11-19 更新 实时
首页/ 技术栈/ 正文

GPT-6.2 Astra 便宜20%-600%,评价不一

技术栈 来源:Latent Space 2026-11-18 18:19 0 阅读 机器翻译
GPT-6.2 Astra 便宜20%-600%,评价不一

OpenAI 今天发布了 GPT-6.2 Astra 和 Sol,比 Anthropic 预计的首个交易日早两天。在 Artificial Analysis 的测试中,Sol 与 9 月的 Astra 打成平手,每任务成本约为其 15%。新的 Astra 在 Intelligence Index 上得分 62,仅次于得分 65 的 Fable 5.2。归根结底,Anthropic 仍拥有全球最好的模型,但就每美元智能(intelligence per $)而言,OpenAI 拥有全球最高效的模型。

这次发布更不寻常的一点是,意见领袖的反馈极其负面……

这引发了防御性的回应……

我们认为,其中一部分是反应过度,也是意见领袖的惯常倾向——要么什么都 🤯🤯🤯,要么就是字面意义上的史上最差模型;但同时也认为部分批评是有道理的,RL 留下的痕迹很明显,甚至可以说是相当怪异(更别提那个糟糕的发布视频了)。不过,我们仍然用它来自动化我们的工作,而这归根结底才是最重要的。

发布很出色,但 Anthropic 的 IPO 不受影响,而这很可能成为本周/今年/这十年的头条新闻(直到 OpenAI IPO……)

今天的 AI 新闻由 SerpApi 呈现!通过 URL 抓取公开网页,输出适合 LLM 使用的 Markdown 或结构化 JSON。用一个 API 即可对网站进行映射(Map)并运行有界爬取(bounded crawl)。可与 100 多个 SerpApi 搜索 API 配合,用于搜索到内容的工作流。

GPT-6.2 Astra 和 Sol

可用性。今天的模型正在向 Plus、Pro、Enterprise 和 Edu 的 ChatGPT Work 与 Codex 推送,并通过 API 以 gpt-6.2-astra 和 gpt-6.2-sol 的形式提供给开发者。OpenAI

每百万 Token 的 API 价格。Astra 为输入 $8、缓存输入 $0.40、输出 $40。Sol 沿用前代定价,即 $2、$0.10 和 $10。两者的上下文窗口均为 1M,输出上限为 128K。Vals · Ultrafast 定价与发布

面向法律领域的 Astra。Enterprise 配置新增了法律检索索引。Artificial Analysis 在 Harvey LAB-AA 上测得 20.8%,而基础版 Astra 为 15.0%。Artificial Analysis

准备度。OpenAI 将 Astra 在 AI 自我改进方面评为 High,这是其对外发布的模型中首次达到该级别。在另一项指标上,Vals 的 RSI-Index 从 27.07% 升至 29.01%,其中 50% 代表其人类参照基准。System card · Vals

FrontierMath Erdős。Epoch 在 68 道题中验证了 7 个 Lean 解答。此前最好成绩为三个。Epoch

这些基准测试衡量的是什么

Agents’ Last Exam 1.1。OpenAI 报告称,Astra 在 xhigh 档位相比 GPT-6 Astra 在 max 档位提升 5.2 分,而每任务成本不到后者的三分之一。报告还称,在该测试中其以约 5% 的成本领先 Fable 5.2 达 5.5 分。在 GDPval v2 上,报告称 Astra 的输出 Token 用量比前代少 22%。OpenAI 的结果

长时工程任务。Mercor 的 OpenWorldEngineer 中,Astra 为 67.2%,Fable 5.2 为 54.6%,GPT-6 Astra 为 44.2%。在 Vals 的多轮 Vibe Code Bench 上,Astra 位列第一,Sol 位列第三。Mercor · Vals

3D 与机器人。Astra 以 1,318 分领跑 AA-3D,高于 Fable 的 1,271 和 Sol 的 1,259。在 Artificial Analysis 的机器人预览中,Astra 得分 79,Sol 得分 78,高于 Fable 的 75。Artificial Analysis

研究型智能体。在 Perplexity 的 WANDR 上,Astra 得分 0.783,每项任务成本约 12.54 美元。Sol 达到 0.746,成本约 2.65 美元。Perplexity

早期测试者并不认同

Matt Shumer 在测试数周后称这些模型“基本不可用”,并表示他的 Opus 和 Fable 智能体拒绝了 Astra 子智能体的许多工作。评测 · 工作流详情

Pietro Schirano 将他早期的许多问题归因于测试框架(harness)。Arena 的 Peter Gostev 在可视化编程上让 Astra 与 Fable 对比测试,并特别点出了 Astra 的 SVG 孔雀。Schirano · Gostev

演示。Gabriel Chua 使用 100 多个 Codex 线程构建了一个可行走漫游的新加坡。Simon Willison 用同一个 three.js 机甲提示词测试了多个模型,包括 Astra、Sol、Carbon 和 Fable。新加坡构建 · 同一提示词对比

Microsoft Ignite 与 Gemini

微软发布自研编程模型。MAI Code 2 面向多步编程,Code 2 Flash 面向更快的编辑。两者均可在 GitHub Copilot 的模型选择器、Copilot Fusion 和 Foundry 中使用。微软还宣布其 Decision 模型进行多模态升级。微软

Copilot Fusion。微软称其多模型框架在 Terminal-Bench 4.1 上达到 71.2%,在该对比中成本比 Claude Fable 5.2 低 88%。Jay Parikh

图像编辑。Artificial Analysis 在其编辑基准测试中将 MAI Image 2.7 排在首位,价格为每 1,000 张图像 34.82 美元。Artificial Analysis

Gemini Tasks。Hydrogen 现已在六个国家为免费 Tasks 提供支持,其中包括需要打电话或发短信的任务。Google 还展示了一款新的 Antigravity 机器人插件,配备 MuJoCo 仿真。Sundar Pichai · Google DeepMind

智能体框架与推理

Hugging Face 的 Pi 适配器。一个在 pi 内部训练的小型适配器,将 Qwen3.8-27B 在 51 个最难的 SWE-bench Pro 任务上的解出数量从 15 提升到 32。在该子集上,它每个任务的工具调用次数减少了 55%。模型卡与配对结果

智能体框架的成本差异悬殊。Baseten 在其涵盖 22 个框架和五个开放模型的基准测试中测得,每个通过任务的成本差距最高达 4.4 倍。Baseten

智能体框架可能会针对其所搭配的模型特化。Google 的 SIM 保留了以往框架编辑的理由和结果。它在 Opus 上演化出的框架在 Gemini 上得分 68.41,而起始框架为 73.89。研究摘要

更多多智能体产品。Cognition 推出了具有持久性的 Devin Bots。Perplexity 报告称,其混合模型团队在 Terminal-Bench 和 WANDR 上击败了领先模型的四个实例。Replit 报告称,在 NumPy 任务上,智能体数量从 16 增加到 32 时出现收益递减。Cognition · Perplexity · Replit

协调问题主导了一次 1,000 个智能体的运行。Hugging Face 表示,任一时刻只有 40–60 个智能体在调用模型。它们每写出一个 Token,就要读取约 130 个 Token,而合并与测试也是瓶颈的成因之一。Thomas Wolf

一小部分缓存未命中就可能占据账单的大头。SemiAnalysis 报告称,在其样本中,1.8% 的未缓存请求贡献了约 15% 的支出。SemiAnalysis

Marin 和 K2 Horizon 公开了各自的训练过程。Marin 发布了其 535B-A23B 训练中的 57 个中间 Mochi 检查点。K2 的报告比较了在相同 22T Token 上训练的四种模型规模,并描述了一个解码适配器,团队称其实现了 2.2× 加速。Marin · K2 Horizon

Zyphra 的长上下文注意力。团队报告称,在将多个模型转换为其亚二次注意力后,在 10M Token 上实现了 9–60× 的加速,而 Qwen 在 1M 长度下于 RULER 上的得分下降不到两个点。Zyphra

推测解码。vLLM 报告称,针对 Kimi K3.1 Pro 采用滑动窗口草稿模型,接受长度最高提升 30%。NVIDIA 也为 Qwen3.8-27B 准备了一个 4B 的草稿模型。vLLM · 草稿模型研究

产业与政策

NVIDIA 在 CEO 交接后的首份财报。在临时 CEO Colette Kress 的带领下,NVIDIA 公布季度营收 1158 亿美元,同比增长 103%,并表示 Vera Rubin NVL72 已出货。董事会表示不打算让 Jensen Huang 复职。NVIDIA · AP

Isomorphic Labs 融资 50 亿美元。Greenoaks 和 Sequoia 共同领投 C 轮。该公司表示将扩大其药物设计平台的规模,并推进其治疗管线。Isomorphic Labs

Genesis AI 融资 5 亿美元。这轮 A 轮融资对该机器人公司的投后估值为 35 亿美元。公司计划用这笔资金部署其 Eno 机器人,并扩大训练与数据采集。Zhou Xian

本地模型与硬件

一个把部分层跑两遍的 9B 模型。SenseNova-Loop-9B 将其 32 层中的 24 层再运行一次。模型作者报告 GAIA 从预览版的 37.86 升至正式版的 69.90。本地用户询问,在长上下文下还能为 KV 缓存留下多少内存。权重 · 讨论

Maxwell V1 Flash。这个 35B/激活 3B 的混合专家(MoE)模型声称,在 256K 下注意力相比稠密 GQA 提速 6.7×。其稀疏注意力仍保留全历史 KV 缓存,而这正是本地用户在投入硬件之前希望得到实测的部分。模型卡 · 讨论

Olala mini 7A1B。OVHcloud 预览了一个小型 MoE,激活参数 1B,上下文窗口 32K。该讨论帖仍在等待权重和可复现的基准测试。OVHcloud · 讨论

Intel 的 Crescent Island 数据。早期简报中那个吸睛的吞吐对比使用的是 FP64 密集型负载。Intel 另一份 LLM 对比则声称,相较 RTX PRO 6000 Blackwell Server Edition,吞吐为其 1.12×、延迟低 29%。硬件讨论

在打印机上跑 Gemma。一位开发者报告,通过一个 PostScript 打印任务运行了 Gemma 3 270M。演示

智能体的实际应用

一个为实时输入打造的电子游戏智能体。miHoYo 的 Gamer 以 5 Hz 采样屏幕,并以 30 Hz 输出键盘和鼠标操作。在作者的评估中,其 RL 版本解出了留出的崩坏解谜任务中的约一半。项目 · 讨论

一个虚构的联盟每晚进行直播。开发者表示,Fock Ball 在物理引擎中模拟比赛,并使用 Claude 完成解说、新闻编辑室等周边工作。开发者帖子

ICLR 评审与 AI 检测器。审稿人报告称,程序主席标记了疑似由 AI 生成的评审意见。该讨论帖提出疑问:检测器如何区分编辑辅助与整篇评审的生成。同行评审讨论

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文([AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse)

OpenAI shipped GPT-6.2 Astra and Sol today, two days before Anthropic’s expected first trading day. In Artificial Analysis’ tests, Sol ties September’s Astra at about 15% of the cost per task. New Astra scores 62 on the Intelligence Index, just behind Fable 5.2 at 65. Net net, Anthropic still owns the world’s best model, but OpenAI owns the world’s most efficient models in terms of intelligence per $.

What’s more unusual about this launch is the STRONGLY negative influencer feedback…

which prompted a defensive response…

We think that some of this is overreacting and normal influencer tendency to either have everything be 🤯🤯🤯 or the literal WORST.MODEL.EVER, BUT also that some of the criticisms are valid, and the RL artefacts are noticeable and downright bizarre (don’t even get us started about the terrible launch video). However, we still use it to automate our work and that’s all that matters in the end.

Great ship, but Anthropic’s IPO is unaffected and that will be probably the headline of the week/year/decade (until OpenAI IPO’s….)

Today’s AI News is brought to you by SerpApi!

Retrieve public web pages by URL as LLM-ready Markdown or structured JSON. Map sites and run bounded crawls with one API. Pair with 100+ SerpApi search APIs for search-to-content workflows.

Check them out to support LS, and contact business@latent.space for feedback!

AI Twitter Recap

GPT-6.2 Astra and Sol

  • Availability. Today’s models are rolling out to ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and to developers through the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI

  • API prices per million tokens. Astra is $8 input, $0.40 cached input and $40 output. Sol keeps its predecessor’s pricing at $2, $0.10 and $10. Both have 1M context and a 128K output limit. Vals · Ultrafast pricing and rollout

  • Astra for Law. The Enterprise configuration adds a legal search index. Artificial Analysis measured 20.8% on Harvey LAB-AA, versus 15.0% for base Astra. Artificial Analysis

  • Preparedness. OpenAI rated Astra High for AI Self-Improvement, a first for one of its externally released models. On a separate measure, Vals’ RSI-Index rose from 27.07% to 29.01%, with 50% denoting its human reference. System card · Vals

  • FrontierMath Erdős. Epoch verified seven solutions in Lean out of 68 problems. The previous best was three. Epoch

What the benchmarks measure

  • Agents’ Last Exam 1.1. OpenAI reports Astra at xhigh gained 5.2 points over GPT-6 Astra at max, using less than a third of the cost per task. It reports a 5.5-point lead over Fable 5.2 at about 5% of the cost on that test. On GDPval v2, it says Astra used 22% fewer output tokens than its predecessor. OpenAI’s results

  • Long engineering runs. Mercor’s OpenWorldEngineer puts Astra at 67.2%, Fable 5.2 at 54.6%, and GPT-6 Astra at 44.2%. On Vals’ multi-turn Vibe Code Bench, Astra was first and Sol was third. Mercor · Vals

  • 3D and robotics. Astra led AA-3D with 1,318, ahead of Fable at 1,271 and Sol at 1,259. On Artificial Analysis’s robotics preview, Astra scored 79 and Sol 78, ahead of Fable at 75. Artificial Analysis

  • Research agents. On Perplexity’s WANDR, Astra scored 0.783 at about $12.54 per task. Sol reached 0.746 at about $2.65. Perplexity

Early testers disagreed

  • Matt Shumer called the models “essentially unusable” after weeks of testing and said his Opus and Fable agents rejected much of Astra’s subagent work. Review · Workflow details

  • Pietro Schirano attributed many of his early issues to the harness. Arena’s Peter Gostev tested Astra against Fable on visual coding and singled out Astra’s SVG peacock. Schirano · Gostev

  • The demos. Gabriel Chua built a walkable Singapore using more than 100 Codex threads. Simon Willison ran the same three.js mech prompt through several models, including Astra, Sol, Carbon and Fable. Singapore build · Same-prompt comparison

Microsoft Ignite and Gemini

  • Microsoft ships its own coding models. MAI Code 2 is for multi-step coding, and Code 2 Flash is for quicker edits. Both are available in GitHub Copilot’s model picker, Copilot Fusion and Foundry. Microsoft also announced a multimodal upgrade to its Decision model. Microsoft

  • Copilot Fusion. Microsoft reports 71.2% on Terminal-Bench 4.1 with its multi-model harness, at 88% lower cost than Claude Fable 5.2 in that comparison. Jay Parikh

  • Image editing. Artificial Analysis ranks MAI Image 2.7 first on its editing benchmark, at $34.82 per 1,000 images. Artificial Analysis

  • Gemini Tasks. Hydrogen now powers free Tasks in six countries, including tasks that need a phone call or a text. Google also showed a new Antigravity robotics plugin with MuJoCo simulation. Sundar Pichai · Google DeepMind

Agent harnesses and inference

Hugging Face’s Pi adapter. A small adapter trained inside pi lifted Qwen3.8-27B from 15 to 32 solves on the 51 hardest SWE-bench Pro tasks. Across that subset it made 55% fewer tool calls per task. Model card and paired results

  • Harness costs vary sharply. Baseten measured up to a 4.4× spread in cost per passing task across its benchmark of 22 harnesses and five open models. Baseten

  • The harness may specialize to its model. Google’s SIM keeps the rationale and outcome of prior harness edits. A harness it evolved on Opus scored 68.41 on Gemini, versus 73.89 for the starting harness. Research summary

  • More multi-agent products. Cognition launched persistent Devin Bots. Perplexity reported that its mixed-model teams beat four instances of the lead model on Terminal-Bench and WANDR. Replit reported diminishing returns from 16 to 32 agents on NumPy tasks. Cognition · Perplexity · Replit

  • Coordination dominated a 1,000-agent run. Hugging Face says only 40–60 agents were calling the model at any moment. They read roughly 130 tokens for every token they wrote, with merges and tests contributing to the bottleneck. Thomas Wolf

  • A small fraction of cache misses can dominate the bill. SemiAnalysis reports that 1.8% of uncached requests accounted for roughly 15% of spend in its sample. SemiAnalysis

  • Marin and K2 Horizon opened up their training runs. Marin released 57 intermediate Mochi checkpoints from its 535B-A23B run. The K2 report compares four model sizes trained on the same 22T tokens and describes a decoding adapter that the team says reaches 2.2× speedup. Marin · K2 Horizon

  • Zyphra’s long-context attention. The team reports 9–60× speedups at 10M tokens after converting several models to its subquadratic attention, with Qwen losing under two points on RULER at 1M. Zyphra

  • Speculative decoding. vLLM reports up to 30% higher acceptance length from a sliding-window draft for Kimi K3.1 Pro. NVIDIA also has a 4B draft for Qwen3.8-27B. vLLM · Draft-model research

Industry and policy

  • NVIDIA’s first earnings after the CEO transition. Under interim CEO Colette Kress, NVIDIA posted $115.8B in quarterly revenue, up 103% year over year, and said Vera Rubin NVL72 had shipped. The board said it did not intend to reinstate Jensen Huang. NVIDIA · AP

  • Isomorphic Labs raises $5B. Greenoaks and Sequoia co-led the Series C. The company says it will scale its drug-design platform and develop its therapeutic pipeline. Isomorphic Labs

  • Genesis AI raises $500M. The Series A values the robotics company at $3.5B post-money. It plans to use the funding to deploy its Eno robot and expand training and data collection. Zhou Xian


AI Reddit Recap

Local models and hardware

  • A 9B model with layers that run twice. SenseNova-Loop-9B runs 24 of its 32 layers a second time. The model’s authors report GAIA rising from 37.86 in the preview to 69.90 in the release. Local users asked how much memory remains for the KV cache at long context. Weights · Discussion

  • Maxwell V1 Flash. The 35B/3B-active MoE claims a 6.7× attention speedup over dense GQA at 256K. Its sparse attention still retains the full-history KV cache, which is what local users want measured before they commit hardware. Model card · Discussion

  • Olala mini 7A1B. OVHcloud previewed a small MoE with 1B active parameters and a 32K context window. The thread is waiting for weights and reproducible benchmarks. OVHcloud · Discussion

  • Intel’s Crescent Island numbers. The early brief’s eye-catching throughput comparison uses FP64-heavy workloads. Intel’s separate LLM comparison claims 1.12× throughput and 29% lower latency versus RTX PRO 6000 Blackwell Server Edition. Hardware discussion

  • Gemma on a printer. A builder reports running Gemma 3 270M via a PostScript print job. Demo

Agents in use

  • A video-game agent built for real-time input. miHoYo’s Gamer samples the screen at 5 Hz and emits keyboard and mouse activity at 30 Hz. Its RL version solved about half of the held-out Honkai puzzle tasks in the authors’ evaluation. Project · Discussion

  • A fictional league runs a nightly broadcast. The builder says Fock Ball simulates the games in a physics engine and uses Claude for surrounding work such as commentary and the newsroom. Builder’s post

  • ICLR reviews and AI detectors. Reviewers report that the program chairs flagged likely AI-generated reviews. The thread asks how a detector can distinguish editing help from whole-review generation. Peer review discussion

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 Latent Space([AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗