GPT-6.2 Astra 和 Sol 现已在 ChatGPT Work 和 Codex 中向 Plus、Pro、Enterprise 和 Edu 用户开放。Pro-500 订阅者可以使用这两款模型的 Ultrafast(超快)模式。Pro-200 订阅者首次获得 Ultrafast 访问权限,从 GPT-6.2 Sol 开始。
开发者可以通过 OpenAI API 访问这两款模型。OpenAI 表示,GPT-6.2 Astra 的 API 定价比其前代产品便宜 20%,输入令牌为每百万 8 美元,缓存输入为 0.40 美元,输出令牌为每百万 40 美元。GPT-6.2 Sol 保持 GPT-6.1 Sol 的定价不变,即输入每百万令牌 2 美元,缓存输入 0.10 美元,输出 10 美元。
OpenAI 训练 GPT-6.2 Astra 以更可靠地处理法律、金融和会计等领域的专业惯例。该公司表示,演示文稿会从模板中提取布局、颜色和排版,并匹配确切的字体大小。幻灯片标题旨在捕捉幻灯片的主要观点,而正文则专注于关键细节。
OpenAI 的公告重点强调其模型的性能价格比。在 Agents' Last Exam(一项测试多步专业工作流程的基准测试)上,根据 OpenAI 的说法,GPT-6.2 Astra 得分比其前代产品 GPT-6 Astra 高出 5.2 分,而成本不到三分之一。它以仅 5% 的成本领先 Anthropic 的 Claude Fable 5.2 达 5.5 分。只有 Google 的 Gemini 4 Carbon 在这项基准测试中与 Astra 持平。
GPT-6.2 Astra 以较低的单任务成本在 Agents' Last Exam v1.1 中获得最高分。Claude Fable 5.2 达到类似的分数,但单任务成本高数倍。| 图片来源:OpenAI
在涵盖合同条款分析等任务的法律基准 Harvey LAB-AA 上,根据 OpenAI 的说法,GPT-6.2 Astra 的通过率几乎是 Claude Fable 5.2 的两倍,而成本约为其六分之一。在包含咨询、金融和设计任务的内部 GDPval-v2 基准测试中,OpenAI 表示 Astra 比其前代产品提高了 6.2 个百分点,同时输出令牌减少了 22%。
OpenAI 还声称在网络设计方面取得了进展。在 3D 基准 AA-3D 上,人类评委比较两个盲测的交互体验。GPT-6.2 Astra 以 1,318 Elo 点排名第一,领先于 Claude Fable 5.2 的 1,271 分和 GPT-6.2 Sol 的 1,259 分。
在 Terminal-Bench-Science(一项在终端中测试科学工作流的基准测试)上,根据 OpenAI 的说法,该模型得分比 GPT-6 Astra 高出 9 个百分点,成本同样不到三分之一。OpenAI 表示,Astra 还在 Vals Science Index 上创下新高,该指数衡量模型解释生物数据和运行实验的能力。
OpenAI 表示,GPT-6.2 Sol 是为希望拥有快速、多功能模型用于日常任务的用户设计的。在 FrontierCode 2.0 编码基准测试中(该测试检查代码更改是否达到真实项目的质量门槛),Sol 的成本仅为 Astra 的五分之一,得分仅落后 0.5 分。OpenAI 表示,该模型也能很好地处理非工作类任务。
在编码基准 FrontierCode 2.0 上,GPT-6.2 Sol 的得分几乎与 GPT-6.2 Astra 相当,但成本仅为其中的一小部分。Claude Fable 5.2 达到最高峰值分数,但单任务成本高数倍。| 图片来源:OpenAI
根据 OpenAI 的说法,GPT-6.2 Astra 和 Sol 在网络安全方面也表现得更好。它们能更有效地发现并利用漏洞,编写更安全的代码,并提供更好的错误修复,在多项测试中具有显著更高的令牌效率。这两款模型都在 OpenAI 的 Preparedness Framework(准备框架) 中达到了网络安全的“Critical”阈值,正如之前的 GPT-6 Astra 所做的那样。
作为回应,OpenAI 正在对潜在风险的网络安全评分请求增加更严格的保障措施。OpenAI 表示,这些保障措施应使滥用变得更加困难,但也限制了合法的漏洞研究和安全测试。
在内部 ExploitBench 测试中,模型必须将已知的浏览器漏洞转化为可行的利用工具。GPT-6.2 Astra 实现了 50% 的成功率,远远超过其前代产品 GPT-6 Astra。结果来自关闭安全措施的测试结果。| 图片来源:OpenAI
在 Artificial Analysis Intelligence Index(汇总了知识工作、编码、科学和一般知识方面的九项评估)中,GPT-6.2 Astra 排名第二,仅次于 Anthropic 的 Claude Fable 5.2。根据 Artificial Analysis 的数据,Astra 比 GPT-6 Astra 高出 5 分,超过了 Google 的 Gemini 4 Carbon(60 分)和 Anthropic 的 Claude Opus 5.5(59 分)。GPT-6.2 Sol 增加了 2 分,达到 57 分,相当于 GPT-6 Astra 的水平,但单任务成本仅为约 15%。
根据 Artificial Analysis 的说法,OpenAI 在成本与性能的帕累托前沿(Pareto frontier,即没有其他模型既更便宜又更好的界线)上占据主导地位。只有 Anthropic 和 Meta 的模型也位于这条前沿线上。OpenAI 也在效率方面占据了大部分顶级位置,这是通过每个任务的输出令牌来衡量的。Claude Fable 5.2 是唯一出现在那里的外部模型。
GPT-6.2 Astra 获得 62 分,在 Intelligence Index 中排名第二,仅次于 Claude Fable 5.2 (65)。OpenAI 模型在成本绩效的帕累托前沿上占据主导地位。| 图片来源:Artificial Analysis
Artificial Analysis 证实了 OpenAI 在科学和事实知识方面的说法。GPT-6.2 Astra 在 Terminal-Bench-Science 0.1 上领先,并在物理推理测试 CritPt v2、事实知识基准 AA-Omniscience 和知识基准 HLE-Diamond 上排名第二,仅次于 Claude Fable 5.2。Astra 还在 3D 设计 (AA-3D) 和机器人技术 (AA-Robotics v0.1) 方面领先,在机器人技术上得分为 79 分,高于 Sol 的 78 分和 Claude Fable 5.2 的 75 分。
然而,编码方面情况不同。根据 Artificial Analysis 的说法,这两款模型在此方面停滞不前。在 Codex harness 最大计算量下测量的 Coding Agent Index 中,GPT-6.2 Astra 保持在 GPT-6 Astra 的位置。GPT-6.2 Sol 实际上比 GPT-6.1 Sol 下降了 3 分。这一下降源于 SWE-Atlas-QnA-AA 上的 14 分下滑,在该测试中,模型比其前代产品更频繁地在答案中遗漏关键事实。
Artificial Analysis 还指出高幻觉率。在最大努力下,GPT-6.2 Astra 有 59% 的时间会产生幻觉,Sol 为 55%。该指标是 AA-Omniscience Hallucination Rate,衡量模型在事实问题上猜测而非承认不知道的频率。这些问题非常困难,因此这些数字并不反映日常幻觉率。
在来自 Epoch AI 自身研究的现实任务中,Astra 落后于 Fable 5.2,仅比 GPT-6 Astra 略有改善。根据 Epoch AI 的说法,该模型在较简单的研究项目中错过了核心问题,并试图修复事物却徒劳无功。在一个关于研究 AI 数据中心的任务中,Astra 过于保守,跳过了关键估算,Epoch AI 表示没有任何承包商可以逃脱这种指责。Epoch AI 认为 Astra 的写作能力相对于其他测试模型来说相对较强,但仍低于其自身的标准。
GPT-6.2 Astra 在 Epoch Capabilities Index 中得分为 174 分,与 Anthropic 的 Fable 5.2 持平。两者均位于推理模型先前趋势线的上方。Astra 的成本显著降低。| 图片来源:Epoch AI
值得注意的是,Astra 能够在单个输出令牌内运行复杂的推理,而没有可见的思维链。Epoch AI 记录 Astra 有 12.3 个串行步骤,高于 GPT-6 Astra 的 10.5 步和当时最强的非 OpenAI 模型 Claude Fable 5.1 的 4.2 步。能够以这种方式推理而不展示其工作的模型更难监控,这是一个安全问题,已在第一代 Astra 模型中出现。
在 FrontierMath Erdős(一项未解决的数学问题基准测试)上,GPT-6.2 Astra 解决了 68 个任务中的 7 个,每个任务都在证明助手 Lean 中正式验证。旧纪录是 GPT-6.1 Sol 创造的 3/68。这七个任务中有一个在此之前没有已知的解决方案。
GPT-6.2 Astra 和 Sol 在 Vals Index 中分别排名第 7 和第 9。Sol 每任务成本为 3.18 美元,不到 Astra 10.68 美元的三分之一,是前十名中最便宜的模型。Claude Fable 5.2 位居榜首。| 图片来源:Vals AI
Vals AI 支持 OpenAI 在 Vals Science Index 上排名第一的说法。Astra 在该排名中领先,并以大约一半的成本和一半的运行时间在 Terminal-Bench Science 和 VCB 1–100 上获得第一名。Sol 在 Terminal-Bench Science 上排名第三,成本约为 Astra 的四分之一。在更广泛的 agentic 编码基准 Terminal-Bench 4 中,Astra 排名第 5,Sol 排名第 7。
根据 Vals AI 的说法,法律测试是这些模型的短板。在 HLAB 基准测试中,Astra 和 Sol 将其前代产品的分数翻倍,但仅排名第 11 和第 21。在法律
查看原文
OpenAI releases two new language models. GPT-6.2 Astra is aimed at demanding professional work, while GPT-6.2 Sol is built for everyday use. OpenAI says both outperform their predecessors on benchmarks and cost less to run.
GPT-6.2 Astra and Sol are available now for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Pro-500 subscribers can use Ultrafast mode with both models. For the first time, Pro-200 subscribers also get Ultrafast access, starting with GPT-6.2 Sol.
Developers can access both models through the OpenAI API. OpenAI says API pricing for GPT-6.2 Astra is 20 percent cheaper than its predecessor's, at $8 per million input tokens, $0.40 for cached inputs, and $40 per million output tokens. GPT-6.2 Sol keeps GPT-6.1 Sol's pricing at $2 for input, $0.10 for cached inputs, and $10 for output per million tokens.
OpenAI calls Astra "the world's best model for scientists" and touts polished slides
OpenAI trained GPT-6.2 Astra to handle professional conventions in fields like law, finance, and accounting more reliably. The company says presentations pull layouts, colors, and typography from templates, matching exact font sizes. Slide titles are meant to capture a slide's main point, while body text stays focused on key details.
OpenAI's announcement leans hard on the price-to-performance ratio of its models. On Agents' Last Exam, a benchmark that tests multi-step professional workflows, GPT-6.2 Astra scores 5.2 points above its predecessor GPT-6 Astra at less than a third of the cost, according to OpenAI. It leads Anthropic's Claude Fable 5.2 by 5.5 points at just 5 percent of the cost. Only Google's Gemini 4 Carbon keeps up with Astra on this benchmark.

GPT-6.2 Astra achieves the highest scores on Agents' Last Exam v1.1 at low cost per task. Claude Fable 5.2 reaches a similar score but costs several times more per task. | Image: OpenAI
On the legal benchmark Harvey LAB-AA, which covers tasks like contract clause analysis, GPT-6.2 Astra more than doubles Claude Fable 5.2's pass rate at about one-sixth of the cost, according to OpenAI. On the internal GDPval-v2 benchmark, which includes consulting, finance, and design tasks, OpenAI says Astra gains 6.2 percentage points over its predecessor while using 22 percent fewer output tokens.
OpenAI says computer use, already a strong suit of the predecessor, got better again. In testing, the model reportedly built a 3D car model, wrote a physics solver in Python, and animated the results in Blender. In another demo, it modeled a remote-controlled car part by part and produced an animated assembly guide.
OpenAI also claims progress in web design. On the 3D benchmark AA-3D, human judges compare two blinded interactive experiences. GPT-6.2 Astra takes first with 1,318 Elo points, ahead of Claude Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
On Terminal-Bench-Science, a benchmark that tests scientific workflows in a terminal, the model scores 9 percentage points above GPT-6 Astra, again at less than a third of the cost, according to OpenAI. Astra also sets a new high on the Vals Science Index, which measures how well models interpret biological data and run experiments, OpenAI says.
Sol nearly matches the original Astra at a fraction of the cost
GPT-6.2 Sol is for users who want a fast, versatile model for everyday tasks, OpenAI says. On the coding benchmark FrontierCode 2.0, which checks whether code changes meet the quality bar of real projects, Sol trails Astra by just 0.5 points at one-fifth the cost. OpenAI says the model also handles non-work tasks well.

On the coding benchmark FrontierCode 2.0, GPT-6.2 Sol scores nearly as high as GPT-6.2 Astra but at a fraction of the cost. Claude Fable 5.2 hits the highest peak score but costs several times more per task. | Image: OpenAI
GPT-6.2 Astra and Sol have also gotten better at cybersecurity, according to OpenAI. They find and exploit vulnerabilities more effectively, write more secure code, and deliver better bug fixes, with notably higher token efficiency in several tests. Both models hit the "Critical" threshold for cybersecurity in OpenAI's Preparedness Framework, just as GPT-6 Astra did before them.
OpenAI is adding stricter safeguards for potentially risky cybersecurity requests in response. These safeguards should make abuse harder but will also limit legitimate vulnerability research and security testing, OpenAI says.

On the internal ExploitBench test, models must turn known browser vulnerabilities into working exploits. GPT-6.2 Astra clears a 50 percent success rate, far outpacing its predecessor GPT-6 Astra. Results come from tests with safeguards turned off. | Image: OpenAI
OpenAI plans to offer the new models through OpenAI Daybreak, which already gives vetted security researchers access to other models with fewer guardrails. The published cybersecurity benchmark numbers come from internal tests with safeguards and classifiers turned off. OpenAI lays out its methodology in the accompanying safety report.
Astra takes second in the Artificial Analysis Intelligence Index
In the Artificial Analysis Intelligence Index, which rolls up nine evaluations across knowledge work, coding, science, and general knowledge, GPT-6.2 Astra finishes second, just behind Anthropic's Claude Fable 5.2. Astra jumps five points over GPT-6 Astra, according to Artificial Analysis, passing Google's Gemini 4 Carbon (60 points) and Anthropic's Claude Opus 5.5 (59 points). GPT-6.2 Sol picks up two points to reach 57, matching GPT-6 Astra's level at only about 15 percent of the cost per task.
OpenAI owns the Pareto frontier on cost versus performance, the line where no other model is both cheaper and better, according to Artificial Analysis. Only Anthropic and Meta models also land on that frontier. OpenAI holds most of the top spots on efficiency too, measured by output tokens per task. Claude Fable 5.2 is the only outside model that shows up there.

GPT-6.2 Astra reaches 62 points, taking second place in the Intelligence Index just behind Claude Fable 5.2 (65). OpenAI models dominate the Pareto frontier on cost-performance. | Image: Artificial Analysis
Artificial Analysis backs up OpenAI's claims in science and factual knowledge. GPT-6.2 Astra leads on Terminal-Bench-Science 0.1 and places second behind Claude Fable 5.2 on the physics reasoning test CritPt v2, the factual knowledge benchmark AA-Omniscience, and the knowledge benchmark HLE-Diamond. Astra also leads in 3D design (AA-3D) and robotics (AA-Robotics v0.1), scoring 79 on robotics against Sol's 78 and Claude Fable 5.2's 75.
Coding is a different story, though. According to Artificial Analysis, both models have flatlined there. In the Coding Agent Index, measured with the Codex harness at max compute, GPT-6.2 Astra stays right where GPT-6 Astra was. GPT-6.2 Sol actually drops three points from GPT-6.1 Sol. That decline comes from a 14-point slide on SWE-Atlas-QnA-AA, where the model omits key facts from its answers more often than its predecessor did.
Artificial Analysis also flags a high hallucination rate. At maximum effort, GPT-6.2 Astra hallucinates 59 percent of the time, Sol 55 percent. The metric is AA-Omniscience Hallucination Rate, which measures how frequently a model guesses on factual questions instead of saying it doesn't know. These are unusually hard questions, so the numbers don't reflect everyday hallucination rates.
Epoch AI finds Astra on par with Fable 5.2 but weaker on real-world tasks
Research organization Epoch AI tested GPT-6.2 Astra on its own benchmarks before launch. The model scores 174 points in the Epoch Capabilities Index (ECI), tying Anthropic's Fable 5.2. Both models beat what the previous trend for reasoning models would have predicted before Astra, according to Epoch AI.
On real-world tasks pulled from Epoch AI's own research, Astra fell behind Fable 5.2 and barely improved over GPT-6 Astra. According to Epoch AI, the model missed the central question on simpler research projects and spun its wheels trying to fix things. On a task about researching AI data centers, Astra played it too safe and skipped key estimates, something Epoch AI says no contractor could get away with. Epoch AI rated Astra's writing as relatively strong compared to other tested models but still below its own bar.

GPT-6.2 Astra scores 174 points in the Epoch Capabilities Index, tying Anthropic's Fable 5.2. Both sit above the previous trend line for reasoning models. Astra is significantly cheaper. | Image: Epoch AI
One thing that stands out is Astra's ability to run complex reasoning within a single output token, with no visible chain of thought. Epoch AI clocks 12.3 serial steps for Astra, up from 10.5 for GPT-6 Astra and 4.2 for Claude Fable 5.1, the strongest non-OpenAI model at the time. Models that can reason this way without showing their work are harder to monitor, a safety concern that already came up with the first Astra model.
On FrontierMath Erdős, a benchmark of unsolved math problems, GPT-6.2 Astra cracked 7 of 68 tasks, each formally verified in the proof assistant Lean. The old record was 3 of 68, set by GPT-6.1 Sol. One of the seven had no known solution before.
Vals AI puts Astra and Sol in the top 10 but finds legal weak spots
GPT-6.2 Astra ranks 7th and GPT-6.2 Sol 9th in the Vals Index, separated by just 1.6 points. The price gap is wide, though. Sol runs $3.18 per task, less than a third of Astra's $10.68, making it the cheapest model in the top 10, according to Vals AI. Astra also got cheaper than its predecessor, scoring 2.6 points above GPT-6 Astra at roughly half of its $21.28 per-task cost. Vals AI says Astra uses the fewest output tokens in the top 10, averaging 91,000 per task.

GPT-6.2 Astra and Sol rank 7th and 9th in the Vals Index. Sol runs $3.18 per task, less than a third of Astra's $10.68, and is the cheapest model in the top 10. Claude Fable 5.2 tops the ranking. | Image: Vals AI
Vals AI backs OpenAI's claim to the top spot on the Vals Science Index. Astra leads that ranking and takes first on Terminal-Bench Science and VCB 1–100 at about half the cost and half the runtime of GPT-6 Astra. Sol comes in third on Terminal-Bench Science at roughly a quarter of Astra's cost. On the broader Terminal-Bench 4 for agentic coding, Astra places 5th and Sol 7th.
Legal tests are where the models fall short, according to Vals AI. On the HLAB benchmark, Astra and Sol double their predecessors' scores but rank only 11th and 21st. On the Legal Research benchmark, the drop is even steeper, with Astra at 30th and Sol at 52nd. Vals AI says the models make nearly twice as many tool calls as their predecessors but write shorter answers.
In September, OpenAI announced a specialized product called "Astra for Law." It does significantly better than the base model on the Legal Research Bench, reaching 6th place. Vals AI tested both models with a one-million-token context window and 128,000 max output tokens at full reasoning effort.
GPT-6.2 Astra is the first released model to hit OpenAI's "High" threshold for AI self-improvement
According to OpenAI, GPT-6.2 Astra is the first externally released model to reach the "High" threshold for AI self-improvement. OpenAI says the model can meaningfully speed up AI research, on par with giving every OpenAI researcher an extra, highly skilled research engineer.
OpenAI calls this out explicitly as a risk. Models that can help build future AI systems could push the field forward faster than safety measures and human oversight can keep up. The industry is currently wrestling with this under the label "RSI".
As countermeasures, OpenAI points to tighter controls in research environments and its monitoring system Control Room, which can track research tasks and shut them down if needed. The company also says it has expanded monitoring that examines how models reason and act and automatically flags unauthorized behavior.
On alignment, meaning how well the model follows user instructions and communicates openly, the new models match their predecessors, according to OpenAI. In two tests, however, they more often failed to flag obstacles blocking their work, like a missing file or an unavailable search tool.