GPT-6 Astra与Claude Fable 5.1等前沿模型横向评测
GPT-6 Astra vs GPT-6.1 Sol vs Gemini 4 Argon vs Claude Fable 5.1: Which Frontier Model Fits Which Job
四款旗舰模型密集发布的横向深度对比,详细拆解了不同工作负载下的选型策略与缓存成本影响,开发者必读。
Anthropic, OpenAI and Google DeepMind shipped 4 frontier-class models within 30 days. Claude Fable 5.1 arrived on September 1. GPT-6 Astra followed on September 3. GPT-6.1 Sol and Gemini 4 Argon landed in the last days of September.
Anthropic、OpenAI 和 Google DeepMind 在 30 天内发布了 4 款前沿级模型。Claude Fable 5.1 于 9 月 1 日上线,GPT-6 Astra 紧随其后于 9 月 3 日发布,GPT-6.1 Sol 和 Gemini 4 Argon 则在 9 月底陆续落地。
We covered each launch on its own. This piece puts them side by side. The benchmark scores overlap more than the launch posts suggest. The prices, access rules and cost per task do not.
我们此前已分别报道了每次发布。本文将它们并列对比。基准测试分数的重叠程度比各篇发布文章所暗示的要高得多。但在价格、访问规则和每项任务的成本方面则并非如此。
One change frames the lineup. OpenAI cancelled GPT-6.1 Astra on September 28 after it failed internal scope and authorization tests. GPT-6 Astra stays OpenAI’s top model for now.
阵容中有一个关键变化:OpenAI 于 9 月 28 日取消了 GPT-6.1 Astra,因其在内部范围和控制权测试中未通过。目前,GPT-6 Astra 仍是 OpenAI 的顶级模型。
Specs, Pricing and Access
规格、定价与访问权限
Astra and Fable 5.1 share the same $10 input and $50 output list price. Sol and Argon list at one-fifth of that. Argon’s price is introductory and doubles later.
Astra 和 Fable 5.1 的输入标价均为 10 美元,输出标价为 50 美元。Sol 和 Argon 的标价仅为该价格的五分之一。Argon 的价格为促销价,后续将翻倍。
| Feature | GPT-6 Astra | GPT-6.1 Sol | Gemini 4 Argon | Claude Fable 5.1 |
|---|---|---|---|---|
| Developer | OpenAI | OpenAI | Google DeepMind | Anthropic |
| Released | Sep 3, 2026 | Sep 29, 2026 | Announced Sep 30, 2026 | Sep 1, 2026 |
| Access | OpenAI API, ChatGPT, Codex | OpenAI API, ChatGPT Work, Codex | Fairwind Program only | Claude API, Bedrock, Google Cloud, Microsoft Foundry |
| Input / output (per 1M) | $10 / $50 | $2 / $10 | $2 / $10 intro, then $4 / $20 | $10 / $50 |
| Cached input (per 1M) | $1.00 | $0.10 | $0.10 intro | $0.25 |
| Long-prompt pricing | Above 272K: $20 / $75 | Above 272K: 2x input, 1.5x output | Not disclosed | Flat to 1M |
| Context window | 1.05M | 1.05M | Not disclosed | 1M |
| Max output per response | 128K | 128K | 1M | 128K |
| Reasoning control | 5 effort levels, low to max | 5 effort levels, low to max | Not disclosed | Effort levels incl. low, medium, high |
| Open weights | No | No | No | No |
| 功能 | GPT-6 Astra | GPT-6.1 Sol | Gemini 4 Argon | Claude Fable 5.1 |
|---|---|---|---|---|
| 开发者 | OpenAI | OpenAI | Google DeepMind | Anthropic |
| 发布时间 | 2026 年 9 月 3 日 | 2026 年 9 月 29 日 | 2026 年 9 月 30 日公布 | 2026 年 9 月 1 日 |
| 访问方式 | OpenAI API、ChatGPT、Codex | OpenAI API、ChatGPT Work、Codex | 仅限 Fairwind 计划 | Claude API、Bedrock、Google Cloud、Microsoft Foundry |
| 输入/输出(每百万 token) | $10 / $50 | $2 / $10 | 促销期 $2 / $10,之后 $4 / $20 | $10 / $50 |
| 缓存输入(每百万 token) | $1.00 | $0.10 | 促销期 $0.10 | $0.25 |
| 长提示词定价 | 超过 272K:$20 / $75 | 超过 272K:输入 x2,输出 x1.5 | 未披露 | 统一至 1M |
| 上下文窗口 | 1.05M | 1.05M | 未披露 | 1M |
| 单次响应最大输出 | 128K | 128K | 1M | 128K |
| 推理控制 | 5 个努力等级,从低到高 | 5 个努力等级,从低到高 | 未披露 | 包含低、中、高的努力等级 |
| 开放权重 | 否 | 否 | 否 | 否 |
Sources: OpenAI GPT-6.1 Sol coverage, Gemini 4 Argon coverage, GPT-6 Astra long-context pricing, Claude Fable 5.1 specs. Standard first-party list prices, short-context tier.
资料来源:OpenAI GPT-6.1 Sol 报道、Gemini 4 Argon 报道、GPT-6 Astra 长上下文定价、Claude Fable 5.1 规格。标准第一方标价,短上下文层级。
The cached-input row matters most for agents. Agents resend system prompts, tool schemas and history on every step. Astra’s $1.00 cache read is 4x Fable 5.1’s and 10x Sol’s.
缓存输入行对智能体(Agents)最为重要。智能体在每个步骤都会重新发送系统提示词、工具模式和历史记录。Astra 的 1.00 美元缓存读取费用是 Fable 5.1 的 4 倍,是 Sol 的 10 倍。
Argon’s 1M output cap is the only structural outlier. The other 3 stop at 128K tokens per response.
Argon 的 1M 输出上限是唯一结构性异常值。其他三款模型的单次响应限制均为 128K token。
Benchmarks: Where Each Model Leads
基准测试:各模型优势领域
No model sweeps the board. Argon leads the knowledge-work and long-horizon coding rows. Astra leads frontier software engineering and computer use. Opus 5.5, not in this lineup, leads Terminal-Bench 4.0.
没有一款模型能通吃所有榜单。Argon 在知识工作和长周期编码行中领先。Astra 在前沿软件工程和使用计算机方面领先。Opus 5.5(不在本次对比阵容中)在 Terminal-Bench 4.0 中领先。
| Benchmark | What it tests | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|---|
| DeepSWE v1.1 | Long-horizon software engineering | 77.9% | 74.1% | 67.4% |
| Vals Index | Finance, coding, legal and tax work | 68.9% | 63.1% | 65.8% |
| FrontierSWE v2 | Frontier software engineering | 55.0% | 65.5% | 56.3% |
| Terminal-Bench 4.0 | Agentic work in a terminal | 57.4% | 58.2% | 57.9% |
| CWE-bench v1 | Vulnerability remediation | 68% (tie) | 68% (tie) | 58% |
| OSWorld-2.0 | Computer use | 69.2% | 72.6% | Not in Google’s table |
| 基准测试 | 测试内容 | Gemini 4 Argon | GPT-6 Astra | Claude Fable 5.1 |
|---|---|---|---|---|
| DeepSWE v1.1 | 长周期软件工程 | 77.9% | 74.1% | 67.4% |
| Vals Index | 金融、编程、法律和税务工作 | 68.9% | 63.1% | 65.8% |
| FrontierSWE v2 | 前沿软件工程 | 55.0% | 65.5% | 56.3% |
| Terminal-Bench 4.0 | 终端中的智能体工作 | 57.4% | 58.2% | 57.9% |
| CWE-bench v1 | 漏洞修复 | 68%(并列) | 68%(并列) | 58% |
| OSWorld-2.0 | 计算机使用 | 69.2% | 72.6% | 不在谷歌的表格中 |
Source: Google DeepMind’s published comparison, as reported in our Gemini 4 Argon coverage. Vendor-reported.
来源:Google DeepMind 发布的对比数据,如我们在 Gemini 4 Argon 报道中所引用的。由供应商报告。
GPT-6.1 Sol is not in Google’s table. OpenAI’s own numbers place it close to Astra:
GPT-6.1 Sol 不在谷歌的表格中。OpenAI 自己的数据显示其表现与 Astra 接近:
- DeepSWE v1.1: Sol matches Astra at roughly one-fifth of the cost.
- OSWorld 2.0 offline set: Sol lands within 2.1 points of Astra at about one-seventh the cost per task.
- AutomationBench 1.0.6: Sol scores 2.2 points above Claude Opus 5.5 at medium effort.
- Terminal-Bench Science 0.1: Astra still leads at 68.1%. OpenAI recommends Astra for the hardest research.
- DeepSWE v1.1:Sol 的表现与 Astra 相当,但成本仅为后者的约五分之一。
- OSWorld 2.0 离线集:Sol 的表现与 Astra 相差仅 2.1 个百分点,且每项任务的成本约为 Astra 的七分之一。
- AutomationBench 1.0.6:在中等工作量下,Sol 的得分比 Claude Opus 5.5 高出 2.2 分。
- Terminal-Bench Science 0.1:Astra 仍以 68.1% 领先。OpenAI 建议在最困难的研究任务中使用 Astra。
Independent signals point the other way on raw intelligence. On the Artificial Analysis Intelligence Index, Astra scores 61. Fable 5.1 scores 5 points higher. On its coding-agent index, Fable 5.1 in Claude Code scores 70 against Astra’s 67. Artificial Analysis also reports that Argon equals Astra on the Intelligence Index.
独立信号在原始智力方面指向相反的方向。在 Artificial Analysis Intelligence Index 上,Astra 得分为 61,Fable 5.1 高出 5 分。在其编码智能体指数上,Claude Code 中的 Fable 5.1 得分为 70,而 Astra 为 67。Artificial Analysis 还报告称,Argon 在 Intelligence Index 上与 Astra 持平。
On ARC-AGI-2, Astra scores 95% and Fable 5.1 scores 90%.
在 ARC-AGI-2 上,Astra 得分为 95%,Fable 5.1 得分为 90%。
Cost per Task: Same List Price, Different Bill
每项任务的成本:相同的标价,不同的账单
Artificial Analysis puts Claude Fable 5.1 at $9.18 per task, against $4.72 for GPT-6 Astra. That is about 1.9x, at identical list prices.
Artificial Analysis 将 Claude Fable 5.1 定为每项任务 9.18 美元,而 GPT-6 Astra 为 4.72 美元。在标价相同的情况下,这大约是 1.9 倍。
The gap comes from token volume, not rates. Cost per task multiplies price by tokens spent. With equal rates, the gap implies Fable 5.1 spent more tokens per task in that run. Anthropic also notes its newer tokenizer produces roughly 30% more tokens for the same text.
差距来自令牌数量,而非费率。每项任务的成本等于价格乘以消耗的令牌数。在费率相同的情况下,这一差距意味着在该次运行中,Fable 5.1 每项任务消耗了更多的令牌。Anthropic 还指出,其较新的分词器对相同文本生成的令牌数大约多出 30%。
The two cheaper models change the picture further:
两款更便宜的模型进一步改变了局面:
- Gemini 4 Argon: Artificial Analysis reports Argon equals Astra’s Intelligence Index at 60% of Astra’s cost per task, using introductory prices.
- GPT-6.1 Sol: On Terminal-Bench Science, OpenAI reports $5.47 per task for Sol against $23.80 for Astra.
- Gemini 4 Argon:Artificial Analysis 报告称,在使用入门级价格时,Argon 的 Intelligence Index 与 Astra 持平,但每项任务的成本仅为 Astra 的 60%。
- GPT-6.1 Sol:在 Terminal-Bench Science 上,OpenAI 报告 Sol 每项任务成本为 5.47 美元,而 Astra 为 23.80 美元。
Caching can reverse the ranking for agents. The per-task figures above do not model heavy cache reuse. A long-running agent rereads the same context on every step. Here is the arithmetic for a 200K-token cached context, before output tokens:
缓存可以逆转智能体的排名。上述每项任务的数据并未模拟大量的缓存重用。长时间运行的智能体在每一步都会重新读取相同的上下文。以下是针对 200K 令牌缓存上下文(在输出令牌之前)的计算:
| Model | Cached rate (per 1M) | Per step | Per 100 steps |
|---|---|---|---|
| GPT-6 Astra | $1.00 | $0.20 | $20.00 |
| Claude Fable 5.1 | $0.25 | $0.05 | $5.00 |
| GPT-6.1 Sol | $0.10 | $0.02 | $2.00 |
| Gemini 4 Argon (intro) | $0.10 | $0.02 | $2.00 |
| 模型 | 缓存费率(每百万次) | 单步费率 | 100步费率 |
|---|---|---|---|
| GPT-6 Astra | $1.00 | $0.20 | $20.00 |
| Claude Fable 5.1 | $0.25 | $0.05 | $5.00 |
| GPT-6.1 Sol | $0.10 | $0.02 | $2.00 |
| Gemini 4 Argon (intro) | $0.10 | $0.02 | $2.00 |
Illustrative math from list cache rates. Astra’s 200K context stays under its 272K long-prompt threshold.
基于列表缓存费率的示例计算。Astra 的 200K 上下文长度低于其 272K 长提示阈值。
In cache-heavy loops, Fable 5.1 reads context at a quarter of Astra’s rate. Measure both on your own traces before you pick on per-task headlines.
在高度依赖缓存的循环中,Fable 5.1 读取上下文的费率仅为 Astra 的四分之一。在选择针对特定任务的 headline 之前,请先在你的自有追踪数据上对两者进行测量。
Which Model for Which Job
哪种模型适用于哪种任务
Pick by workload, not by leaderboard rank. GPT-6.1 Sol is the default for most teams. The other 3 earn their price on narrower jobs.
根据工作负载选择,而非排行榜排名。GPT-6.1 Sol 是大多数团队的默认选择。其他三种模型仅在更狭窄的任务中体现其价格优势。
| Job | Pick | Why | Runner-up |
|---|---|---|---|
| High-volume coding agents, CI bots, PR review | GPT-6.1 Sol | Matches Astra on DeepSWE v1.1 at $2 / $10 and $0.10 cached | Gemini 4 Argon, once public |
| Hardest open-ended engineering | GPT-6 Astra | Leads FrontierSWE v2 at 65.5% | Claude Fable 5.1 |
| Computer use and browser agents | GPT-6 Astra | Leads OSWorld-2.0 at 72.6% | GPT-6.1 Sol, within 2.1 points |
| Long coding sessions inside a harness | Claude Fable 5.1 | Top Artificial Analysis coding-agent score (70) in Claude Code; $0.25 cache reads | GPT-6 Astra (67) |
| Hard science and research | GPT-6 Astra | Leads Terminal-Bench Science at 68.1% | GPT-6.1 Sol at $5.47 per task |
| Legal, finance and business automation | Gemini 4 Argon | Leads Vals Index (68.9%), AutomationBench (51.3%) and Harvey Legal (19.6%) | GPT-6.1 Sol; Claude Fable 5.1 |
| Very long single outputs: big refactors, full reports | Gemini 4 Argon | Only model with 1M output tokens per response | Any of the 3 others, split across turns |
| Vulnerability finding and patching | Gemini 4 Argon or GPT-6 Astra | Tied at 68% on CWE-bench v1 | Claude Fable 5.1 (58%) |
| Tightest budget at frontier quality | GPT-6.1 Sol | Lowest public price; Argon matches it only inside Fairwind | Gemini 4 Argon |
| 任务 | 推荐选择 | 原因 | 备选 |
|---|---|---|---|
| 高容量编码代理、CI 机器人、PR 审查 | GPT-6.1 Sol | 在 DeepSWE v1.1 上与 Astra 持平,费率为 $2 / $10 和 $0.10 缓存 | Gemini 4 Argon,待公开 |
| 最难的开放式工程任务 | GPT-6 Astra | 在 FrontierSWE v2 中以 65.5% 领先 | Claude Fable 5.1 |
| 计算机使用和浏览器代理 | GPT-6 Astra | 在 OSWorld-2.0 中以 72.6% 领先 | GPT-6.1 Sol,差距仅 2.1 分 |
| 沙盒内的长编码会话 | Claude Fable 5.1 | 在 Claude Code 中获得 Artificial Analysis 编码代理最高分(70分);$0.25 缓存读取费率 | GPT-6 Astra(67分) |
| 硬科学与研究 | GPT-6 Astra | 在 Terminal-Bench Science 中以 68.1% 领先 | GPT-6.1 Sol,每项任务 $5.47 |
| 法律、金融和商业自动化 | Gemini 4 Argon | 在 Vals Index(68.9%)、AutomationBench(51.3%)和 Harvey Legal(19.6%)中领先 | GPT-6.1 Sol;Claude Fable 5.1 |
| 超长单次输出:大型重构、完整报告 | Gemini 4 Argon | 唯一支持每次响应输出 1M token 的模型 | 其他任意三种,可拆分到多轮对话 |
| 漏洞发现与修补 | Gemini 4 Argon 或 GPT-6 Astra | 在 CWE-bench v1 上并列第一,均为 68% | Claude Fable 5.1(58%) |
| 前沿质量下的最紧预算 | GPT-6.1 Sol | 最低公开价格;Argon 仅在 Fairwind 内与其持平 | Gemini 4 Argon |
Access decides 2 of these rows. Argon is only available to Fairwind cyber defenders today. Astra’s full offensive-security capability sits behind OpenAI’s Daybreak program; the public release refuses advanced offensive cyber tasks. Anthropic gates its unrestricted twin, Claude Mythos 5.1, behind trusted access programs.
访问权限决定了其中两行的可用性。目前,Argon 仅向 Fairwind 网络防御者开放。Astra 完整的进攻性安全能力位于 OpenAI 的 Daybreak 计划之后;公开发布版拒绝高级进攻性网络任务。Anthropic 将其无限制的孪生模型 Claude Mythos 5.1 置于受信任访问计划之后。
What to Check Before You Switch
切换前需要检查的事项
Most numbers here are vendor-reported, and vendors disagree at the margins. OpenAI reports Astra at 57.9% on Terminal-Bench 4.0. Google’s comparison table lists 58.2%.
此处的大多数数字由供应商报告,且在边缘情况下供应商之间存在分歧。OpenAI 报告称 Astra 在 Terminal-Bench 4.0 上的得分为 57.9%。Google 的比较表格列出的是 58.2%。
- Safeguards affect Fable 5.1 scores: Anthropic ran its benchmarks with production safeguards on. Flagged cyber and biology tasks route to other Claude models, which likely lowered OSWorld and AutomationBench results.
- Argon’s pricing is temporary: $2 / $10 is introductory. It moves to $4 / $20 later, with no end date announced.
- Argon’s context window is undisclosed: Google published the 1M output cap but not the input limit.
- Per-task costs are a snapshot: The $9.18 and $4.72 figures come from Artificial Analysis’ early-September Astra run. Effort settings shift them a lot.
- 安全护栏影响 Fable 5.1 得分:Anthropic 在启用生产环境安全护栏的情况下运行了其基准测试。被标记的网络和生物任务会路由到其他 Claude 模型,这可能导致 OSWorld 和 AutomationBench 的结果降低。
- Argon 的定价是暂时的:2 美元/10 美元为入门价格。后续将调整为 4 美元/20 美元,且尚未公布结束日期。
- Argon 的上下文窗口未公开:Google 公布了 1M 的输出上限,但未公布输入限制。
- 每项任务的成本仅为快照数据:9.18 美元和 4.72 美元的数据来自 Artificial Analysis 于九月初进行的 Astra 运行结果。努力设置(Effort settings)会大幅改变这些数值。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力