只降价20%,账单却少八成,GPT-5.6杀入Claude地盘重算编程账

marsbitPublicado em 2026-08-28Última atualização em 2026-08-28

Resumo

8月24日,OpenAI与AWS联合测试显示,GPT-5.6 Terra在AWS的软件开发平台Kiro上完成任务成本降低了约82%,远超模型本身20%的官方降价幅度。这主要得益于Kiro的“spec-driven”工作流优化,它将模糊任务拆解为清晰需求文档与技术设计,减少了模型在规划、试错和返工中的无效token消耗。 测试基于Terminal-Bench 2.1,评估的是“智能体+模型”组合的整体效率。对比数据显示,相同准确率下,不同框架与模型组合的成本差异显著。例如,Claude Code与Fable 5组合成本为552.67美元,而Codex与GPT-5.5组合成本高达2059.19美元。 今年7月,GPT-5.6系列模型(Sol、Terra、Luna)入驻Kiro,标志着OpenAI首次进入这一由Anthropic主导的开发平台。模型采取渐进式开放,且推理过程为隐藏思维链。价格上,Kiro随后跟进OpenAI调价,Luna计费系数大幅下调至0.1倍,Terra降至1.0倍。 此次成本大幅降低凸显了AI开发的新趋势:评价标准正从单纯的模型性能转向“完成任务的综合成本”。智能体框架、编排策略与模型分工的协同优化,成为降低实际应用开销的关键。对开发者而言,选择模型时更需要关注端到端的执行效率与经济性。

同一个模型,官方价格只降了20%,你的账单却少了八成。

8月24日,OpenAI公布了一项与AWS一起做的测试结果:

在Terminal-Bench 2.1上,GPT-5.6 Terra在Kiro里完成一个成功任务的成本,降低了约82%。

Kiro是AWS的软件开发智能体平台,覆盖IDE、CLI和Web。

GPT-5.6家族的Sol、Terra、Luna三兄弟,已在里面跑了一个多月。

这82%,并非官方降价。

Terra最近一次调价是在7月30日,幅度20%。

7月30日OpenAI调价公告,Terra降20%。

单价只降20%,账单却能砍掉八成。

中间差的那六十个百分点是从哪儿省出来的,才是真正值得我们关注的。

GPT-5.6登陆Kiro是今年7月的事儿。

先是13日,AWS宣布GPT-5.6 Sol、Terra、Luna在Amazon Bedrock正式可用。

紧接第二天,Kiro发博客宣布三款模型上线IDE、CLI和Web。

这是OpenAI模型第一次进Kiro,刚好赶上Kiro公开预览一周年。

先把模型摆上货架,再拉去干活,一个多月后,OpenAI回来交作业:

两家联手把Kiro环境和OpenAI模型一起调了一遍,Terra完成一个成功任务的成本,降了约82%。

省下的

是走弯路的钱

7月30日那次调价,Terra降了20%,可到了Kiro的测试里,单任务成本降了82%。

那多省下的六成,是从哪儿来的?

方向就这么几个:

模型吐的token更少了,工具来回调的次数更少了,失败之后的重试和绕远路更少了。

所以省下的这一大块,并非每次调用的钱,而是那些原本会白花的调用的钱。

道理很简单:一个AI智能体把任务做砸一次,账单照记。它中途选错路径、跑偏三轮再回头,这些token也照收。

真实开发里,钱往往就是这么漏掉的。

OpenAI在官方博客中也点过同一套逻辑,效率来自三层:

发起请求、组织上下文的智能体框架,中间调度请求的编排系统,最后才是GPU上跑的模型本身。

OpenAI拆解GPT-5.6的效率来源:请求从智能体框架出发,经编排系统调度,最后到GPU跑模型,每一层都在省。

省钱还能省到模型分工上。

OpenAI还举过一个用法:编码工作流可以先用Sol把问题想清楚、把计划定下来,再换Luna去实施那些已经定义清楚的改动、写测试、跑评估。

同一条流水线,不同环节配不同档位的智能。

模型只占账单一半

换个框架就换个价

Terminal-Bench 2.1这套题,不是让模型单独答卷。

它把模型丢进一个终端环境,给一个模糊目标,让它自己规划路径、调工具、写脚本、处理报错、反复迭代。

所以跑出来的成绩,是「智能体+模型」这个组合的成绩。

Terminal-Bench 2.1公开榜单,准确率右边多了一项成本。(图源:Terminal-Bench)

榜单里的四行数字最能说明问题:

Claude Code配Fable 5,83.8%,552.67美元;

Codex配GPT-5.5,83.1%,2059.19美元;

Codex配GPT-5.6 Terra,78.4%,421.15美元;

Codex配GPT-5.6 Luna,75.7%,241.45美元。

前两行分数只差0.7个百分点,账单却差了将近4倍。

同一个模型,装进不同的框架,配不同的上下文组织和工具策略,跑出来也根本不是一个价。

按Kiro官方给的数据,Terra在Coding Agent Index上拿77.4分,只比Claude Fable 5的77.2高一点。

它的卖点不在分数,在于这个分数所对应的价钱。

Kiro那套spec-driven的发力点也在这里,核心玩法就一句:不许上来就写代码。

它先把用户那句含糊的目标,拆成正经的需求文档、技术设计和可执行任务清单,再交给模型。

这样,模型到手的不再是一句含糊不清的话,而是一份理清楚的活儿。

熟悉Agent的人会立刻反应过来,这一步省的正是最昂贵的那部分开销。

模型跑偏、返工、推倒重来,烧掉的token往往比正经干活还多。

Kiro还在流程里留了两道闸栏:代码真正修改前,先停下来让人过一眼;活干完之后,再自动跑一轮测试验证对不对。

这两道闸拦下的每一次返工,都能省下真金白银。

Claude在亚马逊的地盘

GPT分走了一半

一年前,Kiro还只是个spec-driven的IDE,模型选择器是Anthropic的主场。

一年后,AWS在自家开发智能体平台上,一口气摆进三档OpenAI模型。

Sol、Terra、Luna跟Claude并排列进同一个下拉菜单,这画面搁一年前很难想象。

虽然GPT-5.6这次是「全家入驻」,但并未「全面开放」。

三款模型是渐进式、实验性开放,面向Pro、Pro+、Pro Max和Power用户,区域只有两个,美国的弗吉尼亚北部和欧洲的法兰克福,支持跨区域推理。

还有一条挺多人不适应的:这批模型在Kiro里走的是隐藏思维链,你看不到它的推理步骤,只看得到最终结果。

习惯了盯着Agent一步步推理的人,会觉得像把活扔进了一个不透明的盒子。

官方的说法是,这是预期行为,不影响输出质量。

三档模型在Kiro里明码标价。

7月14日刚上线那会儿,同样的任务,Sol扣2.4倍,Terra扣1.2倍,Luna扣0.6倍。

7月30日OpenAI那轮降价落地,第二天Kiro跟进:Luna从0.6倍一路砍到0.1倍,Terra从1.2倍降到1.0倍,只有Sol未动。

AWS的态度摆在了明面上:一个开发平台,不能只绑一家模型。

同一个选择器里,两家前沿模型开始互相压价。

模型的评价标准也跟着变了:分数高不再默认能赢,花钱少同样能赢。

对开发者来说,选模型以前问的是「这个模型多少钱一百万token」,现在得问「让它把这件事做完,我到底要花多少」。

参考资料:

https://x.com/OpenAIDevs/status/2091966982015103068

https://openai.com/index/gpt-5-6-in-kiro/

本文来自微信公众号“新智元”(ID:AI_era),作者:ASI启示录,编辑:元宇

Criptomoedas em alta

Perguntas relacionadas

Q根据文章,GPT-5.6 Terra模型的官方价格在7月30日降低了多少?最终在Kiro平台上完成一个任务的平均成本降低了多少?

AGPT-5.6 Terra模型在7月30日的官方价格降低了20%。在AWS的Kiro平台上,通过模型与平台的协同优化,完成一个成功任务的成本降低了约82%。

Q文章中提到账单成本大幅降低(82%)的关键因素是什么?请简述这些因素。

A账单成本大幅降低的关键因素并非仅仅来自模型单价的下降(20%),更多是来自效率优化。这主要包括:模型输出的token更少,工具调用次数减少,任务失败后的重试和绕路减少。本质上,节省的是那些因任务失败、路径选择错误而“白花”的token费用。OpenAI指出效率来自三层优化:智能体框架、请求编排系统和模型本身。

Q在Terminal-Bench 2.1的榜单对比中,为什么Claude Code配Fable 5与Codex配GPT-5.5的准确率相近,但成本相差近4倍?

A因为Terminal-Bench 2.1测试的是“智能体+模型”组合的整体表现。虽然两者最终准确率相近,但不同的智能体框架(如Fable 5与Codex)在组织上下文、调度工具、规划任务路径等方面的策略和效率不同。Codex配GPT-5.5的组合在完成任务过程中可能消耗了更多的token(例如,更多次数的无效尝试、返工),导致总成本远高于Claude Code配Fable 5的组合。

QAWS的Kiro平台采用了什么核心方法来降低AI编程智能体的成本?

AKiro平台采用了“spec-driven”(规范驱动)的核心方法。它不直接让模型根据模糊指令写代码,而是先将用户模糊的目标拆解成详细的需求文档、技术设计和可执行的任务清单,再交给模型执行。这样模型接收到的是清晰、结构化的任务,大幅减少了因理解偏差而导致的跑偏、返工和token浪费。平台还设置了人工确认和自动化测试两道“闸栏”,进一步拦截错误,节省成本。

QGPT-5.6系列模型(Sol, Terra, Luna)在Kiro平台上线后,对开发者选择模型的标准产生了什么影响?

AGPT-5.6系列模型上线Kiro并与Claude模型同台竞争后,改变了开发者选择模型的标准。过去开发者主要关注“每百万token的价格”,而现在更需要关注“完成特定任务的总成本”。模型的竞争维度从单纯的“能力分数”扩展到了“成本效率”,即花更少的钱把事情办成同样变得具有竞争力。这促使开发者在评估模型时,必须综合考虑其在实际工作流中的整体效率和花费。

Leituras Relacionadas

From Contract to Cryptocurrency Payment: Sberbank Unveils Legal Scheme for Settlements with Foreigners

Sber Bank plans to launch international business settlements in digital currencies via its SberBusiness app by the end of 2026. The bank's deputy chairman, Anatoly Popov, announced this ahead of the Eastern Economic Forum, stating the goal is to simplify digital currency use for businesses in cross-border trade. The legal basis is Federal Law No. 282-FZ "On Digital Currency and Digital Rights," effective September 1, 2026, which permits foreign trade crypto settlements through licensed intermediaries. A published guide outlines the process: a company signs a contract, transfers rubles to a licensed intermediary, who then buys and sends cryptocurrency to the recipient, with reporting for regulators generated automatically. This service is positioned as an alternative channel for foreign trade, especially where traditional bank transfers face sanctions. Sber aims to integrate crypto payments into standard business tools within SberBusiness, avoiding the need for specialized technical knowledge. Development will depend on the practical application of the new law. From a macro perspective, this initiative is seen not just as a technical innovation but also within the context of sanctions pressure on crypto markets. The article notes that licensed intermediaries could become targets for secondary sanctions, as seen with UK actions against crypto exchanges in summer 2026. The long-term viability of Sber's scheme may depend more on political dynamics than technology, following a historical pattern where new payment channels face regulatory countermeasures.

cryptonews.ruHá 1h

From Contract to Cryptocurrency Payment: Sberbank Unveils Legal Scheme for Settlements with Foreigners

cryptonews.ruHá 1h

Don't Trust, Verify: Malicious AI Links Expose a Nightmare Reality for Crypto Industry Workers

Generative AI is becoming increasingly prevalent, with professionals in the digital assets and blockchain space regularly using it. However, this makes them prime targets for attackers seeking to steal sensitive, often irrevocable, information. Refi Hub co-founder Numa Lunah recently reported being "hacked" through a malicious link sent in a chat with the AI model Claude, which appeared to be a legitimate transcription app download. The link installed malware that attempted to steal all his data. While Numa claimed no sensitive data was leaked—as he wiped and reinstalled his laptop's OS—he later discovered a corrupted `SKILL.md` file in his backups, disguised as a style guide. This file contained hidden instructions to reload the malware and steal credentials whenever the AI accessed it. This incident highlights a new attack vector: LLMs providing malicious outputs. Microsoft Defender experts have previously warned about the evolution of cryptojacking attacks from SEO poisoning to "poisoning" LLM responses from models like Gemini, Claude, Copilot, and ChatGPT. Threats include malicious artifacts via context windows, chatbot-recommended download links, fake AI-branded installers, and infected source code or agent skills. For crypto professionals, the risk is heightened because they often manage irreplaceable secrets like seed phrases, private keys, exchange API keys, and wallet data. The article argues that the most dangerous vulnerability is human trust and complacency. A fundamental shift in security culture is needed: treating every AI suggestion as potentially hostile, regardless of its source. This isn't paranoia but a survival necessity. The key takeaway is that the threat lies not in vulnerable code but in the human instinct to trust convenient answers, a flaw no software patch can fix. Numa was saved only because he meticulously reviewed every configuration file before letting the AI interact with it—a level of caution most users likely neglect.

cryptonews.ruHá 3h

Don't Trust, Verify: Malicious AI Links Expose a Nightmare Reality for Crypto Industry Workers

cryptonews.ruHá 3h

Trading

Spot

Artigos em Destaque

Como comprar BILL

Bem-vindo à HTX.com!Tornámos a compra de Billions Network (BILL) simples e conveniente.Segue o nosso guia passo a passo para iniciar a tua jornada no mundo das criptos.Passo 1: cria a tua conta HTXUtiliza o teu e-mail ou número de telefone para te inscreveres numa conta gratuita na HTX.Desfruta de um processo de inscrição sem complicações e desbloqueia todas as funcionalidades.Obter a minha contaPasso 2: vai para Comprar Cripto e escolhe o teu método de pagamentoCartão de crédito/débito: usa o teu visa ou mastercard para comprar Billions Network (BILL) instantaneamente.Saldo: usa os fundos da tua conta HTX para transacionar sem problemas.Terceiros: adicionamos métodos de pagamento populares, como Google Pay e Apple Pay, para aumentar a conveniência.P2P: transaciona diretamente com outros utilizadores na HTX.Mercado de balcão (OTC): oferecemos serviços personalizados e taxas de câmbio competitivas para os traders.Passo 3: armazena teu Billions Network (BILL)Depois de comprar o teu Billions Network (BILL), armazena-o na tua conta HTX.Alternativamente, podes enviá-lo para outro lugar através de transferência blockchain ou usá-lo para transacionar outras criptomoedas.Passo 4: transaciona Billions Network (BILL)Transaciona facilmente Billions Network (BILL) no mercado à vista da HTX.Acede simplesmente à tua conta, seleciona o teu par de trading, executa as tuas transações e monitoriza em tempo real.Oferecemos uma experiência de fácil utilização tanto para principiantes como para traders experientes.

709 Visualizações TotaisPublicado em {updateTime}Atualizado em 2026.06.02

Como comprar BILL

Discussões

Bem-vindo à Comunidade HTX. Aqui, pode manter-se informado sobre os mais recentes desenvolvimentos da plataforma e obter acesso a análises profissionais de mercado. As opiniões dos utilizadores sobre o preço de BILL (BILL) são apresentadas abaixo.

活动图片