GPT-6 Astra Just Launched — and You Still Can't Use It

What GPT-6 Astra actually is on day one

GPT-6 Astra is OpenAI’s hosted flagship reasoning model, available only through the OpenAI API and ChatGPT surfaces. It is the same Astra family whose internal version produced the ten mathematical results on August 1, triggered the cyber-capability warning on August 7, and was confirmed at the Critical cyber threshold in the September 1 “Path to Astra” update. There is no separate multilingual Astra model and no downloadable checkpoint. This guide covers direct OpenAI use of Astra only.

What follows is documentation-verified against OpenAI’s model page, pricing page, launch post, and the three Astra safety and research posts, and is written for cloud integration, not local setup.

Timeline: same family, not separate models

• August 1, 2026 — Internal-Astra math. OpenAI’s math release credits “an internal version of Astra, our next major model.” First public naming of the family; no customer product yet.
• August 7, 2026 — Cyber pause and warning. OpenAI said it could not rule out critical cyber capabilities for the upcoming Astra and paused internal activities that did not meet strengthened controls. Model not released.
• September 1, 2026 — Path to Astra. The safety update confirmed Astra meets the Critical cybersecurity threshold, layered on stronger safeguards, and limited advanced cyber work to alpha testers, then Daybreak Blue. The New Stack’s launch analysis provides independent reporting on the sequence.
• September 3, 2026 — GPT-6 Astra launch. The launch post and model page make Astra available to enterprises in the Trusted Access Program, with API and ChatGPT plan access “coming in the coming days.”

Stories calling Astra “the math model” or “the cyber model” are describing one of the earlier steps above. GPT-6 Astra is the customer-facing release; access still depends on OpenAI’s account rollout and entitlement.

How you use GPT-6 Astra today

GPT-6 Astra is a closed hosted model. There is no local route and no downloadable weights.

• API model ID: gpt-6-astra on /v1/responses or /v1/chat/completions. Fine-tuning is not supported on the model page.
• Rollout language: “rolling out today for enterprises in our Trusted Access Program, with access through API and our Plus, Pro, Business and Enterprise plans coming in the coming days.”
• Context and output: 1,050,000-token context window, 128,000 max output tokens, knowledge cutoff April 30, 2026.
• Modalities: text and image in, text out. Audio and video not supported.
• Reasoning effort: low, medium, high, xhigh, max.
• Tools (Responses API): web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search.
• Versioning: the model page documents snapshot support. Before relying on a version-pinning workflow, confirm the concrete snapshot ID shown for your account.

Because Astra is closed and hosted, your only real choice is which OpenAI surface to call it from.

Pricing: the rate and the resolved discrepancy

The current OpenAI pricing page now lists gpt-6-astra in the Standard, Batch, Flex, and Fast mode tables. Earlier task notes flagged that the pricing page had not yet caught up to the launch; that discrepancy is resolved as of publication — the model is listed, and the rates below are read directly from the pricing page.

• Standard (short context, ≤272K input): $10.00 input, $1.00 cached input, $12.50 cache writes, $50.00 output per 1M tokens.
• Standard (long context, >272K input): $20.00 input, $2.00 cached input, $25.00 cache writes, $75.00 output.
• Batch and Flex: 50% of Standard.
• Fast mode: 2x the applicable rate; unavailable with EU data residency for Astra.
• Regional uplift: +10% for eligible data-residency endpoints on models released on or after March 5, 2026.

The model page restates the headline Standard rate as “$10 input / $50 output.” Treat the pricing page as canonical for cache writes, long-context tiers, and service tiers.

What launch benchmarks do and do not prove

The launch post is a vendor result page. Two patterns in OpenAI’s own writing explain why those numbers do not generalize.

First, scores move with harness choices. In an OpenAI evaluation explanation, GPT-5.6 Sol scored 13.3% on the ARC-AGI-3 public set with the official harness and 38.3% with a Responses API harness that retained reasoning across turns and used compaction. Same model, same benchmark, different setting, and the headline number changes by a factor of three. Vendor leaderboard scores measure a model-plus-harness bundle, not the model alone. The New Stack’s launch analysis is an additional independent report on the launch claims.

Second, some headline benchmarks have a funding and access conflict. Epoch AI’s FrontierMath Tier 4 (v2) methodology states: “FrontierMath was developed with funding from OpenAI, who has exclusive access to a subset of the benchmark.” When the dataset holder has exclusive access and the lab reporting the result funds the dataset, independent replication from outside is not possible.

Third, the launch results are reported by the lab whose model is being tested. Treat them as claims to verify, not facts to adopt. Your own workload is the only thing that tells you whether Astra beats your current model on the work you actually do.

Safeguards can interrupt your work

This is operational. OpenAI’s September 1 safety update says extra safety checks can slow, pause, or stop legitimate work, including defensive cybersecurity. It says legitimate activity can be flagged as potential cyber misuse or unauthorized behavior. The New Stack’s launch analysis reports the same launch-day operating constraint.

The same post says ChatGPT and Codex users may be prompted to review the action, while on the API surface “the task will stop.” Long-running agent tasks — tool loops held open for minutes or hours — should be designed so the run can terminate without warning and resume or roll back from outside the model call.

A simple decision and pilot process

1. Confirm access first. API, ChatGPT, and Codex access are rolling out on different surfaces. Verify your account can use the model before designing around it.
2. Record the exact model identifier. If your account exposes a concrete snapshot, record it with your pilot results rather than assuming an alias is immutable.
3. Use a consistent harness. If you compare models, hold the API settings, retained context, compaction, tools, prompts, and graders constant. The ARC-AGI-3 article demonstrates why this matters.
4. Plan for interruption. Treat any long-running agent task as if it will be stopped. Persist state outside the model call so resume or rollback is possible without starting over.
5. Pilot on your own workload. Run three to five representative tasks on Astra and on your current model behind the same harness. Record latency, cost, and quality with your own graders. Do not import vendor percentages.
6. Check the pricing shape. $10/$50 Standard for short context; double the input rate and 1.5x the output rate past 272K input tokens. Add the +10% uplift for data residency if applicable.

Done means

• GPT-6 Astra, the August 1 internal math Astra, the August 7 cyber pause, and the September 1 Critical-cyber Astra are all the same family; the customer-facing model launched September 3.
• The API ID is gpt-6-astra; it is closed and hosted, with no local or weights route.
• Verified numbers: 1.05M context, 128K max output, text and image in, text out, $10/$50 Standard, with Batch/Flex at 50% and Fast at 2x.
• The pricing-page discrepancy is resolved — gpt-6-astra is now listed at the rates the model page advertises.
• Launch scores are vendor results that moved materially with harness settings on ARC-AGI-3, and FrontierMath has a documented OpenAI funding and exclusive-access conflict.
• Safeguards can slow, pause, or stop legitimate work; on the API surface an interrupted task simply stops.
• You have a pilot plan that records the exact model identifier and harness, plans for interruption, and measures cost and quality on your own workload.

What this article does NOT cover

• Local or self-hosted deployment, fine-tuning, or any path to model weights — none is available.
• The internal-Astra mathematical proofs from August 1, beyond noting they came from “an internal version of Astra.”
• Alternative providers, regional availability beyond the data-residency note on the pricing page, or AWS billing terms.
• Specific benchmark percentages beyond the 13.3-to-38.3 ARC-AGI-3 public-set figure and the FrontierMath funding/access note, both sourced above.
• A step-by-step SDK recipe; this is a buying and integration brief, not a coding tutorial.

Related guides

• Enterprise model harness strategy
• How to read an LLM pricing table
• AI model matrix quickstart
• Write the rollback before you deploy
• Tests passed but the fix is not live

Sources

• GPT-6 Astra model page (OpenAI API docs)
• OpenAI API pricing page
• GPT-6 Astra launch analysis (The New Stack)
• FrontierMath Tier 4 (v2) methodology (Epoch AI)
• OpenAI launches GPT-6 Astra and says welcome to the AGI era (The New Stack)
• OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra (9to5Mac)