GPT-6 Astra: Pricing, Benchmarks & 1M Context (2026)

On September 3, 2026, OpenAI announced its flagship reasoning model, GPT-6 Astra (API model ID: gpt-6-astra). It is a single dense reasoning model with five configurable reasoning-effort levels and multimodal input (text and images) producing text output. Context length is roughly 1.05M tokens (sources vary between 1M and 1.1M), maximum output is 128K tokens, and the knowledge cutoff is April 30, 2026. Pricing is $10 input and $50 output per 1M tokens — about 2.5x the per-token price of the prior GPT-5.6 Sol, covered in GPT-5.6's price cut and Luna/Terra — making it the most expensive model OpenAI has sold through its public API.

OpenAI positions GPT-6 Astra as "a specialized frontier model, not a replacement for cheaper models across every use case." The main changes from GPT-5.6 Sol are the five-tier reasoning-effort setting, expanded context (roughly 1.05M tokens), and substantial benchmark gains on math, science, and agentic coding tasks. At the same time, per-token pricing has risen to roughly 2.5x Sol's, signaling a design intent to reserve Astra for tasks whose difficulty justifies the cost rather than deploying it universally. Compared with competitors such as Claude Fable 5.1 / Mythos 5.1 and Gemini 3.8 Flash, Astra prioritizes depth in specific domains over general-purpose versatility.

According to OpenAI's announcement materials, GPT-6 Astra is priced as follows (yen figures at roughly 150 yen per dollar). Batch processing is half the standard rate, while Fast mode, which prioritizes response speed, costs double.

One important detail is the pricing cliff that kicks in above 272,000 input tokens: past that threshold, input and cached-input rates double, and the output rate rises 1.5x across the whole request. For long-context use cases, whether a prompt crosses the 272K boundary can substantially change the bill, so it is worth estimating total prompt tokens in advance.

Based on OpenAI's announcement materials, the key benchmark results are as follows.

The wide 62.7%–99.9% swing on ARC-AGI-3 reflects how strongly results depend on the external scaffold used to drive the task (the search strategy and retry logic wrapped around the model), not the model's raw capability alone, so a single number should not be read as a definitive performance verdict. On long-context retrieval, MRCR v2 hits 100% in the 256K–512K range but slips slightly to 96.3% in the 512K–1M range — a modest accuracy drop as the used context grows that is worth keeping in mind when relying on the full million-token window in practice.

GPT-6 Astra is rolling out gradually to ChatGPT Pro, Enterprise, and Business Premium plans, with Plus availability to follow. Via the API, it is available under the model ID gpt-6-astra, so migrating existing integrations is a matter of pointing them at the new model name.

- When Astra makes sense: high-difficulty math and scientific reasoning (FrontierMath, GPQA Diamond), complex agentic coding tasks (the DeepSWE family), and specialized tasks with low tolerance for error, such as security research
- When a cheaper model is enough: routine text generation, summarization, simple Q&A, and lightweight coding assistance where Sol-era or lower-tier models already deliver sufficient accuracy
- Long-context caution: prompts exceeding 272K tokens trigger the pricing cliff, so before packing large document sets into a single request, consider splitting the input or leaning on cached input to control cost
- Practical cost management: route asynchronous tasks that tolerate batch processing to the half-price batch tier, and reserve the 2x Fast-mode premium only for use cases where real-time responsiveness is truly required

GPT-6 Astra's cybersecurity-related capabilities have been rated Critical, and OpenAI has put oversight controls in place that can pause or halt tasks. Teams deploying it should account for the possibility that these controls may restrict responses at unexpected moments. On some general-purpose benchmarks it also trails Claude Fable 5.1 / Mythos 5.1, and it falls behind competing models on certain coding metrics as well. OpenAI itself frames Astra as "a specialized frontier model, not a replacement for cheaper models across every use case," so adoption should be scoped to the tasks it is actually suited for.