What is GPT-6 Astra and when did it launch?

GPT-6 Astra is OpenAI’s frontier foundation model released on September 4, 2026, following a developer preview on September 3. It succeeds GPT-5.6 Sol, offering improvements in automated reasoning, computer and browser control, software engineering, and scientific problem-solving.

How does recurrent depth differ from test-time thinking tokens?

Traditional reasoning models like o1 and o3 emit explicit chain-of-thought tokens in text format. Recurrent depth applies recursive computational passes over internal latent state vectors within the transformer layers, allocating additional compute to ambiguous tokens without inflating visible output length.

What are the official benchmark scores for GPT-6 Astra?

OpenAI and independent evaluators confirmed 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 with Provider Adapter (62.7% standard), 95.9% on BenchCAD for 3D CAD reconstruction, 100% on ExploitBench, and 79.4% on SWE-bench Verified.

What does GPT-6 Astra cost via the OpenAI API?

Standard mode costs $10.00 per million input tokens and $50.00 per million output tokens. Fast mode costs $20.00 input and $100.00 output per million tokens. Cached prompts cost $1.00 per million read tokens and $12.50 per million write tokens. Batch and Flex requests receive a 50% discount.

How does Artificial Analysis evaluate GPT-6 Astra?

Artificial Analysis benchmarked Astra across its Intelligence Index and found that while base token prices increased 2.5x compared to GPT-5.6 Sol, Astra uses significantly fewer output tokens per task, reducing the net operational cost difference on complex tasks.

GPT-6 Astra Deep Report: Benchmarks, Architecture & API Guide

A technical analysis of OpenAI’s GPT-6 Astra, released on September 4, 2026. Covers the 1,050,000-token context window, 97.6% FrontierMath score, 99.9% ARC-AGI-3 adapter result, recurrent depth reasoning, and complete developer API pricing.

On September 4, 2026, OpenAI officially made **GPT-6 Astra** publicly available to ChatGPT Plus, Pro, Business, and Enterprise subscribers, alongside API deployments across the OpenAI developer platform, Microsoft Azure, and Amazon Bedrock under the model identifier `gpt-6-astra`. The release marks a divergence from the incremental parameter expansions of the GPT-4 and GPT-5 families. Astra pairs an expanded **1,050,000-token context window** with an internal execution architecture based on **recurrent depth reasoning**. This system allocates variable test-time compute directly within latent representations rather than relying exclusively on sequential text chain-of-thought tokens. This technical report reviews the empirical benchmark results, architectural mechanics, independent evaluations from Artificial Analysis and the ARC Prize Foundation, cybersecurity evaluations, and production developer workflows. --- ## Technical Specifications at a Glance The table below outlines the ...

GPT-6 Astra Deep Report: Recurrent Depth, FrontierMath Saturation, and OpenAI API Economics