OpenAI 發布 GPT-6 Sol 與 Luna:API 定價降 50%,效能基準公開

2026年9月23日 05:18
站內 AI 整理稿

OpenAI has released GPT-6 Sol and GPT-6 Luna, 2 new models in its GPT-6 family.They sit below GPT-6 Astra, which launched earlier this month.OpenAI trained both with methods similar to Astra’s.The aim is to bring Astra’s advances to faster, more affordable models.Deployable today?Yes.

Both models are live in the OpenAI API as gpt-6-sol and gpt-6-luna.They are API-only models, so there are no weights to self-host.Three tiers, one recipe The GPT-6 family now has 3 tiers.Astra is the top model for the hardest work.Sol targets complex coding and professional tasks at lower cost.

Luna targets fast, high-volume everyday work.OpenAI team states better caching and inference let it serve these models more cheaply.It is cutting Sol and Luna API prices by 50% against their GPT-5.6 promotional pricing.ModelInput (per 1M tokens)Output (per 1M tokens)GPT-6 Astra$10.00$50.

00GPT-6 Sol$2.00 (was $4)$10.00 (was $20)GPT-6 Luna$0.10 (was $0.20)$0.50 (was $1.20) One detail is worth noting.Luna’s output price falls from $1.20 to $0.50, a cut of about 58%, not 50%.Benchmarks: what OpenAI reports Professional work: On AutomationBench 1.0.6, Sol at xhigh effort scores 33.

2% at $0.27 per task.Claude Opus 5 at max effort scores 26.9% at 11.1x that cost.Low-effort Astra scores 30.3% at 3.9x Sol’s cost.Luna at high effort gains 5.4 points over its predecessor at 58% lower cost per task.On Agents’ Last Exam, Sol at max effort scores 56.4%.

That beats Claude Opus 5’s best score at 60% lower cost per task.Coding: On DeepSWE v1.1, Sol at max effort scores 68.8%.That is 1.1 points behind Claude Fable 5 at xhigh, at about 80% lower cost per task.Luna at max effort scores 66.6%, comparable to Opus 5 and Fable 5 at medium effort.

In those comparisons, Luna costs 93% less per task than Opus 5 and 96% less than Fable 5.On FrontierCode 1.1 Main, which grades whether code is ready to merge, Sol matches Claude Fable 5.1 at xhigh at much lower cost.Computer use: On OSWorld 2.0 offline, Sol at xhigh scores 60.5% versus 60.

3% for Opus 5 at medium.Sol’s cost per task is about 80% lower.Luna at max beats GPT-5.6 Sol at medium for 1/10 of the cost.Factuality: OpenAI’s internal test uses de-identified ChatGPT conversations where users flagged model errors.Sol makes about half as many mistakes as its predecessor.

Luna at higher effort matches GPT-5.6 Sol at about 1/100 of its cost.OpenAI also carried Astra’s communication style over.Expect clearer, slightly shorter answers with less jargon, especially in coding conversations.

Prompt caching for long-running agents Agents resend the same instructions, tools and history on every turn.GPT-6 ships an improved prompt caching system with higher cache hit rates by default.Cached input reads get discounts of up to 90%.

Eligible shared prefixes reused within a 30-minute window now qualify.New controls for developers: A Prompt Caching Dashboard tracks hit rates over time.A diagnostics tool explains misses, for example "reason": "toolschanged".Explicit breakpoints let you choose where a cached prefix ends.

Reasoning effort can change mid-conversation via configurationupdate without breaking cache.allowed_tools restricts callable tools while keeping definitions stable.Prewarming prepares known context before the first user request.The full prompt caching guide covers each pattern.

GitHub reports these changes cut the share of prompt tokens needing fresh processing by more than 50%, helping Copilot respond faster.Availability API: gpt-6-sol and gpt-6-luna.ChatGPT Work and Codex: Plus, Pro, Business, Enterprise and Edu users.Free and Go: Luna in the ChatGPT desktop app.

Not yet in Chat.The ChatGPT rollout is gradual through launch day.Interactive Explainer window.addEventListener('message',function(e){var d=e.data;if(d&&d.mtpEmbed==='gpt6-sol-luna'&&d.height){var f=document.getElementById('mtp-g6-frame');if(f&&e.source===f.contentWindow){f.style.height=d.

height+'px';}}}); Key Takeaways Sol costs $2/$10 and Luna $0.10/$0.50 per 1M tokens.Sol at xhigh beats Opus 5 max on AutomationBench at 9% of the cost.Luna scores 66.6% on DeepSWE v1.1, costing 93% less per task than Opus 5.Cached input reads get up to 90% off, with new cache controls.

Both are live in the API, ChatGPT Work and Codex; not yet in Chat.Check out the Technical Blog.All credit goes to the researcher of this project.Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter.Wait!are you on telegram?

now you can join us on telegram as well.Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.?Connect with us The post OpenAI Releases GPT-6 Sol and Luna: 50% Cheaper API Pricing and Benchmarks appeared first on MarkTechPost.

Related

相關文章

量子位生成式AI

DeepSeek新論文公開Agent訓練!梁文鋒署名

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400"> DeepSeek新論文公開Agent訓練!梁文鋒署名 克雷西 2026-09-23 15:29:50 來源:量子位 每秒能產生5000+個沙盒 克雷西 發自 凹非寺 量子位 | 公眾號QbitAI 大模型訓練拼的是算力,Agent訓練拼的是環境。 環境怎麼造?梁文鋒署名的DeepSeek最新論文,把技術細節公開了。 DeepSeek做的這個系統叫DSec(DeepSeek Elastic Compute),乾的事情就是給Agent訓練批量製造沙盒。 它每秒能產生5000+個沙盒,一天能達到300萬個,峰值同時運行38萬個。 支撐這個規模的單集群也非常龐大,大約有160個節點、3萬核CPU和250TB內存。 為啥訓個Agent會這麼費勁? 因為大模型訓練的環境就是GPU集群,喂數據算梯度,但Agent完全不同。 它得在沙盒裡寫代碼、跑編譯、開瀏覽器,甚至裝操作系統……每執行一步都改變環境狀態,隨時可能把環境搞崩。 所以每輪訓練都得給它一個全新的、乾淨的沙盒,而且隨用隨拋、訓完就扔。 所以,問題兜兜轉轉,還是回到了基礎設施—— 這些基礎設施需要在每秒5000個的速度下,給每個沙盒裝好一整套操作系統和工具鏈。 同時,還不能讓幾十萬個併發沙盒把集群的內存和CPU擠爆。 具體怎麼辦,論文把這整套工程的全貌攤開了。 Agent訓練需要「一個世界」 DSec要解決的第一個核心問題是,不同類型的Agent任務對沙盒環境的要求差異極大,而且這些環境必須在同一個平臺上統一調度。 一個刷OJ題的Agent,只需要一個無狀態的函數調用環境,跑完拿到輸出就行,連文件系統都不

剛剛
量子位生成式AI

Qwen一號位定了!劉大一恆接棒

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400"> Qwen一號位定了!劉大一恆接棒 鷺羽 2026-09-23 15:27:49 來源:量子位 與稚暉君同屆天才少年 鷺羽 發自 凹非寺 量子位 | 公眾號QbitAI 定了!林俊暘離職半年,新任千問負責人終於就位—— 前華為天才少年劉大一恆。 阿里雲棲大會已經在官網嘉賓一欄中,將他掛在了C位。 緊隨蔡崇信、吳泳銘之後,title是: 阿里巴巴ATH事業群Token Foundry Qwen LLM項目負責人。 開幕式上,他也剛剛在吳泳銘主旨演講之後,帶來15分鐘的主題演講,標題是「Qwen:邁向真實世界智能體」。 重視程度不言而喻。。。 so為什麼是他? 這位新帥的來頭,可能比外界預想的更有料。 前華為天才少年接棒 劉大一恆,四川自貢人,本碩博均就讀於四川大學計算機學院。 師從呂建成教授,研究方向集中在自然語言生成與理解。 曾在ACL、ICML、NeurIPS、EMNLP、AAAI等頂會及其它期刊發表論文共計百餘篇,谷歌學術引用超六萬次,並以通訊作者身份獲得了NeurIPS 2025最佳論文獎。 另一位通訊作者正是前任負責人林俊暘。 劉大一恆同時也是Qwen初期核心成員。 2021年加入達摩院後,負責Qwen項目最早期的預訓練工作。 隨後幾年更是高產不斷: 主導開發Qwen1至Qwen3.5系列語言模型及Qwen-Math系列,核心參與Qwen-Coder、VL、Omni等模型開發,累計參與300多款Qwen模型開源。 換句話說,Qwen從早期研發到持續迭代,整個發展過程他其實一直都在。 只不過這一次,他終於正式走向臺前。 事實上,真正讓劉大一恆首

剛剛
量子位生成式AI

Jev vs Decitron:同為決策AI,為什麼不是一回事?

它來自TypeSafe AI。這支有OpenAI背景的團隊沒有繼續卷生成和推理,而是換了個方向:讓AI直接做判斷。他們甚至把口號直接寫成:Decisions, not strings(要決策,不要文本)。Jev的走紅,也讓“Decision”(決策)重新成為AI圈的熱門詞。

剛剛