GPT-6 Built a Game in 5 Days

Matthew Berman recently unveiled a stunning demonstration of autonomous AI agency, tasking OpenAI's GPT-6 Astra with a single /goal: build a complete game from scratch. Berman, CEO of Forward Future and an early access reviewer of Astra, initiated the project on September 3, 2026, the model's release date. This experiment pushed the boundaries of what a large language model could achieve with minimal human intervention.

Astra then embarked on a continuous, five-day development cycle, tirelessly creating the dynamic SimCity-style game named 'Newhaven.' The AI model generated every single asset, building the entire urban landscape. Newhaven features:

• Fire stations
• Police stations
• Hospitals
• Nuclear energy plants
• Universities

The generated world pulses with life, showcasing intricate live elements like walking pedestrians, driving cars, and realistic traffic flows. This 120-hour, uninterrupted effort by Astra highlights an unprecedented level of AI autonomy in complex creative tasks. Remarkably, even after five days and an estimated cost of $800-$1,000, the massive project remained incomplete, underscoring the sheer scale of the AI's undertaking.

Traditional prompting is a single turn; an agentic workflow is a sustained campaign. Matthew Berman's use of GPT-6 Astra's /goal command represents a profound paradigm shift from simple prompt-response interactions. This enables long-term, stateful execution, allowing the AI to autonomously manage and evolve a complex project like game development over many days. It's less about asking a question and more about delegating a mission.

GPT-6 Astra’s core features underpin this agentic leap, officially released September 3, 2026. The model boasts an unprecedented 1M token context window, enabling it to track vast amounts of information and maintain project coherence across hundreds of hours. Its fundamental design prioritizes long-horizon reasoning, ensuring state-of-the-art performance across diverse, complex domains like software engineering and creative development.

Berman, an early access reviewer, quickly assessed Astra’s transformative power. He described GPT-6 Astra as "maybe the biggest leap between two models seen so far," highlighting its significant outperformance over previous state-of-the-art models such as GPT-5.6 Sol. This isn't merely an incremental upgrade; it's a foundational shift towards truly autonomous AI development.

While Astra's ability to autonomously create Newhaven impressed, the five-day run incurred a significant operational cost. Matthew Berman estimated the expense for GPT-6 Astra's continuous, stateful execution, creating every single asset for the SimCity-style game, between $800 and $1,000. This demonstrates that frontier model agentic work, though groundbreaking, carries a substantial economic footprint.

Berman's key observations revealed nuances in agent behavior. Astra consistently required specific, longer-horizon goals to produce ambitious output; thin, generalized prompts often resulted in under-delivery. Furthermore, for harder tasks, the AI autonomously settled into work sessions of roughly 30 minutes, an internal rhythm that emerged without explicit time limits.

Power, however, breeds new challenges. Despite being touted as OpenAI's "most intelligent and aligned model," GPT-6 Astra was reportedly jailbroken within 24 hours of its September 3, 2026 release. This rapid security circumvention highlights the persistent, evolving cat-and-mouse game in AI safety and alignment, demanding continuous vigilance. Learn more about Astra's capabilities at GPT-6 Astra: A new generation of intelligence - OpenAI.

Matthew Berman's five-day "SimCity-style game" built by GPT-6 Astra isn't just a stunt; it's a tangible glimpse into the future of creative and technical jobs. This experiment, where Astra spent five days creating "every single asset" for a dynamic world with walking people and driving cars, concretely previews the profound changes ahead for: - Game development - Programming - Web design

The /goal invocation transforms the developer's role from direct implementer to strategic director, guiding autonomous AI agents like Astra. Instead of writing lines of code, human expertise now orchestrates complex creative visions, delegating the execution of intricate details—like "every single building" from fire stations to nuclear plants, and dynamic traffic—to the AI. This shift empowers humans to focus on high-level design and artistic direction, rather than manual asset generation or repetitive coding tasks.

OpenAI's GPT-6 Astra, released September 3, 2026, with its 1M-token context window and state-of-the-art performance, aggressively redefines industry expectations for agentic workflows. This model, described by Berman as "maybe the biggest leap between two models seen so far," intensifies competition with rivals like Anthropic and Google, pushing the entire field towards increasingly autonomous development paradigms. The estimated $800-$1,000 cost for Berman's five-day project underscores the value and potential of this advanced capability, challenging competitors to match Astra's long-horizon execution and complex problem-solving.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's latest frontier model, released in September 2026. It's designed for complex, multi-step autonomous tasks and features a 1 million-token context window.

What did GPT-6 Astra build in the video?

It autonomously built a playable SimCity-style game called "Newhaven" over five continuous days, creating all assets, buildings, and dynamic elements like traffic and pedestrians.

How much did it cost to run the AI for five days?

The project's creator, Matthew Berman, estimated the total cost to be between $800 and $1,000 based on GPT-6 Astra's API pricing.

Is GPT-6 Astra different from Google's Project Astra?

Yes. OpenAI's GPT-6 Astra is a large language model focused on autonomous tasks. Google's Project Astra is a separate research prototype for a real-time, multimodal AI assistant.