Unverified leak points to an OpenAI Astra launch next week
The account behind the claim calls Astra OpenAI's largest pretrain since GPT-4.5 and names "mewfour" as its release candidate.
By Ryan Merket · Published
Primary source: X / @synthwavedd
Why it matters
Astra would mark a return to large-scale pretraining as OpenAI's product lineup shifts toward named capability tiers. The claim remains uncorroborated and carries no performance or pricing data.
An unverified leak posted Thursday claims OpenAI is preparing to release a model called Astra during the week of August 10th, potentially putting another major model into the market roughly a month after OpenAI launched GPT-5.6.
leo (@synthwavedd) made the claim in a two-post thread on X at 7:00 a.m. on August 6th. The account described Astra as a newly pretrained model and the largest OpenAI has trained since GPT-4.5. It also identified an internal checkpoint called "mewfour" as the release candidate.
The post does not establish how leo obtained the information, and its second installment promotes a Discord community rather than supplying documentation. OpenAI has not confirmed the Astra name, the mewfour codename, the claimed training scale or the proposed release window. The account's assertions should be treated as a leak claim until OpenAI publishes a product announcement, documentation or a model card.
Astra would follow a rapid run of OpenAI releases
The timing would be unusually tight. OpenAI released GPT-5.6 on July 9th, after beginning a limited preview on June 26th. OpenAI describes GPT-5.6 Sol as its current flagship, alongside the lower-cost Terra and Luna tiers. Its API model catalog continues to recommend GPT-5.6 Sol for complex reasoning and coding.
That official lineup gives no indication of where Astra would sit. The leak calls Astra a "new pretrain," language that points toward the underlying pretrained model rather than a narrow reasoning, coding or voice release. If that description is accurate, Astra could serve as a new base for later post-trained products even if OpenAI markets the public release under a different name. The post does not specify whether Astra would appear in ChatGPT, Codex, the API or a limited research preview.
OpenAI's current naming system adds another unresolved issue. With GPT-5.6, OpenAI said the generation number and the Sol, Terra and Luna capability tiers would identify its models. Astra does not fit that published structure. Internal codenames frequently differ from public product names, making it possible that Astra refers to a training project rather than the label customers would see.
The GPT-4.5 comparison concerns training scale
The leak's central comparison is GPT-4.5, which OpenAI introduced on February 27th, 2025. OpenAI presented GPT-4.5 as an expansion of unsupervised pretraining, designed to improve general knowledge, pattern recognition and conversational judgment without relying on the deliberate reasoning process used by its o-series models.
OpenAI also called GPT-4.5 very large, compute-intensive and expensive to serve. That history matters because "largest model since GPT-4.5" is a claim about the resources and scale behind training, not proof that Astra would outperform GPT-5.6 or other current models. The leak provides no parameter count, training-compute figure, benchmark result, inference price or context-window specification.
A large new pretrain would still carry strategic weight. OpenAI said when it released GPT-4.5 that stronger pretrained models could provide a better foundation for reasoning systems and tool-using agents. Astra, under the account's description, would represent another investment in that base-model route after OpenAI spent much of its recent release cycle packaging intelligence into specialized capability and cost tiers.
The "mewfour" detail is the most specific part of the claim and remains unsupported outside the account's post. Calling a checkpoint a release candidate implies that training and internal testing have advanced far enough for OpenAI to select a prospective production build. It does not establish that deployment reviews, safety evaluations or product integration have finished.
The claimed window runs from August 10th through August 16th. Until OpenAI publishes first-party material, Astra remains a purported internal model with an asserted launch date, rather than an announced addition to OpenAI's lineup.