OpenAI's Secret Models: From Rumors to a Millennium Prize

Episode 1 · OpenAI's future model 'Bel' reportedly finishes pretraining in August (2026-08-28, 2 posts)

OpenAI has run unusually heavy pretraining since late 2025; a leak claims Sol finished in March, Astra in April/May, and the mysterious long-term model 'Bel' completed pretraining in August.

• OpenAI reportedly running many pre-trains; far-future model codenamed 'Bel' — haider1 · 2026-08-28
• Rumor: OpenAI's "Bel" model finished pre-training in August — haider1 · 2026-08-29

Episode 2 · OpenAI's Next-Gen Model Astra Reportedly Trained on Over 100K GPUs (2026-09-02, 10 posts)

OpenAI's next-generation model, internally codenamed Astra, has reportedly completed training at the Stargate site in Abilene, Texas, using over 100,000 GPUs across roughly two months of pretraining — a run Emad Mostaque reshared claims will be the first billion-dollar training run. Blogger imjustnewatai says Sam Altman confirmed the training is done and that Astra is OpenAI's first model significantly supervised by a previous-generation model during training. None of this is officially confirmed.

Confirmed

• The Information reports OpenAI considered naming Astra 'GPT-6' but dropped the idea (m3, m10); m6 adds the name may narrow to 'GPT-5.7 Astra' or 'GPT-Astra', saving the true generational leap for a later codename Bel
• Multiple posts (m2, m7, m9) agree on the scale: 100K+ GPUs, OpenAI's largest training run to date
• Internal tester Lentils80 (relayed by @OkDisplay3159, m1) disclosed two checkpoints: ultima-alpha, a likely public release candidate, and vega-alpha, a cybersecurity-focused enterprise version

Unconfirmed

• op7418 claims Astra is already deployed but not yet open, will offer 100% capacity to all paying users including Plus, with rollover reset credits — unverified
• Long-horizon task demos (8-day runs, 1,600 sub-agents, a Xiaomi 17 Ultra case) come from secondhand accounts
• haider1's claims of 10 research-level math/CS advances and OpenAI internal research-intern benchmark parity are opinion plus rumor; the $1B cost figure lacks an official source

Why it matters

• If the 100K-GPU, $1B scale holds, Astra marks a new order of magnitude in training cost and validates Stargate's role
• The old-model-supervises-training paradigm, if confirmed, would be a new scaling approach
• Dropping the GPT-6 name shows naming itself has become part of OpenAI's expectation management
• maksymandr's PostTrainBench Lite results look impressive but await independent verification

• OpenAI reportedly passed on the GPT-6 name for Astra, saving the 6-worthy jump for Bel — Angaisb_ · 2026-09-02
• Astra hailed as a watershed model, reportedly passing OpenAI's internal AI-research intern benchmark — haider1 · 2026-09-03
• Insider leak: OpenAI's Astra tested as 'ultima-alpha' and 'vega-alpha' checkpoints — Ok_Display_3159 · 2026-09-03
• OpenAI considered naming Astra 'GPT-6' but passed, The Information reports — rohanpaul_ai · 2026-09-04
• OpenAI weighed the 'GPT-6' name for Astra before dropping it, says The Information — rohanpaul_ai · 2026-09-04
• GPT-6 Astra scores impressively on PostTrainBench Lite benchmark — maksym_andr · 2026-09-04
• Leak claims GPT-6 Astra trained on 100,000+ GPUs at OpenAI's Stargate site — BLUECOW009 · 2026-09-04
• Astra reportedly trained on 100k GPUs at Stargate Texas in first $1B training run — ai · 2026-09-04
• Leaked GPT-6 Astra demo shows 8-day task runs, 1600 sub-agents, and 3D modeling — op7418 · 2026-09-04
• OpenAI "gpt-next" Leak: Astra Trained on 100k GPUs With Old Models Supervising — imjustnewatai · 2026-09-04

Episode 3 · OpenAI Researcher Says GPT-6 Astra Will Be Obsolete in Weeks (2026-09-04, 4 posts)

On September 4, OpenAI researcher roon said on X that the newly released GPT-6 Astra will be "obsolete" within weeks. The remark was widely relayed and seen as a rare candid statement from an OpenAI insider about the pace of model iteration—given how fast R&D is moving, any flagship release may only be a fleeting milestone. Several bloggers went further: the comment adds plausibility to earlier rumors of "a new round of large-scale pretraining at OpenAI" and "a new 10T-parameter model," suggesting OpenAI may have another, bigger flagship model held back for later this year; @markk even speculated about moves such as Grok 4.7 from competitors based on this. OpenAI has not officially responded or confirmed anything. Confirmed

• roon himself posted on X the remark that Astra would be "obsolete within weeks" (relayed by @Tolopono and others with consistent content).
• The remark continues roon's long-held view that "the pace of model iteration outstrips any current release."

Unconfirmed

• The existence of a "10T new pre-trained model" and any release date this year are rumors with no official evidence.
• Relay posts differ on Astra's ownership: most refer to OpenAI's GPT-6 Astra, while the conversation version relayed by @UltraRareAF refers to Google's Astra—a factual discrepancy that remains to be verified.
• @markk's inference that Grok 4.7 is coming is personal speculation.

Why it matters

• An internal researcher publicly downplaying the company's latest flagship may signal a more aggressive release cadence and reflects the frontier-lab norm of "obsolete upon release," which is relevant to model procurement and product planning.

• OpenAI researcher roon says GPT-6 Astra will be obsolete within weeks — Tolopono · 2026-09-04
• OpenAI Researcher roon: GPT-6 Astra Will Be Obsolete in Weeks — Tolopono · 2026-09-04
• OpenAI Employee Hints Astra May Be Obsolete in Weeks, Fueling New Big Model Rumors — haider1 · 2026-09-04
• Rumor: OpenAI employee hints major new model coming, Astra possibly obsolete within weeks — mark_k · 2026-09-04

Episode 4 · OpenAI reveals internal model behind Navier-Stokes breakthrough is significantly stronger than GPT-6 Astra (2026-09-09, 12 posts)

In announcing its solution to the Navier-Stokes millennium problem, OpenAI disclosed that the proof was produced by a group of agents using its next-generation internal model — significantly stronger than the publicly released GPT-6 Astra, which had launched just four days earlier — and that the model's training is still ongoing. The public-facing Astra is therefore not OpenAI's capability ceiling.

Confirmed

• Per Axios, the model OpenAI used on Navier-Stokes is significantly more capable than the released GPT-6 Astra.
• OpenAI's official wording: the proof was produced by agents using its next-generation model, significantly stronger than GPT-6 Astra; relayed by Omar Sanseviero of dair-ai.
• Matthew Berman adds that Astra was released only four days prior, and that the next-gen model is still in training.
• Polymarket, VraserX and others relayed the same official disclosure consistently.

Unconfirmed

• A leak claims OpenAI restarted its largest frontier RL training run on August 28, and an internal checkpoint from that run solved the problem at a cost of millions of dollars; single-source, unverified.
• Per a relay by r0ck3t23, unconfirmed reports say the internal system used 10,000 agents over 88 hours, and the model is unavailable to ordinary users.
• haider1 and imjustnewatai relay rumors of a new pretraining model codenamed Fable, intended to surpass Astra.

Why it matters

• kimmonismus notes OpenAI having stronger internal models is no secret, but the official "significantly stronger" phrasing is bold.
• rohanpaulai estimates OpenAI Frontier Labs' internal models lead what the public can access by over a year.
• VraserX argues this gap between the strongest internal and public models widens the perception gap between the public and frontier labs — public benchmarks and APIs may systematically understate OpenAI's true frontier capability.

• Report: OpenAI's Millennium Problem model is internal, "significantly more capable than GPT-6 Astra" — kimmonismus · 2026-09-09
• "Significantly more capable than GPT-6 Astra": OpenAI's internal model gap — kimmonismus · 2026-09-09
• dair-ai amplifies OpenAI's next-gen model claim behind Navier-Stokes solution — dair_ai · 2026-09-09
• Commentary Claims OpenAI Frontier Labs Run Models a Year Ahead of the Public — rohanpaul_ai · 2026-09-09
• OpenAI used a model 'significantly more capable' than Astra for Navier-Stokes, per Axios — 141_1337 · 2026-09-09
• OpenAI Quietly Discloses a Model Far Beyond GPT-6 Astra, Released Just 4 Days Ago — MatthewBerman · 2026-09-09
• OpenAI reveals internal AI model significantly more capable than GPT-6 Astra — Polymarket · 2026-09-09
• Leak: OpenAI's new internal model solved Navier-Stokes after largest RL run restart — scottleibrand · 2026-09-09
• Rumors: OpenAI has an internal model more capable than GPT-6 'Astra' — haider1 · 2026-09-09
• Rumor: OpenAI has an internal model stronger than GPT-6 Astra — imjustnewatai · 2026-09-10
• OpenAI claimed 10,000 agents cracked 90-year-old Navier-Stokes problem in 88 hours — r0ck3t23 · 2026-09-10
• OpenAI says its Navier–Stokes system is far more capable than GPT-6 Astra — VraserX · 2026-09-10

Episode 5 · Report: OpenAI Training New Model That Beats Its Own GPT-6 Astra (2026-09-09, 4 posts)

Citing OpenAI's blog, users report a new internal model started training on August 28 and already surpasses GPT-6-Astra and Google's Astra on math benchmarks within days, fueling speculation that GPT-6.5 pretraining is underway.

• OpenAI May Already Be Pretraining GPT-6.5 That Beats Astra on Evals — willdepue · 2026-09-09
• OpenAI model trained in just 11 days beats Google's Astra on math — adonis_singh · 2026-09-09
• Speculation swirls: OpenAI's mystery model may be math-specialized pretraining beating GPT-6 Astra — giffmana · 2026-09-09
• OpenAI's new model, in training since Aug 28, reportedly beat GPT-6-Astra in just one week — alexcovo_eth · 2026-09-10