What Is GPT-6 Astra? OpenAI’s Latest GPT-6 Features, Pricing, API, Release Date, and How to Use It in ChatGPT

OpenAI announced GPT-6 Astra on September 3, 2026, and put the API model gpt-6-astra into the developer docs the next day. The official one-liner: “our strongest model, built for the hardest end-to-end work.” Developers should ignore the slogan and ask: how it differs from GPT-5.6 Sol, what it costs, which ChatGPT surface actually exposes it, and whether to change the default route.

This piece follows the OpenAI model page, the pricing page, and the ChatGPT help center. Prices and entry points are as of documents visible on September 8, 2026. During the rolling launch, trust the in-product model picker.

What GPT-6 Astra is

GPT-6 Astra is OpenAI’s strongest public text flagship. It is not another “smoother chat” model. The docs place it on the hardest end-to-end work: complex reasoning, software engineering, Computer Use, research, and document writing. Input is text and images; output is text. Audio and video are not in this model’s modalities.

Versus the previous tier, GPT-5.6 Sol, Astra’s change is task duration and the tool loop: it is built to chain repo-scale coding, desktop actions, retrieval, and writing into one request, instead of asking you to click “continue” every two minutes. That is the same industry line as Claude Fable 5 putting long-running agents on stage—competition moved from “one good answer” to “finish the job unattended.”

The official knowledge cutoff is April 30, 2026. When you need newer facts, use tools such as Web search. Do not assume it knows the world after May.

Release date and rollout

OpenAI also said it will offer the model later through Azure and AWS Bedrock. Until those marketplace entries exist, do not hard-code those IDs in production routing. On the Codex side, they are experimenting with “notes across multiple context windows”: long sessions are no longer crushed into one summary, and older windows stay searchable. That is a product signal for coding marathons, not an SLA.

Features: reasoning tiers, tools, and Computer Use

Astra’s capability list compresses to three layers: think deeper, use tools, and hold a long context.

• Reasoning tiers: reasoning.effort supports low, medium, high, xhigh, and max. Gateways often default to low. Raise the tier for hard cases; do not send every request at max.
• Specs: about 1.05 million tokens of context and up to 128,000 output tokens. Very long prompts trigger the surcharge in the next section.
• Structured features: streaming, function calling, and Structured Outputs work; fine-tuning is not supported.
• Responses API tools: Web search, File search, image generation, Code interpreter, Hosted shell, Apply patch, Skills, Computer Use, MCP, Tool search.

For iOS / macOS developers, the real news is that Computer Use + Hosted shell + Apply patch can edit files, run commands, and inspect results in one session. Once the model is strong enough, the bottleneck becomes whether the execution environment falls asleep mid-run—close the laptop lid and the agent dies.

Self-reported scores OpenAI published (not independently re-run under one harness): Terminal-Bench 4 about 57.7, DeepSWE about 74.1, GPQA about 96.0, BrowseComp about 91.5. Treat them as a direction signal, not contract acceptance criteria. Same rule as Anthropic’s coding tools ranking: a benchmark says “it can chew a harder repo,” not “your monorepo merges overnight.”

How it maps to GPT-5.6 and GPT-6 Pro

The September wave is easy to mix up. Astra is the underlying flagship; GPT-6 Pro is the product name on the Chat surface.

If a Plus user cannot find Astra in the regular chat box, that does not mean it is closed. OpenAI put Plus Astra in Work (get things done) and Codex (write code), and reserved GPT-6 Pro in standard Chat for higher plans. Enterprise workspaces start with it off; an admin must open model access.

API: model ID, endpoints, and calling notes

The stable ID to put in config is gpt-6-astra. Snapshot aliases currently point at the same string. When you need a pinned version, read the Snapshots table on the model page. Do not invent your own date suffix.

Documented primary endpoints include Chat Completions (v1/chat/completions) and Responses (v1/responses). Tool loops, Computer Use, and MCP go through Responses. Free-tier RPM/TPM is “unsupported.” Quotas start at Tier 1 and rise with usage tiers.

const response = await client.responses.create({
  model: "gpt-6-astra",
  reasoning: { effort: "high" },
  input: "在仓库里定位支付超时，给出最小修复 diff，并说明验证命令。"
});

Production advice: Sol for the default loop, Astra for the upgrade. Trigger on eval scores or human spot-checks. Do not fire max because a user typed one sentence. Model swaps must share one harness—same principle as the LLM pricing and performance guide.

Pricing: standard, cache, long context, Fast / Batch

These are the USD per million tokens numbers on OpenAI’s pricing page as of writing. Tool calls (search, Computer Use, and the rest) are extra.

Rough math: the same 1M input + 100K output is about $6 on the Sol promo, about $15 on standard Astra, and about $30 on Fast Astra. Agent loops multiply by rounds. For a fuller “API vs cloud host vs local” frame, see the recent AI Agent compute cost comparison.

How to use it in ChatGPT (by plan)

The most common launch-day confusion: “Officially Plus can use it—why can’t I pick it in Chat?” The answer is a different surface, not a broken account.

Plus: use Work and Codex

1. Open the ChatGPT desktop app (web / mobile Work runs in the cloud; Codex is mainly desktop).
2. Switch from Chat to Work in the top-left, or open the Codex view.
3. In that surface’s model switcher, choose GPT-6 Astra.
4. CLI users should upgrade Codex CLI to 0.153.0 or newer and update to the latest Desktop. Older CLIs hide Astra entirely.

Work and Codex quotas are counted separately from regular Chat. Long jobs bill by “how much work got done,” not “how many messages you sent.” When quota runs out, switch to Sol / Terra or wait for reset. Switching surfaces is not an unlimited refill. For Codex weekly-limit mechanics, see what to do when the Codex weekly limit is exhausted.

Pro / Business / Enterprise

GPT-6 Pro (powered by Astra) is in standard Chat for Pro, Business, and Enterprise. Business Pro message quota is shared with GPT-5.6 Sol Pro; switching back and forth does not grant a second pool. Enterprise starts disabled; a workspace admin must turn on model access.

Free and Go: standard Chat stays on GPT-5.6 Luna as the front line. There is no Astra / GPT-6 Pro.

When to switch to Astra

• Stay on Sol: daily pairing, short diffs, high-frequency tool loops, cost-sensitive batch extraction. Promo pricing runs at least through November 21. It is still the default workhorse.
• Step up to Astra: cross-module refactors, let the model run tests and then patch, Computer Use on the desktop, swallow a long document in one pass, research-style browsing. Budget in engineer-days, not chat-message counts.
• Do not upgrade for the name: a Plus user who copies the same sentence from Chat into Work does not automatically get smarter—they just burn quota faster.
• Fix the harness when you compare Claude / Gemini: same repo, same tests, same timeout. Do not let model evals fight daily development for the same machine.

Why long-running jobs belong on Cloud Mac

Astra’s Computer Use, Hosted shell, and long Codex sessions are as hard on the runtime as Fable 5: the session cannot drop, you need a real macOS toolchain, disk and memory have to be enough, and it is best if CI lives on the same box.

• Stay awake: hour-scale and overnight jobs cannot survive laptop sleep. Desktop Work may still sync context to the cloud, but xcodebuild and the simulator can only be accepted on a Mac.
• Real toolchain: after Apply patch edits Swift, tests must run on the same machine. A GitHub-hosted macOS cold start stretches the agent feedback loop.
• Side-by-side evals: Astra vs Sol vs Fable 5 needs the same scripts in parallel. Local GPU/RAM folklore blows up first.

The model ID may change again next year. The host and workflow do not need a rewrite. Put Codex / Agent loops on Vuncloud Cloud Mac and keep the laptop for reviewing diffs—that is the shared physical layer of 2026 flagship models, not an add-on for one vendor.

FAQ

What is GPT-6 Astra?

OpenAI’s flagship reasoning model, announced on September 3, 2026. It targets hard reasoning, coding, Computer Use, research, and long documents. The API ID is gpt-6-astra.

When was it released?

Announced on September 3; API live on September 4. ChatGPT plans roll out by surface. Trust the help center and the in-product picker.

How much does the API cost?

Standard short context: $10 input / $1 cached input / $12.50 cache write / $50 output (per million tokens). Input over 272K bills the whole request at long-context rates. Batch/Flex is half price; Fast is double. The free API tier is not supported.

Can Plus use it?

Yes, but in Work and Codex, not the regular Chat picker. GPT-6 Pro in standard Chat is for higher plans.

Are Astra and GPT-6 Pro the same?

Same flagship underneath. Chat shows GPT-6 Pro; Work, Codex, and the API use Astra / gpt-6-astra.

How long is the context?

About 1.05 million tokens of context, up to 128,000 output tokens; knowledge cutoff 2026-04-30. Input over 272K costs more.

Closing

Three sentences: Astra is OpenAI’s September 2026 flagship; you pay for the long tool loop, not a chat voice; Plus users go to Work / Codex and should not wait in regular Chat. Keep the default route on Sol and step up for hard cases. Price the 272K switch and the Fast multiplier first, then decide whether a repo-scale job should run overnight.

Models will keep changing names. What you actually need is an execution environment that stays up, and an upgrade rule written in plain language.

Further reading

• How to buy AI Agent compute more cheaply in 2026: cloud, API, and local comparison
• LLM pricing and performance guide
• Codex weekly limit exhausted: what to do and alternatives
• Claude Fable 5 vs Opus 4.8: long-running Agent upgrade breakdown
• 2026 AI coding tools ranking