OpenAI Releases GPT-6 Astra With 1.05-Million-Token Context Window and Native Reasoning

OpenAI officially released GPT-6 Astra on September 4, 2026, making the flagship system available for API integration through multiple cloud providers with a 1,050,000-token context window and support for up to 128,000 completion tokens. The model is priced at $10.00 per million input tokens and $50.00 per million output tokens, targeting long-horizon agentic workflows, complex software development, and scientific research.

According to technical specifications published on OpenRouter, the model processes multimodal inputs including standard text, images, and PDF documents, while delivering text-based completions. OpenAI designed the system specifically for multi-step tasks requiring deep reasoning, browser navigation, and programmatic computer interaction.

Context Windows and Input Processing

The model’s 1,050,000-token context window allows developers to load extensive codebases, multi-chapter reference manuals, and research corpora directly into active memory. Up to 128,000 tokens can be generated in a single response, accommodating full application scaffolds, comprehensive scientific papers, and detailed execution logs.

Input processing covers:

• Multi-page PDF documents and structured reports.
• Static imagery and visual user interface screenshots for interactive workflows.
• High-volume plain text prompts and conversation histories.

All outputs from the model are generated as text, with support for structured JSON schemas through standard formatting parameters.

Provider Availability and Pricing Structures

Standard API access for GPT-6 Astra is listed at $10.00 per million input tokens and $50.00 per million output tokens. Ancillary rates include $1.00 per million tokens for cache reads, $12.50 per million tokens for cache writes, and $10.00 per 1,000 calls for web search functionality.

OpenRouter lists several provider endpoints offering distinct price, latency, and throughput profiles:

• OpenAI Flex: Listed at $5.00 per million input tokens and $25.00 per million output tokens, featuring a 5.43-second P50 latency and throughput of 64 tokens per second.
• Azure: Listed at $10.00 per million input tokens and $50.00 per million output tokens, with a 4.96-second P50 latency and 41 tokens per second throughput.
• OpenAI (Standard): Listed at $10.00 per million input tokens and $50.00 per million output tokens, with a 3.16-second P50 latency and 34 tokens per second throughput.
• Azure (US): Listed at $11.00 per million input tokens and $55.00 per million output tokens, with an 8.80-second P50 latency and 22 tokens per second throughput.
• OpenAI Fast: Listed at $20.00 per million input tokens and $100.00 per million output tokens, delivering a 0.70-second P50 latency and a throughput of 14 tokens per second.

Across platform usage, automated prompt caching reduced the effective weighted average input cost to $2.65 per million tokens, based on recorded cache hit rates averaging between 87% and 88% on primary endpoints.

Benchmark Performance Across Reasoning and Agentic Tasks

Evaluations indexed by Artificial Analysis place GPT-6 Astra within the upper percentiles across reasoning, agentic execution, and scientific domains. The model recorded an Intelligence Index of 54.7, ranking higher than 95 percent of evaluated systems, and an Agentic Index of 51.6, ranking above 91 percent of compared models.

On domain-specific evaluations, the model achieved the following scores:

• GPQA Diamond: 96.1% on graduate-level scientific reasoning benchmarks.
• Artificial Analysis Long Context Reasoning (AA-LCR): 80.7% on long-context retrieval and inference.
• Artificial Analysis Coding Index: 76.9%, outperforming 96 percent of compared systems.
• AA-Omniscience Accuracy: 62.6% correct response accuracy.
• SciCode: 56.5% on Python programming tasks for scientific computation.
• Humanity’s Last Exam (HLE): 54.7% across advanced academic disciplines.
• GDPval-AA: 54.2% on economically valuable professional tasks.
• AA-Omniscience Non-Hallucination Rate: 48.7% on avoidance of incorrect factual claims.
• CritPt: 31.7% on research-level physics reasoning.

API Integration, Reasoning Controls, and Tool Calling

GPT-6 Astra supports step-by-step internal deliberation via reasoning tokens. Developers can inspect the intermediate thinking process using the reasoning_details array returned in API responses, and conversation state can be maintained by passing complete reasoning blocks back to the endpoint on subsequent turns.

The API accepts standard configuration parameters:

• max_completion_tokens: Sets an upper bound on response generation up to 128,000 tokens.
• reasoning: A mapping that toggles thinking tokens, configures reasoning effort, limits reasoning length, or excludes internal chains from the final response.
• seed: Enables deterministic sampling across repeated calls.
• response_format: Constrains the output to specific JSON schemas.
• tools and tool_choice: Manages function calling and external execution schemas.

Developers interacting with OpenAI-compatible endpoints can invoke the model using standard HTTP requests:

curl -N https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
    "model": "openai/gpt-6-astra",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Analyze the attached dataset and generate a summary report."}
    ]
  }'

In addition to standard Chat Completions, the model is accessible through the OpenAI Responses API format and the Anthropic Messages API format.

Early Workload Distribution and Deployment

Platform metrics show that high-volume developer platforms and autonomous agent frameworks adopted GPT-6 Astra upon release. Public activity metrics indicate that Nous Research’s Hermes Agent generated 2.24 billion tokens, OpenAI’s Codex platform generated 1.05 billion tokens, and code editor Cursor generated 1 billion tokens, followed by omp.