GPT-6 Astra: Features, Pricing and How It Compares

GPT-6 Astra is OpenAI's flagship model, released on September 3, 2026, with the API model ID gpt-6-astra. As of October 17, 2026, it has a 1,050,000-token context window, costs $10 per million input tokens and $50 per million output tokens at standard rates, and ranks seventh on the Artificial Analysis Intelligence Index.

This guide covers what the model is, its specs and prices, what changed for developers, and where it sits on public leaderboards. Figures come from OpenAI's pages and from Artificial Analysis (AA), both read on October 17, 2026. Where a number is the company's own measurement, this guide says so.

GPT-6 Astra release date and timeline

OpenAI's API changelog shows Astra arriving in a short run of launches. The GPT-5.6 family came first, then Astra, then the GPT-6 Sol and Luna models, and then GPT-6.1 Sol.

July 9, 2026: GPT-5.6 family (Sol, Terra and Luna)
September 3, 2026: GPT-6 Astra released
September 22, 2026: GPT-6 Sol and GPT-6 Luna
September 29, 2026: GPT-6.1 Sol, with multi-agent beta support, and Ultrafast mode added for Astra

Source: OpenAI API changelog.

GPT-6 Astra specs

The model page lists the values below. OpenAI describes Astra as "Our most capable model for the most demanding work," aimed at complex reasoning, coding, computer use, research and document creation.

| Specification | GPT-6 Astra | |---|---| | Model ID | gpt-6-astra | | Context window | 1,050,000 tokens | | Max input | 922,000 tokens | | Max output | 128,000 tokens | | Knowledge cutoff | April 30, 2026 | | Inputs | Text and image | | Outputs | Text only | | Reasoning effort levels | low, medium, high, xhigh, max (no none) |

Source: OpenAI model page for gpt-6-astra.

GPT-6 Astra pricing

Standard API prices are per million tokens. The table shows Astra next to the other GPT-6 models OpenAI lists, all at short context.

| Model | Input | Cached input | Output | |---|---|---|---| | GPT-6 Astra (up to 272K input) | $10 | $1 | $50 | | GPT-6 Astra (over 272K input) | $20 | $2 | $75 | | GPT-6.1 Sol | $2 | $0.10 | $10 | | GPT-6 Sol | $2 | $0.20 | $10 | | GPT-6 Luna | $0.10 | $0.01 | $0.50 |

Astra's cache writes cost $12.50 per million tokens at short context and $25 at long context. A few billing rules apply, according to the model page and the pricing page:

A prompt over 272K input tokens is charged 2x the input and cache rates, and 1.5x the output rate, for the whole request.
Batch and Flex are 50% of the Standard price.
Fast mode costs 2x.

Source: OpenAI API pricing.

OpenAI's guide also says that Astra's estimated cost per task is lower than earlier models, even though its per-token price is higher, because it uses fewer output tokens. That is a vendor claim. Per-task cost is what matters for your workload, so measure it on your own prompts.

What is new for developers

OpenAI's latest-model guide lists several API features that come with Astra:

Async tool calling, enabled with async: true.
Mid-turn steering, so you can adjust a run while it is in progress.
Mid-conversation reasoning effort changes through a configuration_update item. OpenAI says this keeps the prompt prefix intact for caching.
Asynchronous misalignment monitoring.
Fast and Ultrafast modes for speed-sensitive work.

Astra is supported on Chat Completions, Responses and Batch. It is not supported on Realtime, Assistants, Fine-tuning, Embeddings, Images, Audio, Videos or Moderation.

Rate limits on the Standard tier are below. OpenAI cut its usage tiers from five to three (Build, Launch and Grow) on October 6, 2026.

| Tier | Requests per minute | Tokens per minute | |---|---|---| | Build | 5,000 | 1,000,000 | | Launch | 10,000 | 4,000,000 | | Grow | 15,000 | 40,000,000 |

Source: OpenAI model page for gpt-6-astra.

Migration notes for existing code

If you are moving code to Astra from an earlier OpenAI model, check these items from the migration notes:

Astra does not support effort none. Use low instead.
Tool calling requires the Responses API.
Remove temperature, top_p and top_logprobs when the effort level is not none.
If you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m".
Fast mode has no EU data residency. Ultrafast supports US residency and global processing only.

Source: OpenAI latest-model guide.

How GPT-6 Astra ranks on independent and vendor benchmarks

Artificial Analysis is an independent benchmarking site. On October 17, 2026, it gave GPT-6 Astra an Intelligence Index score of 53, ranked seventh of 217 models. Astra's output speed is 57.4 tokens per second, against a median of 83.4 for its price tier. Its blended price is $7.70 per million tokens, and one run of the index costs $3.26 on Astra.

The table below is a selected snapshot from the AA leaderboard on the same day. Parentheses show the reasoning effort AA used. Several models tie at 51 to 53, so the ranks are close.

| Model (effort) | Intelligence Index | Cost per index task | |---|---|---| | Claude Opus 5.5 (max) | 58 | $5.98 | | Claude Sonnet 5.5 (max) | 56 | $5.46 | | Claude Fable 5.1 (max) | 53 | $7.63 | | GPT-6 Astra (max) | 53 | $3.26 | | Gemini 4 Argon (high) | 53 | $1.99 | | GPT-6.1 Sol (max) | 52 | $0.72 |

Source: Artificial Analysis leaderboard.

Vendor benchmarks are a different kind of evidence. Anthropic's launch table, published September 22, 2026, compares Claude Opus 5.5 with Astra. These are Anthropic's numbers, not independent results:

| Benchmark (Anthropic, vendor-reported) | Claude Opus 5.5 | GPT-6 Astra | |---|---|---| | Terminal-Bench 4.0 | 66.4% | 57.9% | | FrontierCode v1.1 | 54.4% | 53.3% | | GDPval-AA v2.1 (Elo) | 1846 | 1542 | | AutomationBench | 40.0% | 41.4% | | Humanity's Last Exam with tools | 67.7% | 57.2% | | Terminal-Bench-Science 0.1 | 58.7% | 64.6% |

Anthropic says its benchmark margins are a less reliable guide to real-world differences. It also notes that the Astra Terminal-Bench figures are "as reported by OpenAI," and that AutomationBench counted safeguard interventions as failures. Source: Anthropic, Claude Opus 5.5.

Google's Gemini 4 Argon page includes Astra as a rival in its own table. There, Astra leads on FrontierSWE v2 (65.5%), Terminal-Bench Science 0.1 (68.1%) and the OSWorld-2.0 offline subset (72.6%). Gemini 4 Argon leads on the Vals Index, at 68.9% against Astra's 63.1%. The OpenAI pages read for this guide did not include Astra benchmark scores, so this guide quotes none.

Is GPT-6 Astra better than Claude Opus 5.5?

It depends on the measure. On the AA index, Claude Opus 5.5 at max effort scores 58 and Astra at max scores 53, and Opus costs more per index run. On Anthropic's table, Opus 5.5 leads four of the six rows, and Astra leads AutomationBench and Terminal-Bench-Science. Neither source tests your own workload, so run both on the tasks you care about. For a broader view, see Claude Opus 5.5: pricing, benchmarks and what's new.

What to watch for with Astra

OpenAI's guide also lists behaviors to plan for. These come from the company, not from an outside test:

Astra is more likely to ask clarifying questions.
It is more sensitive to instructions in AGENTS.md files and skills.
It tends toward heavy formatting.
It delegates to subagents less often.
It tests its work thoroughly.

If you use prompts written for older models, read them again before you switch. Astra's instruction handling may change how they behave.

Where voice fits

Astra returns text only. If your product needs spoken audio, you need a separate speech model to turn that text into sound. OndaLabs is an AI voice studio. Its text to speech includes 30 prebuilt voices, usable in every supported language, plus 37 language pages. Onda Lite costs 1 credit per character and is available on all plans. Onda Flash costs 1.5 credits per character and is available on paid plans. Speech to text costs 50 credits per minute. Plan prices are on the pricing page.

LLM prices are per million tokens and OndaLabs prices are per credit, so the two are not directly comparable. The comparison that matters is the one for your own script or transcript volume.

For more on how the top models compare, read The best AI models right now: October 2026 roundup.

Frequently asked questions

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's most capable model in its API, released September 3, 2026, under the model ID gpt-6-astra. OpenAI positions it for complex reasoning, coding, computer use, research and document creation. It accepts text and image input and returns text only.

How much does GPT-6 Astra cost?

At standard rates on short context, Astra costs $10 per million input tokens and $50 per million output tokens, with cached input at $1. Prompts over 272,000 input tokens cost $20 for input and $75 for output. These prices were read on October 17, 2026 and can change.

Is GPT-6 Astra better than Claude Opus 5.5?

Not on every measure. On the Artificial Analysis Intelligence Index on October 17, 2026, Claude Opus 5.5 at max effort scored 58 and GPT-6 Astra at max scored 53. Anthropic's own table shows Opus 5.5 ahead on four of six benchmarks it lists. Test both on your own tasks before choosing.

What is the difference between GPT-6 Astra and GPT-6.1 Sol?

Both have a 1.05 million token context window and both support effort levels from low to max. GPT-6.1 Sol is priced lower, at $2 per million input tokens and $10 per million output tokens at standard short-context rates. On the AA index, Sol at max effort scored 52 at $0.72 per task, against Astra's 53 at $3.26.

Does GPT-6 Astra work with the Realtime API?

No. On the OpenAI model page read for this guide, Astra is supported on Chat Completions, Responses and Batch. It is not supported on Realtime, Assistants, Fine-tuning, Embeddings, Images, Audio, Videos or Moderation.

Can GPT-6 Astra generate speech?

No. Astra's outputs are text only, so a spoken version needs a separate text to speech model. OndaLabs offers text to speech with 30 prebuilt voices, and its plans and credit costs are listed on the pricing page.