scaling01 Estimates GPT-6 "Astra" at 4-5T Params with Looped Layers
On September 28, leaker scaling01 published his central estimate for OpenAI's next model, GPT-6 "Astra": roughly 4.2T total parameters, 120B activated per token, 112 layers, with 50% of layers looped once. The next day he revised his view, saying the parameter count is unlikely below 3T or above 6T, with 4-5T most plausible and 4.5-4.6T the most common result in his simulations. All claims remain personal speculation; OpenAI has not confirmed anything.
Confirmed
- scaling01's central estimate: 4.2T total parameters, 120B activated, 112 layers, 50% of layers looped once
- He declined to publish his estimation method but said evidence supports the speculation, with 4.5-4.6T most common in simulations
- In follow-up discussion he added: if total parameters were much smaller (3T), layers might need to loop 2-3 times; otherwise an active ratio as high as 2% would be required, which he considers impractical
- September 29 revision: 4-5T most likely, and he admits a habit of overestimating model sizes
- He further speculates a sparse MoE with many experts and a low active ratio
Unconfirmed
- The model's actual parameter count, layer count and looping mechanism are unverified by OpenAI, and the estimation method remains unpublished
- Whether it is indeed a sparse MoE is further speculation by the leaker
Why it matters
- Looped architectures and sparse MoE are hot topics in frontier-model discussion; if GPT-6 truly loops layers, it would trade fewer parameters for greater effective compute depth, informing views on next-generation scaling
2026-09-28 ~ 2026-09-29 · 6 related posts
Primary sources
- Leaker claims GPT-6 Astra is a 4.2T-param, 120B-active looped model; Opus 5.5 may not be looped — scaling01 · 2026-09-28
- Leaker: GPT-6 at ~3T params would need 2-3x layer looping to stay plausible — scaling01 · 2026-09-28
- [source] scaling01 narrows next-gen model size estimate to 4-5T parameters — scaling01 · 2026-09-29
3 near-duplicate retellings: scaling01 · scaling01 · scaling01