OpenAI
GPT-6.2 Astra Pro
最大上下文
1.1M
最大輸出
128K
發布日期
暫無
gpt-6.2-astraGPT-6.2 Astra 是 OpenAI GPT-6.2 系列的旗艦模型,專為最艱難的專案打造。它適用於真實工程、專業工作與科學研究,並針對網站、簡報和 3D 設計提供更強的視覺推理能力。OpenAI 表示其評估成績創新高,且每次任務成本不到 GPT-6 Astra 的三分之一。
上下文視窗
1.1M Token
最大輸出
128K Token
發布日期
2026年11月17日
模態
| Token 階梯 | 輸入價格 | 輸出價格 | 快取讀取 | 快取建立 5m |
|---|---|---|---|---|
| <=272K | $8/M | $40/M | $0.4/M | $10/M |
| >272K | $16/M | $60/M | $0.8/M | $20/M |
推理
工具呼叫
Temperature 參數
附件
知識庫
Endpoint 協議
GPT-6.2 Astra 在複雜工程與專業工作中的優勢。
適合需要擬定計畫、檢查階段性成果並逐步修訂的複雜任務。
能在現有專案中實作並測試橫跨多個檔案的相關修改。
能分析螢幕截圖與視覺參考圖,作為版面、字體排印及互動介面設計的依據。
105 萬 Token 的上下文視窗可同時容納多份來源文件,並為深入分析保留空間。
運用 GPT-6.2 Astra 改造程式碼庫、整理多份文件或製作瀏覽器原型。
盤點呼叫位置、遷移共用介面,並在提交審查前執行相關測試以找出回歸問題。
比較多份報告的結論,標示不一致的數據,整理成可追溯原文的摘要。
結合參考簡報或螢幕截圖,將專案摘要整理成符合既有設計、可供審閱的投影片。
依需求製作可點擊的介面或簡單的 3D 場景,再根據螢幕截圖和互動回饋調整。
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)了解上下文限制、支援的輸入、計費級距與 TokenHub API 串接方式。
它是 OpenAI 的 GPT-6.2 旗艦模型,適用於複雜軟體工程、多步推理,以及需要處理文件與視覺參考的專業工作。
OpenAI 公布的上下文視窗為 105 萬 Token,單次最多輸入 92.2 萬 Token、輸出 12.8 萬 Token。
可以。TokenHub 支援文字、圖片與 PDF 輸入,模型以文字作答。附件格式以所選端點的文件為準。
使用 TokenHub API 金鑰,並將 model 欄位設為 gpt-6.2-astra。API 串接區段提供所選端點的基礎 URL 與請求範例,包括 Responses 和 Chat Completions。
TokenHub 按每百萬 Token 列出輸入、輸出、快取讀取與快取寫入價格。輸入超過 27.2 萬 Token 時,適用價格表中的較高級距。
涉及跨模組程式碼變更、複雜視覺任務或綜合分析多份來源文件時,優先考慮 Astra。日常開發和可重複的工作流程可從成本較低的 Sol 開始。
透過一個 API Key 接入 GPT-6.2 Astra 與更多 AI 模型。
GPT-6.2 Astra 影片與社群討論
收錄 GPT-6.2 Astra 的發布貼文、實測示範與開發者討論。
X (Twitter)
OpenAI ·
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. https://openai.com/index/introducing-gpt-6-2-astra-sol https://t.co/jLCyVIws2Y
https://pbs.twimg.com/media/HXbQg3xacAAib6t.jpg查看原文OpenAI ·
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra. You’ll see the difference in things like slideshows, websites, and 3D designs. https://t.co/FcbBmZtbpF
https://pbs.twimg.com/media/HXWFd1_bIAAi6NF.pnghttps://pbs.twimg.com/media/HXWFfGPboAAcE96.pnghttps://pbs.twimg.com/media/HXWFgH8aUAAzQ0x.png查看原文OpenAI Developers ·
Upgrades, people. Upgrades. The wait is over, you can now use GPT-6.2 Astra and GPT-6.2 Sol in Codex and the API. https://t.co/Uz6JF5XAF1
https://pbs.twimg.com/media/HXWMJzlWQAA4bE2.jpg查看原文Reddit
Introducing GPT-6.2 Astra and Sol
czk_21 ·
查看原文Title: What do you think about GPT-6.2 Astra and GPT-6.2 Sol?
Most_Researcher_3010 ·
OpenAI has introduced GPT-6.2 Astra and GPT-6.2 Sol. Astra focuses on advanced reasoning and coding, while Sol focuses on everyday coding and Al workflows. I'm curious about one thing: Would you prefer an Al that is better at deep reasoning or one that helps you complete everyday tasks faster? Why?
查看原文Video roundup of Astra 6.2 reviews are in!
Ok_Improvement5078 ·
展開全文
I went ahead and put together a round up for all major reviews of Astra 6.2. Seems like most are negative. Positive Universe of AI: “GPT-6.2 Astra & Sol Are OUT! (So Why Is Everyone Upset?)” https://benchpress.me/e/gpt-6-2-astra-sol-are-out-so-why-is-everyone-upset-4168d904 The video leads with OpenAI’s best number: Astra beats Claude Fable 5.2 on Agents’ Last Exam at about 5% of the cost. From there it works through the launch-day backlash and asks whether it’s deserved. Arena: “GPT 6.2 vs Claude Fable 5.2: Astra & Sol First Impressions” https://benchpress.me/e/gpt-6-2-vs-claude-fable-5-2-astra-sol-first-impressions-676dcca7 Peter Gostev doesn’t pick an overall winner: Astra is precise and detailed at a competitive price, and Fable is more creative and varied in style. He calls Sol the “really big story,” since it gets close to frontier quality at a fifth to a tenth of the price. Negative Fun AI Builds With Mark: “GPT-6.2 Astra Left Me Speechless” https://benchpress.me/e/gpt-6-2-astra-left-me-speechless-309af1bf After building four games with Astra, Mark calls the results “shockingly bad” and the model “a bit of a joke.” His tactical shooter test had clunky reload animations, awkward hand rendering and a repetitive gameplay loop, and he puts Anthropic’s Opus and Sonnet “light years ahead.” Nerd Snipe: “Astra 6.2 Is A Step Back” https://benchpress.me/e/astra-6-2-is-a-step-back-76dcade1 Theo and Ben find Astra 6.2 nearly unusable: it has strong capabilities but is very hard to work with. Theo describes a 7-day run that landed zero improvements and made the compiler 56% slower, and he now switches to Opus 5.5 at the end of long runs just to get a readable summary. Every: “We Tested OpenAI’s GPT-6.2 Astra For 2 Weeks” https://benchpress.me/e/we-tested-openai-s-gpt-6-2-astra-for-2-weeks-8a964c63 The Every team finds that Astra and Sol handle bounded tasks efficiently, but harder work still needs more judgment and follow-through. Writing is a step back from Astra 6 and too compressed to follow, and in real coding tasks the model often stops too early. Nick Saraev: “GPT-6.2 Astra Just Dropped, And It’s Not Great” https://benchpress.me/e/gpt-6-2-astra-just-dropped-and-it-s-not-great-6cd1a5e1 Nick calls the release underwhelming but pushes back on the pile-on: “it’s a magical talking computer, doesn’t need to be perfect.” His message is to reset expectations rather than write the model off.
YouTube
GPT-6.2 Astra and GPT-6.2 Sol—Models built for the real world
OpenAI
查看原文Building 3D video sets with GPT-6.2 Astra
OpenAI
查看原文GPT 6.2 vs Claude Fable 5.2: Astra & Sol First Impressions
Arena
查看原文