Google's API leaked an "RSI model" — what the evidence actually shows

It is a strange kind of news story: no announcement, no paper, no demo. Just a string in an API response, and a very fast cleanup operation.

What actually leaked

According to reporting from 36Kr, the leak was not a roadmap or an internal memo. It was the kind of JSON payload any developer gets back when querying a model list — and it contained the name rsi-model-liverl-le, most plausibly read as "RSI Model LiveRL LE".

The parameter set attached to it is unremarkable on its face: an input ceiling of 1,048,576 tokens, an output ceiling of 65,536 tokens, and internal tags referencing 10 exclusive training slots numbered 00 through 09. Google's reaction was the interesting part — affected keys were revoked overnight, access cut before anyone could query the thing properly.

Neither Google nor DeepMind has issued a statement. That silence is currently doing a lot of work: one camp reads it as confirmation, the other as routine key hygiene after a staging label escaped into production.

A model ID is not proof of recursive self-improvement

Be precise about the term. Recursive self-improvement describes a system that improves its own capabilities fast enough that the loop outpaces the humans watching it — each generation of the model making the next generation cheaper, faster or smarter to build. It is the difference between a tool that helps you write code and a tool that writes the next version of itself.

A string in an API response does not demonstrate that. Token ceilings describe a serving configuration. Ten numbered slots could be a job queue, a shard naming scheme, or a load-testing environment. Anyone who has shipped software knows that internal identifiers are aspirational at least as often as they are descriptive — engineers name things after what they are aiming at.

So: genuinely interesting leak, unclear what it proves. The honest answer is that we do not know whether Google has a working RSI loop, a research track pointed at one, or a label someone forgot to strip before a deploy.

The numbers that do exist — and they are the real story

What is verifiable is narrower and considerably more interesting. AlphaEvolve, DeepMind's automated algorithm-discovery system, produced a 32.5% speed-up in flash attention kernels and a 23% speed-up in matrix-multiplication kernels, cutting Gemini training time by roughly 1%.

One percent does not sound like much. Applied to a training run that consumes weeks of wall-clock time and a very large amount of compute, it is not nothing — and more importantly, it was found by a machine searching a space of possible programs that no human team would have swept by hand. That is the on-ramp to RSI as it actually looks in 2026: not a model that redesigns itself overnight, but a system that quietly makes each successive training run better.

The industry context has shifted too. What was a startup pitch — Recursive Superintelligence, Auto-Research and similar efforts — is now discussed as something running inside frontier labs. Sergey Brin reportedly pushed DeepMind to accelerate Gemini's development. Anthropic CEO Dario Amodei has publicly acknowledged RSI emerging across the industry. And in the wake of the September leaks, Amodei, Sam Altman, Elon Musk and Demis Hassabis have all publicly agreed on the need to pace frontier development. That agreement, more than any leaked identifier, is the signal worth watching.

Same week, the price of a million tokens

Note the leaked model's input ceiling: 1,048,576 tokens. A year ago that was a headline. Today it is table stakes — and because it is table stakes, the number that matters is not the ceiling. It is what a million tokens costs.

Run the arithmetic on a full context window — feed the leaked model's own 1,048,576-token input ceiling through each price list and the spread is brutal. One million-token pass costs $10 on GPT-6 Astra, $0.50 on Gemini 3.8 Flash, and $0.15 on DeepSeek-V4.1-Flash at off-peak rates. That is a 20× gap between Google's Flash tier and the two $50-per-million-output flagships, and roughly 67× between those flagships and DeepSeek's cheapest window. On the output side, GPT-6 Astra's $50 per million is more than 83× DeepSeek's off-peak $0.60.

In euro terms — converted at roughly €1 = $1.08, and drifting with the rate — GPT-6 Astra works out near €9.30 in / €46.30 out per million tokens, Gemini 3.8 Flash near €0.46 / €2.78, and DeepSeek off-peak near €0.14 / €0.56. European business customers add VAT on top; the ordering does not change.

For context: blended enterprise inference costs have fallen roughly 80% over the past year, to somewhere around $25–30 per million output tokens. Picking a model by parameter count is over. Buyers now compare reasoning performance per euro and per token — which is exactly why a leaked training-slot tag is less consequential to your business than a price list.

The European angle: enforcement replaced the grace period

A year ago, a story like this would have triggered a consultation in Brussels. Not now. Since 2 August 2026, the European Commission's AI Office and national regulators directly enforce the AI Act's Article 50 transparency rules — deepfake labelling, clear disclosure that a user is interacting with an AI system — backed by financial penalties. The voluntary Codes of Practice and transitional compliance grace periods are gone.

The other deadline moved in the opposite direction. The Digital Omnibus on AI pushed mandatory compliance for Annex III high-risk systems — hiring tools, credit scoring, biometrics — out to December 2027.

For a European company, the practical consequence is this: nothing in the currently enforced regime obliges a lab to publish an internal research milestone like RSI. What is enforced is output-side transparency — whether you label generated content and disclose automated interaction. Your compliance file is built from documented capabilities and your own disclosure practices, not from a vendor's blog post about how clever its training loop is.

There is a second, quieter risk. If a system really is improving itself, the model you validated in January is not the model you are serving in June. Model version pinning in your API calls stops being housekeeping and becomes an audit requirement.

What to actually do about this

• Do not re-architect anything. There is no public RSI model to call. The Google model a European developer can reach today is Gemini 3.8 Flash, on standard Flash pricing.
• Pin your model versions and log which snapshot answered each request. Cheap insurance if a provider swaps weights underneath a stable alias.
• Route by price. Bulk extraction, classification and OCR do not need a $50-per-million-output model. DeepSeek-V4.1-Flash's off-peak window, GLM-5.3-Flash's free tier and Mistral's OCR 4.1 cover a lot of ground for a fraction of the flagship rate.
• Keep a local baseline. A 1,048,576-token context will not fit on a 16 GB consumer card. In AI Arena we deliberately run local models at far shorter contexts so we know exactly when a workload has to leave the desk and become an API line item.

What we can say with confidence is narrow: a model identifier leaked, Google moved fast, and no one has confirmed what it is. What we can say with confidence and a price list is that the economics of running these models changed far more in the last twelve months than any internal naming scheme suggests. Watch the announcements about pacing — four labs agreeing on something is rarer than any leak.