OpenAI GPT-6 Astra Architecture, Stargate 100K GPU Benchmarks & Plus Guide
⚡ Key Intelligence Takeaways
- Pre-training Compute: 100,000+ GPUs clustered at the Texas Stargate facility, actively supervised by GPT-5.6 Sol.
- Spatial & CAD Breakthrough: SOTA 95.9% on BenchCAD (vs Claude Fable 5.1 at 84.3%), generating production 3D CAD parametric models.
- Scientific Shell Power: Sets a new world record on Terminal-Bench Science at 64.6% (vs 52.6%).
- Autonomous Desktop Control: Native accessibility-tree integration with Blender, Excel, and KiCad with persistent multi-session memory.
- 100% Allocation for ChatGPT Plus: Consumer Plus subscribers ($20/mo) can dedicate 100% of normal message volume directly to Astra.
The Stargate Texas Training Cluster: 100,000+ GPUs
OpenAI has officially launched GPT-6 Astra, its flagship foundation intelligence model. In his public address, OpenAI President Greg Brockman formally characterized the deployment as the operational arrival of the AGI era.
Astra was pre-trained across an unprecedented distributed supercomputing cluster of more than 100,000 next-generation accelerators at Stargate in Abilene, Texas. Rather than utilizing standard next-token prediction, pre-training was actively supervised by GPT-5.6 Sol. Sol verified synthetic reasoning paths, eliminated flawed mathematical conjectures, and verified multi-step terminal tool executions before weights were finalized.
This automated synthetic oversight directly empowered Astra to establish a proven prime gap bound of 186, resolving an analytic number theory conjecture that had stood unsolved for decades.
Figure 1: Official launch benchmark showdown comparing GPT-6 Astra against Anthropic Claude Fable 5.1, Claude Opus 5, and GPT-5.6 Sol.
Empirical Benchmark Showdown: Astra vs. Frontier Rivals
Across verified evaluations, Astra establishes clear separation in visual induction, 3D spatial engineering, and original mathematics:
| Benchmark | Target Capability | GPT-6 Astra | Claude Fable 5.1 | GPT-5.6 Sol |
|---|---|---|---|---|
| BenchCAD | 3D Vision-to-CAD Synthesis | 95.9% | 84.3% | 71.2% |
| Terminal Science | Scientific Shell Tool Operations | 64.6% | 52.6% | 46.1% |
| Terminal-Bench 4.0 | Autonomous Software Engineering | 57.9% | 55.8% | 48.2% |
| ARC-AGI-3 | Abstract Visual Inductive Logic | 99.9% | 68.2% | 7.8% |
| FrontierMath Tier 4 | Original Research Level Proofs | 98.0% | 87.8% | 83.0% |
| ExploitBench | Zero-Day Cyber Exploitation | 100.0% | 82.0% | 74.0% |
Autonomous Computer-Use in the Desktop App
Astra's primary user-facing leap is native OS computer-use inside the official ChatGPT Desktop application:
- Blender 3D Modeling: Translates natural language engineering briefs into procedural Python execution within Blender, controlling lighting rigs, camera angles, and geometric meshes.
- Complex Financial Modeling in Excel: Audits nested workbooks, resolves circular formulas, and reconciles financial accounts without hallucinations.
- EDA & Hardware Design in KiCad: Parses schematic netlists, optimizes PCB trace routing, and outputs verified Gerber manufacturing files.
- Persistent Multi-Session Memory: Automatically persists execution state. Shutting down your machine preserves project context without needing full re-indexing.
Pricing, Access Tiers & Token Economics
In a major departure confirmed by OpenAI Product Lead Thomas Sottiaux, Astra is rolling out directly to consumer accounts:
- ChatGPT Plus ($20/month): 100% of normal usage allowance can be dedicated to Astra with zero forced fallbacks.
- ChatGPT Pro, Business & Enterprise: Higher throughput pools, admin controls, and zero data retention (ZDR).
- API Standard Pricing (
gpt-6-astra):- $10.00 per 1M Input Tokens
- $50.00 per 1M Output Tokens
- $1.00 per 1M Cached Input Tokens (Read)
- $12.50 per 1M Cache Write Tokens
- Fast Mode API: 2.5× faster tokens per second at 2× price premium ($20 input / $100 output) for trading and real-time agent swarms.
Where Astra Falls Short: Realistic Boundaries
- Terminal Coding Lead is Modest: Leading Claude Fable 5.1 by just 2.1% on Terminal-Bench 4.0 (57.9% vs 55.8%) means Claude remains fiercely competitive in CLI coding.
- Claude Opus Algorithmic Consistency: Claude Opus 5 retains a slight edge on complex single-pass algorithmic synthesis.
- Daybreak Blue Cyber Gating: Scoring 100% on ExploitBench triggered OpenAI's "Critical" risk threshold, strictly gating cyber capabilities behind the Daybreak Blue protocol.
Read the Full Technical Report on SoftReviewed
👉 Read the complete, authoritative benchmark showdown and architecture analysis on SoftReviewed:
OpenAI GPT-6 Astra: Complete Benchmark Showdown & Architecture Breakdown
Comments
Post a Comment