GPT-6 Astra Benchmarks: 99.9% ARC-AGI-3, 4x Human Excel Speed

The article evaluates OpenAI's newly announced GPT-6 Astra model through benchmark results and product details shared via OpenAI's official X post and accompanying video.

Benchmark Results

The author highlights several key benchmarks where Astra demonstrates significant performance gains over prior models:

Financial Modeling World Cup Challenge (based on the 2023 Microsoft Excel World Championship): Astra completes tasks at roughly four times the speed of the human champion .

ARC-AGI-3 (novel game world rule inference): Astra scores 99.9% , compared to GPT-5.6 Sol at 7.8% and Claude Opus 5 at 30.2%.

SRE-Bench (binary reverse engineering without source code): Astra solves 88.0% in one attempt, versus Sol at 55.9% and Opus 5 at 12.5%.

ExploitBench (turning known vulnerabilities into working exploits): Astra achieves 100% .

Terminal-Bench Science (end-to-end research workflows via code and terminal): Astra reaches 64.6%, nearly triple Sol's 22.4%.

Overseas developers have also integrated Astra into modeling software and Unreal Engine, producing results comparable to a small team's one-to-two-month effort.

Product Availability and Pricing

Astra will be accessible through ChatGPT Plus, Pro, Business, and Enterprise tiers, as well as via the OpenAI API and AWS. Early enterprise access is granted through the Daybreak project. API pricing is set at $10 per million input tokens and $50 per million output tokens . A Fast Mode offers 2.5× speed at double the standard price.

Developer Tooling

For developers, Astra functions both as a chat model and as the underlying engine for Codex-style engineering agents capable of reading repositories, editing files, running tests, generating pull requests, and explaining failures. The article notes that domestic developers facing network or payment barriers can use services like Code80, which proxies real subscription accounts via API endpoints.

Frequently Asked Questions

Biggest change for programmers: AI shifts from generating code snippets to handling entire tasks, requiring clearer requirements, acceptance criteria, and boundary definitions.

Token pricing relevance: Per-token cost becomes a secondary metric; agent workflows should be evaluated on total task cost including call count, failure rate, human intervention, and rework time.

Security focus areas: Execution permissions, network access, sensitive data access, command auditing, and sandbox isolation are critical as model agency increases.

Domestic integration: Use official APIs or proxies like Code80, with emphasis on key management, permission governance, and audit logging.

}