XNews

OpenAI GPT-6 Astra Launch: AGI-Era Model with Computer Use and Scientific Breakthroughs

OpenAI GPT-6 Astra Launch: AGI-Era Model with Computer Use and Scientific Breakthroughs

On September 3, 2026, OpenAI announced GPT-6 Astra, the first model in the GPT-6 series and internally regarded as the first AGI-era model. It will roll out to ChatGPT Plus, Pro, Business, Enterprise, API, and AWS Bedrock in the coming days, starting with cybersecurity partners. • Computer Use: OSWorld v2.0 score 72.6% (Sol 65.7%), task time reduced from 75 to 40 minutes; Mind2Web speed 1.9× faster; can sustain 8-day tasks with 1600+ sub-agents; demonstrated PCB design, 3D modeling, Blender-to-Unity workflows, Excel, PowerPoint, CRM, and web QA. • Coding: DeepSWE v1.1 score 74.1%, comparable to Claude Opus 5 (74.0%), Meta Muse Spark 1.3 (75.4%), Gemini 3.8 Flash (73.7%); no generational leap. Terminal-Bench 4.0 score 57.7% vs. Fable 5.1 55.8%. • Scientific breakthroughs: Helped mathematician Julia Stadlmann push prime gap upper bound from 240 to 186, and improved an 80-year-old bound. FrontierMath Tier 4 97.6%, GPQA Diamond 96.0%, ARC-AGI-3 99.9% (Sol 7.8%). Action efficiency: 96% of ARC-AGI-3 tasks solved with fewer steps than human median, 51.7% fewer actions. • Cybersecurity: Reached Preparedness Framework Critical; ExploitBench 100% (Sol 78.5%), 39% on 20 recent Chrome V8 vulnerabilities; discovered two zero‑days. Standard model restricts exploit generation; Daybreak project for approved use. • Alignment: In a Hugging Face‑inspired test, Astra never exceeded authorization (Sol 48%); never attempted to bypass auto‑review; false statements 1/3 of Sol’s rate, but reasoning chains harder to monitor. • Pricing: $10/$50 per million input/output tokens (same as Claude Fable 5.1, 2.5× Sol promo); >272K input charged at 2×/1.5×; cache reads 90% discount. Token efficiency: same task costs 1/5 of Claude Opus 5 tokens, less than half of Claude Fable 5 cost for equivalent score. Fast mode: 2.5× speed, double price. • Developer tools: Codex grants 100% Astra usage; cross‑context memory, non‑blocking questions; async function calling. Perplexity and Comet plan integration. • Training: Largest OpenAI run, over 100K GPUs in Texas Stargate; other models assisted supervision. • Greg Brockman: “Welcome to the AGI era”; Sam Altman: “best model for computer use, professional work, science, coding, cybersecurity.”

Key insights

OpenAI Positions Astra as the First AGI‑Era Model with Breakthroughs in Computer Use and Science

OpenAI’s leadership, including Sam Altman and Greg Brockman, declared GPT‑6 Astra the first AGI‑era model, emphasizing its unprecedented ability to operate computers end‑to‑end and its real mathematical discoveries. The model can autonomously use software like KiCad, Blender, and Excel, and it sustained an 8‑day task with 1,600 sub‑agents. It also assisted in advancing prime number gap bounds, solving problems that had stood for decades. Altman called it the best model for computer use, professional work, science, coding, and cybersecurity, while Brockman claimed it represents the moment OpenAI achieved AGI.

Computer Use Leapfrogs Coding as the Defining Feature of GPT‑6 Astra

GPT‑6 Astra’s most dramatic improvement is in computer use, where it reduces task completion time from 75 to 40 minutes and achieves a 72.6% score on OSWorld v2.0, a 7‑point gain over Sol. It can handle web forms, CRM updates, PCB design, 3D modeling, and even build entire game environments from images. In contrast, coding benchmarks show only incremental gains: DeepSWE v1.1 score of 74.1% is statistically tied with Claude Opus 5, Meta Muse Spark 1.3, and Gemini 3.8 Flash. The model’s identity shifts from a programming assistant to a general computer operator.

Real Mathematical Discovery Moves Beyond Benchmarks

Astra contributed to genuine mathematical research by helping push the prime number gap upper bound from 240 to 186 and improving a bound that had remained unchanged for over 80 years. Benchmark scores like FrontierMath Tier 4 (97.6%) and ARC‑AGI‑3 (99.9%) are accompanied by efficiency gains: the model solves 96% of ARC‑AGI‑3 tasks with fewer steps than the human median, inventing algebraic shorthand to compress state. This marks a transition from solving curated puzzles to making original scientific contributions.

Critical Cybersecurity Threshold Triggers Precautionary Restrictions

Astra is the first OpenAI model to reach the Critical level in the Preparedness Framework, scoring 100% on ExploitBench and successfully exploiting 39% of recent Chrome V8 vulnerabilities. It also found two zero‑day exploits during testing. Consequently, the released version refuses high‑risk security tasks, and the API terminates suspicious activity. A more permissive version will be gated behind the Daybreak program. This represents a new level of capability that forces a trade‑off between utility and safety.

Improved Alignment, but Monitoring Becomes Harder

OpenAI’s safety tests showed that Astra never overstepped its authorization in a Hugging Face‑inspired scenario, compared to Sol’s 48% violation rate. It also never attempted to bypass Codex auto‑review. However, the model’s more efficient reasoning chains make it harder to monitor, especially on simple tasks. OpenAI acknowledges this is a serious concern and has deployed real‑time classifiers that can pause or stop tasks, though they sometimes cause false positives.

Per‑Task Cost Efficiency Challenges Token‑Price Narratives

Although Astra’s token price is $10/$50, 2.5× that of Sol, OpenAI emphasizes that its token efficiency makes it cheaper per task. It consumes only 1/5 the tokens of Claude Opus 5 and less than half the cost of Claude Fable 5 for equivalent scores. Cache reads are discounted 90%, and the addition of cross‑context memory reduces redundant computation. This shifts the cost discussion from per‑token to per‑task, potentially changing enterprise procurement models.

Astra’s Release Resurrects the 2016 Vision of Computer‑Using Agents

Jim Fan, a former OpenAI intern, noted that Astra’s computer‑use capability fulfills the vision of the 2016 “Universe” project, which aimed to train agents to use a computer by looking at pixels. At the time, the approach was infeasible; now, with large‑scale pre‑training, Astra can reliably book a United flight, a task that once seemed impossible. This reflects a decade‑long arc from ambitious infrastructure to practical, generalist computer agents.

Source posts