last updated 20 November 2026 and states plainly what a distillation job can use today:

• Supported teacher models: GLM 5.2 MaaS, z-ai/glm-5.2-maas — one entry, and it is not a Gemini model.

• Supported student models: Qwen 3 at 4B, 8B, 14B and 32B, and Gemma 3 at 1B, 4B, 12B and 27B. No Gemini student either.

• Billing: the tuning job is billed on training tokens, and separately for the API calls the teacher model makes to generate the labelled data. Google's example output is a distillation_labelled_dataset.jsonl in the customer's own Cloud Storage bucket.

Read that next to the screenshot and the shape of the story changes. Distillation on Google's platform did not go away — it was narrowed, and it moved off Gemini as a teacher and onto a third-party open-weights model served as a managed service, with open-weights students. That is consistent with the post's claim that a Gemini-teacher path closed. It is not proof of why, and it is certainly not proof that two competitors' objections caused it. It is simply what the documentation says on the day the post went up plus two.

The terms already banned it in writing

There is a second, quieter piece of context, and it makes the "handing out" framing harder to accept at face value. Google's Gemini API Additional Terms, live today, say this under Use Restrictions: "You may not use the Services to develop models that compete with the Services (e.g., Gemini API or Google AI Studio). You also may not attempt to reverse engineer, extract or replicate any component of the Services, including the underlying data or models (e.g., parameter weights)."

A distillation service that uses Gemini 3.1 Pro to generate training data for a student model is, on its face, in tension with that clause unless Google carved out an explicit exception for it. An early-access program is exactly the mechanism a company would use to grant that exception narrowly, to named projects, under a separate agreement — which is what "requesting access" in the release note implies. So the most defensible reading is not "Google was giving the model away and got scared." It is "Google piloted a controlled exception to its own anti-distillation terms, and the pilot ended."

Worth noting as well: Gemini 2.5 Flash, the student in the screenshot, is itself on the way out. Google's pricing page now marks gemini-2.5-flash deprecated and says that from 1 December 2026 all API requests to it will be routed automatically to gemini-3.5-flash-lite and billed at that model's rates. A distillation recipe pinned to a teacher and a student that have both since moved is not a program anyone can be said to have "pulled" in November — it is a program that aged out of its own inputs.

What the bill would have looked like

If the program existed, the economics explain both the interest and the nervousness. Take a modest distillation run: 100,000 prompts, roughly 300 input tokens each, with the teacher producing about 1,500 output tokens per example. That is 30 million input tokens and 150 million output tokens of teacher work, before any tuning compute.

• On Gemini 3.1 Pro at Google's published rates — $2.00 per million input tokens and $12.00 per million output for prompts under 200k — the teacher alone comes to about $60 of input and $1,800 of output, roughly $1,860.

• On GLM 5.2, the teacher Google documents today, at $1.40 input and $4.40 output, the same run comes to about $42 plus $660, roughly $702.

• The student side is trivial by comparison: distilling into Gemini 2.5 Flash's price class was always the cheap half of the exercise.

Those are arithmetic on published per-token prices with the token counts stated above, not a quote from any product. But they make the shape obvious: a frontier model as a distillation teacher is the single most expensive line in the job, and it is the line where the teacher vendor has the most to lose. Both models sit on one key at OrcaRouter with provider list price passed through at 0% markup, so if you are pricing a run like this yourself the two teacher options are google/gemini-3.1-pro-preview at $2.00/$12.00 and z-ai/glm-5.2 at $1.40/$4.40 per million tokens — the same numbers Google publishes, with no second contract to sign for the second one.

What would settle it

Four things would move this from a claim to a record, and none of them exist yet.

• A dated capture. One Wayback snapshot of the distillation page after 9 August 2026 gives the removal a date. Without it, every timeline in circulation is invented.

• Google's deprecation or changelog note. Google normally records withdrawal of an early-access service. There is no such entry between 8 May and 19 November 2026 — only the original announcement, still standing.

• Any statement from Anthropic or OpenAI. The complaint is the load-bearing part of the post and the part with no artifact behind it at all.

• Anyone with access who will describe the queue. Early access means named projects and a separate agreement. One such customer describing what the console actually offered — teacher list, quotas, whether reasoning traces came back as the screenshot claims — would settle more in a paragraph than a month of archive queries.

Until then, the defensible summary is this. A Google documentation page did offer Gemini 3.1 Pro as a distillation teacher to early-access customers. That page is gone, and its neighbours are not. The feature Google documents today uses GLM 5.2 as the teacher and open-weights models as students. Whether the withdrawal was a retreat under competitive pressure or a pilot reaching its planned end is unknown, and the post that started this does not contain the answer — it contains the question.

What to do with this if you build

The practical takeaway does not depend on resolving the rumour. If you were relying on a frontier Gemini model as a distillation teacher, that path is not documented as available today, and a program that once existed is not a contract you can point at. If you are starting a distillation job now, the documented teacher is a third-party model, and the students are open weights — a combination worth testing on your own task before you commit a training budget to it.

For everything else — calling Gemini 3.1 Pro, Carbon or the GLM teacher side by side to see which one actually produces the training data you want — one endpoint and one key is the cheap way to find out. Failover across providers means an experiment does not die because one route is rate-limited, and the routing DSL lets you put two teachers against the same prompt set in a single run rather than two. That is the part of this story a reader can act on this week; the rest is a screenshot and a 404.

A screenshot of the OrcaRouter model page for Gemini 3.1 Pro Preview, captured 20 November 2026, showing the model listed with its OpenAI-compatible and Gemini endpoint types, a 1,048,576-token context window, a 65,536-token maximum output and per-million-token pricing of $2.00 input and $12.00 output.

Compared in this article4

Detected from this article · Benchmarks: Artificial Analysis · updated daily