Gemini 4 Flash Leak: Five Hydrogen Checkpoints, One Speed Range, and No Model ID

Five checkpoints, one codename, and not a single model string. That is the whole of the Gemini 4 Flash leak as it stands on 5 November 2026: a claim, relayed across X, that Google has five new “hydrogen” checkpoints for a Flash-tier model of its next generation, running somewhere between 300 and 500 output tokens per second. If it holds up it would be the cheapest route to Gemini 4-class behaviour, sitting under Gemini 4 Argon and above Gemini 3.8 Flash, the cheap workhorse Google has sold since 2 September. But nothing in the claim has been published by Google — no model ID, no price, no context window, no benchmark, no screenshot, not even a consistent measurement of the speed figure that carries the story. What follows is what can be checked against primary sources, what cannot, and the one number that actually matters to anyone choosing a model this month.

What was posted, and how thin the chain is

The trail runs through a handful of accounts over about ten days. On 25 October, reports of limited grey testing appeared alongside a gemini-3.8-flash-router string found in Google Antigravity metadata; the string names a routing alias, not a model, and a developer who surfaced it, Pankaj Kumar, said explicitly that any connection to a Flash-tier Gemini 4 was his own guess. On 27 October an X account called White — roughly fifteen hundred followers, no track record of primary documents — posted that Gemini 4 Flash had “surfaced internally” under the codename Hydrogen. It carried no screenshots and no sourcing. Lumina, an account that tracks models in testing, amplified it as “apparently codenamed” Hydrogen. The same day, @lyraxana attributed a very fast model of around 300 tokens per second on LMArena to the same thing; on 4 November the same account said Hydrogen was being A/B tested in AI Studio and on LMArena, again with no screenshots. The post this article responds to is the 4 November entry from @chetaslua: Gemini has five new Hydrogen checkpoints, it is a Flash model, and the speed is around 300 to 500 tokens per second, followed by an invitation to guess which checkpoint looks best.

That account is not nobody. @chetaslua has a record of pulling unreleased Gemini builds out of AI Studio's A/B pool and posting the artefacts — the Gemini 3.0 Pro one-shot operating-system demos that circulated in the summer came from there — so the post is worth reading rather than dismissing. But a track record is not evidence about this model. Across the entire chain, nobody has posted a model ID, a rate card, a context window, a benchmark, a checkpoint hash, or a single output. Every claim resolves to one of three accounts, each relaying the others. The count of five checkpoints is not independently sourced anywhere; it is a number in a post.

A checkpoint is not a model, and five of them is a training signal

The vocabulary matters more than it looks. A checkpoint is a snapshot of weights saved during or after a training run, pulled out for evaluation, and labs produce them continuously — dozens per programme, most never named, none of them products. Five new checkpoints over roughly nine days describes a post-training sweep: someone is testing variants, comparing them, and discarding most. It tells you Google is iterating on a Flash-sized Gemini 4 model. It does not tell you that five things are coming, and it certainly does not tell you five models exist. The only reading that survives contact with how labs work is “the Flash tier of Gemini 4 is being actively tuned.”

The primary check is blunt. Google's own Gemini API model reference lists exactly one Gemini 4 model — gemini-4-argon, with a 1M-token context — and no gemini-4-flash, no Hydrogen, no Flash variant of the 4 line at all. The current Flash entry is Gemini 3.8 Flash, described there as Google's most intelligent Flash model. No public catalogue carries a Hydrogen entry either, which is what you would expect: a model that exists only as internal checkpoints has nothing to serve and nothing to bill.

Google has also never used the name. The codename pattern is real — Gemini 4 shipped as Argon, the gated cyber model announced on 28 October is Carbon, and community trackers have logged Barium-b and Xenon alongside them — but a naming convention is a container for rumours, not proof that any particular one is filled. Hydrogen fits the pattern neatly. That is precisely why it spreads.

Why 300 to 500 tokens per second is a stranger number than it looks

Two things are wrong with taking that range at face value, and only one of them is about honesty.

The first is arithmetic. A range that wide, reported alongside a count of five checkpoints, is more plausibly the spread between those checkpoints than a measurement of one model. Five variants of different sizes or at different training stages will not decode at the same rate, and a tester sampling all five would naturally produce a band rather than a point. Read that way, “300 to 500 tokens per second” is not a specification at all — it is a description of a sweep.

The second is instrumentation, and it is the reason speed claims about Google models deserve suspicion in general. Artificial Analysis currently measures Gemini 3.8 Flash at 113 output tokens per second and Gemini 4 Argon at 68. Yet September coverage of the same 3.8 Flash release confidently cited figures above 300 tokens per second. Same model, same month, a factor of nearly three — because a token rate depends entirely on where you stand: streamed through AI Studio, sampled on an arena, or averaged over an API under load. Nobody in this chain has said which. Until someone does, a Hydrogen figure of 300 is consistent with two opposite conclusions: that it is roughly four times faster than Argon, or that it is roughly where the shipping Flash already sits.

Throughput is also the wrong headline for the workloads a Flash model is bought for. Argon's time to first answer token on Artificial Analysis is 39.5 seconds; a high thinking level means the wait comes before the stream, not during it. An agent that fires twenty calls per task does not care that each stream runs at 400 tokens per second if each one idles for half a minute first. If the interesting thing about Hydrogen is that Google has made a Gemini 4 model cheap and quick off the mark, that is a latency claim, and no one has made it.

What a Gemini 4 Flash tier would actually have to beat

Here is the ground truth it would land on, read from Artificial Analysis on 5 November and from Google's published rate card:

• Output price per 1M tokens — Gemini 3.8 Flash $3.75 vs Gemini 4 Argon $10.00 today, $20.00 from 1 December

• Input price per 1M tokens — $0.75 vs $2.00 today, $4.00 from 1 December

• Intelligence Index v5.0.1 — 31 for Gemini 3.8 Flash vs 50 for Gemini 4 Argon

• Output speed — 113 tokens per second vs 68, both as measured by Artificial Analysis

• Time to first answer token — the wait before the stream starts, where Argon reads 39.5 seconds

• Context — 1,048,576 tokens on Argon, unified across input, thinking and output; Google sells no free tier for it

• Price expiry — Argon's introductory rate ends 30 November, twenty-five days out, and doubles on 1 December

Put a monthly bill against those numbers. A workload burning 50M input and 10M output tokens costs $75 on Gemini 3.8 Flash. The same workload costs $200 on Gemini 4 Argon at today's introductory rate, and $400 once the standard rate begins. That gap — more than five-fold once the standard rate begins — is the entire reason a Flash tier of Gemini 4 is interesting, and it is the thing the leak conversation keeps skipping in favour of token rates. If Hydrogen arrives at Flash pricing with anything near Argon's index, the bill stays at $75 and the capability goes up; if it arrives at Argon's price, it is a faster premium model and the cheap-tier story was wrong. Nothing posted so far distinguishes those two futures.

This is also where a routing layer earns its place rather than being bolted on. Because prices move, the useful property is not picking the right model once but being able to move between them without a rewrite. On OrcaRouter the provider's list price passes through with no markup added, so Argon's 1 December doubling lands at the same number here on the same day, and both Gemini 4 Argon and Gemini 3.8 Flash sit behind one key — a swap is a config change, not an integration. The routing DSL is the other half: put the cheap Flash tier first for a task class and fail over to Argon when the cheap call does not clear its bar, which is the only sane way to adopt an unproven model without betting a production path on it.

What Google has said, and what would make this real

Google's public position has not moved. On 24 September, DeepMind chief Koray Kavukcuoglu said Gemini 4 had entered post-training and that Google intended to release “as soon as possible,” with a timeline “much earlier” than the end of 2026. Gemini 4 Argon was announced on 30 September, reached general availability on 13 October, and Google has said nothing about a Flash sibling since — no name, no date, no price, no acknowledgement that Hydrogen exists as a label. The nearest thing to a schedule is not a Google statement at all: a Manifold market asking whether Google will release a model officially named “Gemini 4 Flash” by 13 November sits at 11%, and its resolution rules are worth reading because they encode exactly the distinction this article is making — a public preview or paid-tier release counts, while closed testers, an announcement, or a model ID spotted in code does not.

That gives a clean falsification list. Any two of the following would turn this from a rumour into a story:

• A gemini-4-flash string in Google's Gemini API model reference or changelog

• A published price line, which is the one thing a leak usually cannot fake and a launch cannot avoid

• An arena alias carrying a checkpoint hash — the standard of evidence this story lacks. The gated Carbon model left exactly that trail: a checkpoint logged on 30 October under the alias “Argon 4,” hash 01a161e8-cebc-7960-a363-fc6a0e8ab276. One trail has a hash. This one has a post.

• A catalogue entry at any serving platform, which is what makes a model callable rather than discussed

One name collision is worth flagging before it wastes anyone's afternoon. Searching for a “Hydrogen” model now also returns Hydrogen-1.0, an open-weight science model released on 2 November by Deep Principle. It has nothing to do with Google, Gemini or this leak. The shared word is the only connection.

What to do with this today

Do not put Gemini 4 Flash on a roadmap, and do not delay a decision waiting for it. If throughput is what you need now, Gemini 3.8 Flash is the shipping answer at 113 tokens per second and $0.75/$3.75, and it is the model any Hydrogen would have to displace on price rather than on speed. If depth is what you need, Gemini 4 Argon is callable today, but budget for 1 December now rather than discovering the doubling on the invoice — and read the migration notes before you wire it up, because Argon ignores temperature, top_p and top_k, so sampling settings carried over from another model silently do nothing, and its long think times mean client timeouts need raising or background mode enabling.

Keep the model string out of your code. Whatever this turns out to be, the useful posture is the one where adopting it costs an afternoon rather than a sprint: one endpoint, two model names, a routing rule that can be edited in a line, and failover so an unproven tier never becomes a single point of failure. That is worth doing regardless of whether Hydrogen is real, because the leak cycle around Google's Flash line is not going to slow down.

The honest summary is short. Google is very likely tuning a Flash-tier Gemini 4 model; the codename is plausible and the cadence fits; the specific claims in circulation — five checkpoints, 300 to 500 tokens per second — are unverifiable, partly self-contradictory, and come from accounts relaying each other without a single primary artefact. The next real signal will not be another post. It will be a string in a model list or a price on a page. Five checkpoints do not move a product roadmap. One model ID does.