Specifications
Gemini 4 Pro
One internal checkpoint escaped in mid-September 2026, and almost everything the internet believes about Gemini 4 Pro traces back to it. Here is what was read off that snapshot, and how much of it should be trusted.
The claim register
Each row below is one assertion, with the label describing how it reached the public. Only the rows marked Confirmed can be checked against a Google source today, and those rows describe what is absent rather than what the model does.
| Attribute | What is claimed | Confidence | Source |
|---|---|---|---|
| Release status | Not released. No Google announcement, model card, model ID or documentation page exists.Verified against the Gemini API changelog and the Vertex/Gemini Enterprise model list as of 20 September 2026. | CONF | Google AI for DevelopersGoogle Cloud documentation |
| Model name | Unconfirmed. Reported as "Gemini 4 Pro" after the Pro line reportedly skipped 3.5.Google has never used the string "Gemini 4" in public communication. | RUM | NokiaPowerUserGeeky Gadgets |
| Release window | October 2026, after a slip from the original June 2026 I/O target and a subsequent July window. | RUM | Geeky GadgetsMindStudioEmergent |
| Internal checkpoint | A first internal Gemini 4 Pro checkpoint surfaced in mid-September 2026.A checkpoint is a training snapshot, not a shippable model. Specs seen on one can change before launch. | LEAK | NokiaPowerUserNokiaPowerUser |
| Max output tokens | 256,000 output tokens.Would be roughly a 4× jump over the 64k output ceiling of the current Pro line. | LEAK | NokiaPowerUser |
| Context window | Multi-million tokens. No specific figure has leaked. | RUM | NokiaPowerUserBeatAPI |
| Reasoning modes | A selectable "high-effort" compute mode alongside standard inference. | LEAK | NokiaPowerUser |
| New capabilities | SVG generation, 3D modelling and interactive simulation output.Observed in Arena.ai side-by-side tests attributed to an anonymous checkpoint. | LEAK | NokiaPowerUserNokiaPowerUser |
| API pricing | Unknown. A $2.25 per-million-input-token figure circulates with no traceable origin. | RUM | BeatAPIkie.ai |
| API availability | No model ID, request schema, quota, region or service tier is known. | CONF | Google AI for Developerskie.ai |
What the checkpoint actually showed
The mid-September 2026 leak took the form of an internal checkpoint, with specifications subsequently read from a leaked image and from side-by-side comparisons run on Arena.ai against an unlabelled model. Three details are worth separating from the rest.
A 256,000-token output ceiling
This is the most concrete number in the whole leak. Output limits — how much the model can produce in a single response, as distinct from how much it can read — have been a quiet bottleneck for agentic and long-form code generation. A ceiling at 256k would be a step change rather than an increment, and it is specific enough to be falsified on launch day.
A high-effort compute mode
Reported as a selectable mode rather than an always-on behaviour. That framing matters: it implies a pricing and latency tier rather than a raw capability jump, and it follows the industry pattern of exposing inference-time compute as a user-facing dial.
SVG, 3D and interactive simulation output
The softest of the three claims, and the one most likely to be a demo artefact rather than a product feature. Generating valid SVG is something current frontier models already do with mixed success; "3D modelling" and "interactive simulations" are descriptions of outputs observers saw, not of a documented capability.
Why "4" and not "3.5"
The version jump is itself a piece of evidence. Between February and September 2026 the Flash line advanced through 3.5, 3.6, 3.7 and 3.8 while the Pro line stayed on 3.1. If a Gemini 3.5 Pro had been planned and then shelved, the next Pro release would have a numbering problem: shipping "3.5 Pro" months after "3.8 Flash" would read as a regression.
Jumping to 4 solves that, and it is the simplest explanation for why reporting settled on the name. It remains an inference. Google could ship 3.9 Pro, 4.0 Pro, or a name nobody has guessed.
Frequently asked questions
What is Gemini 4 Pro?
Gemini 4 Pro is the expected name for Google’s next Pro-tier Gemini flagship. It has not been announced. The name comes from reporting that the Pro line skipped version 3.5 and will jump to 4; Google has never used the string publicly.
What are the leaked Gemini 4 Pro specs?
A checkpoint that surfaced in mid-September 2026 was reported to show a 256,000-token output limit, a multi-million-token context window, and a selectable high-effort compute mode. Observers also reported SVG generation, 3D modelling and interactive simulation output. These are readings from an internal training snapshot and may not match a shipped model.
Is a 256k output limit believable?
It would be a large jump, roughly 4× the current Pro line’s output ceiling, but output limits have risen with every Gemini generation and the figure is not physically implausible. Treat it as the single most specific leaked number — and therefore the one most worth checking against the model card on launch day.
Can I access Gemini 4 Pro now?
No. There is no model ID, no preview programme and no waitlist. Any site offering Gemini 4 Pro API access today is selling access to something else. The model does not exist outside Google.
What is a checkpoint, and why does it matter?
A checkpoint is a saved snapshot of a model partway through training or post-training. Capabilities, limits and even naming routinely change between a checkpoint and a public release, and some checkpoints are never shipped at all. A leaked checkpoint is real evidence that work is happening; it is weak evidence about final specifications.