Leaked
Leaked image puts output ceiling at 256k tokens and adds a high-effort mode
Specifications read off a leaked image: a 256,000-token output limit, a multi-million-token context window, and a selectable high-effort compute mode.
Specifications attributed to the leaked checkpoint were read off a circulated image, and separately inferred from Arena.ai comparisons against an unlabelled model.
The three concrete claims:
- 256,000 output tokens. Roughly a 4× increase over the current Pro line’s ceiling, and the single most specific number in the whole leak.
- Multi-million-token context. No figure attached — the phrasing is directional, not measured.
- A high-effort compute mode. Described as selectable, which reads more like a pricing and latency tier than a capability.
Observers also reported SVG generation, 3D modelling and interactive simulation output in the Arena.ai sessions. Attribution to Google in anonymous arena testing is inferred from output style, not confirmed, so that layer of the claim is weaker than the rest.
The output figure is the one worth remembering, because it is precise enough to be checked against the model card on launch day.