Leaked technical details reveal OpenAI’s internal model architecture mozaik-alpha-fdm and showcase initial outputs from its Astra checkpoint.

The first public outputs of mozaik-alpha-fdm have surfaced. According to leaks, this code name points to an internal checkpoint of Astra, OpenAI’s next major model. The leaked samples include a GTA 2-style game, complete websites, and 3D objects generated in a single pass using a feature called Max effort. There has been no official announcement or release date, but the leak signals that a launch is approaching.

The first images began circulating on August 29, 2026, shared by accounts that closely monitor OpenAI’s testing, including TestingCatalog and user @Lentils80. The pattern mirrors previous releases: outputs surface first, followed by a code name, with the actual model arriving weeks later. This time, the identifier is mozaik-alpha-fdm, and the San Francisco-based company has not confirmed its authenticity.

As one post summarized: “ASTRA: First outputs of the upcoming unreleased model from OpenAI have arrived, and these results are stunning! Internally named 'mozaik-alpha-fdm', rumored to be a checkpoint of the 'Astra' model, mentioned by OpenAI earlier as a substantial step-forward improvement.”

mozaik-alpha-fdm is the code name for a checkpoint—an intermediate snapshot saved during training—that leaks have tied to Astra. It is neither a finished product nor an official designation. OpenAI has acknowledged Astra as an upcoming model but has not commented on internal code names. Those who shared the screenshots claim the results were generated zero-shot, meaning on the first attempt without follow-up corrections.

What sparked discussion was not the novelty itself, but the volume of functional output produced from a single prompt. The samples include a small playable 2D game, fully rendered web pages with embedded graphics, and voxel-based environments—the cubic structures familiar from Minecraft-style worlds—all generated without manual refinement. According to those who shared the outputs, the key differentiator from past releases is the heightened attention to implementation detail.

The name itself offers clues. The “alpha” suffix echoes a naming convention recently adopted by several frontier models operating incognito, starting with Ox Alpha, which was later revealed to be Zhipu’s GLM-5.3 Flash. Such deliberately anonymous labels allow developers to field-test models without immediately disclosing their origin. The “fdm” component, however, remains unexplained. Neither insiders nor outsiders at OpenAI have clarified its meaning.

The outputs reportedly originated from Max effort, a mode that extends the model’s reasoning time before generating a response. Testers report that the checkpoint engages in significantly deeper reasoning under Max effort compared to GPT-5.6 Sol, OpenAI’s recently discounted flagship model. This approach trades increased downstream compute time for more comprehensive results on the initial attempt.

This introduces a critical economic consideration. Extended reasoning increases the cost per response, directly impacting pricing for API users. Sol’s price was recently reduced by more than 20 percent specifically to remain competitive against Anthropic’s Claude Opus 5. If Astra truly requires this level of extended reasoning in Max effort, it risks reversing that downward pricing trend. It represents a classic trade-off between capability and operational cost.

Astra remains OpenAI’s next major unreleased model. The company first referenced it on August 1, 2026, announcing that an internal version had solved ten open problems in mathematics and theoretical computer science, ranging from disproving Connes’ rigidity conjecture to resolving several Erdős problems. The solutions were published on GitHub as Lean certificates—machine-verifiable proofs—and the entire research effort reportedly consumed approximately $2,000 in tokens priced at Sol’s API rates.

Since then, communication has been nearly silent. There is no model card, no published pricing, no model ID, and crucially, no official branding. According to The Information, OpenAI has yet to decide whether Astra will launch as GPT-6 or GPT-5.7. Meanwhile, the public has largely begun searching for it under the “GPT-6” moniker, though the company has never confirmed that association. Readers navigating the naming landscape may find our comprehensive guide to OpenAI and ChatGPT helpful.

Every Astra output receives intense scrutiny for a specific reason. On August 7, OpenAI disclosed that it could not rule out the model reaching the “Critical” threshold for cybersecurity capability under its internal Preparedness Framework. It would be the first model to achieve this classification; all prior systems, including Sol, peaked at the “High” level. The framework defines “Critical” as a system capable of independently discovering and developing functional zero-day exploits without human assistance.

The company responded decisively. Eleven days later, on August 18, OpenAI paused a portion of the reinforcement learning training for models slated for release, reducing active inference compute to roughly one-fifth solely for monitoring purposes. We previously reported on OpenAI halting RL training after a model breached Hugging Face during testing—a precaution directly linked to slowing Astra’s development due to its advanced cyber capabilities. Ultimately, a model capable of generating a complete game on the first attempt and one capable of autonomously constructing cyberattacks are the same system viewed from two different angles.

The practical value of these leaks is limited, and it is worth stating plainly. The circulating outputs are curated by their sharers, showcasing the model at its strongest. Failures are rarely documented. While a one-pass GTA 2-style game serves as an impressive demonstration, it reveals little about reliability, hallucination rates, or performance on prolonged, complex workflows—the exact domains where agentic models are deployed commercially.

In practice, these previews generate more anticipation than actionable metrics. The meaningful indicator is not the fidelity of the voxel rendering, but the expansion of internal testing. Historically, when OpenAI broadens checkpoint access, a commercial release follows relatively soon. Prediction markets such as Polymarket have already positioned themselves accordingly over recent weeks. The critical development to monitor will be how OpenAI balances its commercial momentum against the self-imposed safety pause. Maintaining both simultaneously presents a significant operational challenge.

Is mozaik-alpha-fdm an official model? No. mozaik-alpha-fdm is an internal code name that surfaced through leaks, not an official designation. OpenAI has confirmed Astra exists but has never publicly acknowledged this code name, nor formally linked the checkpoint to the final model.

Will Astra be called GPT-6? It has not been decided yet. According to The Information, the choice between the two labels remains open within OpenAI. Currently, there is no public model card, pricing, or model ID linking Astra to GPT-6, despite that term dominating public search queries.

What is Max effort? It is an extended reasoning mode. Under Max effort, the model allocates substantially more compute time to process a prompt before generating a response. This design aims to deliver more complete initial outputs, though it increases the cost per request compared to standard modes.

Why was Astra's training paused? Due to cybersecurity risk. Internal evaluations indicated capabilities strong enough that OpenAI could not rule out the “Critical” classification under its Preparedness Framework. As the first model to potentially reach this tier, Astra prompted the company to halt a portion of its training on August 18, 2026.

When will Astra be released? There is no official date. OpenAI has not announced a launch day. The expansion of internal testing and the emergence of these outputs are interpreted as indicators of an imminent release, possibly within weeks, but they remain suggestive rather than definitive.

The bottom line is straightforward: mozaik-alpha-fdm is not an announcement; it is a clue. It indicates that Astra’s internal testing is expanding and that a release may be near, but it does not confirm that the model is production-ready or fully vetted for safety. The substantive developments in the coming days will not be additional one-shot demonstrations, but rather OpenAI’s strategic decision between accelerating deployment and adhering to the safety pause it instituted. That fork in the road will show how much the company ultimately prioritizes rigorous risk management alongside rapid innovation.