简而言之：在困难编码任务（SWE-Bench Verified）上，微软的 MAI-Code-1-Flash 比 Claude Haiku 4.5 少用 60% 的 token，在 SWE-Bench Pro 上领先 16 分，并具备自适应解答长度控制。关键洞见在于：每美元智能，即每消耗一个 token 所获得的价值，才是评估生产环境中 AI 模型的正确指标。
昨天，微软在模型发布卡片上新增了一项指标，这项指标很可能会成为标准。1
平均 token 用量。
在第一行中，微软的模型在 SWE-Bench Verified 上达到 71.6，而所用 token 仅为 Claude Haiku 4.5 消耗量的约三分之一。
如今基准测试从两个不同维度来衡量：整体性能，以及达成该智能水平所需的成本。
这再次表明，补贴2、tokenmaxxing3，以及许多用例中不计代价追求性能的时代已经结束。
即便是全球最有价值的公司，也无力为每一种可能的用例都负担最先进的智能。4 Uber 在四个月内就花光了预算，随后对员工的 AI 支出设了上限。5 Salesforce 在 Anthropic tokens 上花费了 3 亿美元，并已冻结工程岗位招聘。6
这个全新的双重基准回答了买家唯一的问题：我的每一美元能换来多少智能？
Artificial Analysis 已经在对这一点进行基准测试了。7 GPT 5.5 与 Claude Opus 4.8 在 Intelligence Index 上仅相差一个点，均在 60 左右。运行该指数在 GPT 5.5 上花费 $3,357，在 Opus 4.8 上花费 $4,685。同样的答案，贵了 40%。
模型公司如今必须在两个维度上展开竞争。应用层则会在更高一层竞争，即按成果计费——一个已关闭的工单、一个已交付的 PR，或一个已解决的客服案例，实际成本究竟是多少。
如今，技术栈中的每一层都必须按照客户的思维方式来定价：按结果，而非按 token。
-
推出 MAI-Code-1-Flash —— 微软发布了一款新的编程模型，并在发布说明中标注了平均 token 用量。↩︎
-
微软取消 Claude Code 许可证，将开发者转向 GitHub Copilot CLI — 在工程使用量超出预算后，微软取消了其体验与设备部门（Windows、Microsoft 365、Outlook、Teams、Surface）的 Claude Code 许可证。↩︎
-
Uber 在 4 个月内耗尽预算后限制员工 AI 支出 — Uber 在四个月内耗尽预算后限制员工 AI 支出。↩︎
-
Salesforce 在 AI 上花费 3 亿美元，冻结工程招聘 — Salesforce 在 AI 上花费 3 亿美元，冻结工程招聘。↩︎
-
AI 模型与 API 提供商分析 — 对 AI 模型成本的独立分析。↩︎
In short : Microsoft's MAI-Code-1-Flash achieves 60% fewer tokens than Claude Haiku 4.5 on hard coding tasks (SWE-Bench Verified), a +16-point lead on SWE-Bench Pro, and adaptive solution length control. The key insight: intelligence per dollar, value per token spent, is the right metric for evaluating AI models in production.
Yesterday Microsoft added a new metric to a model release card, one that will likely become a standard.1
Average token usage.
In the first row, the Microsoft model hits 71.6 on SWE-Bench Verified using about a third of the tokens Claude Haiku 4.5 burns.
Benchmarks are now measured on two different dimensions, the overall performance & the cost to achieve that intelligence.
This is yet another sign that the era of subsidies2, tokenmaxxing3, & all-out performance for many use cases is over.
Even the most valuable companies in the world cannot afford state-of-the-art intelligence for every conceivable use case.4 Uber capped employee AI spending after blowing through its budget in four months.5 Salesforce is spending $300M on Anthropic tokens & has frozen engineering hires.6
This new dual benchmark answers the buyer’s only question : what is my intelligence per dollar?
Artificial Analysis already benchmarks this.7 GPT 5.5 & Claude Opus 4.8 land within a point of each other on the Intelligence Index, around 60. Running the index costs $3,357 on GPT 5.5 & $4,685 on Opus 4.8. Same answer, 40% more expensive.
Model companies must now compete on both dimensions. The application layer will compete one level up, on dollars per outcome, what a closed ticket, a shipped PR, or a resolved support case actually costs.
Every layer in the stack now has to price the same way the customer thinks : per result, not per token.
-
Introducing MAI-Code-1-Flash — Microsoft announces a new coding model with average token usage on the release card. ↩︎
-
The Unsustainable Subsidy — The era of AI subsidies is ending. ↩︎
-
Tokenmaxxing — Models that game benchmarks with extra tokens are losing their edge. ↩︎
-
Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI — Microsoft cancelled Claude Code licenses across its Experiences and Devices division (Windows, Microsoft 365, Outlook, Teams, Surface) after engineering usage outran budgets. ↩︎
-
Uber caps employee AI spending after blowing through budget in 4 months — Uber caps employee AI spending after blowing through budget in four months. ↩︎
-
Salesforce Spends $300M on AI, Freezes Engineering Hires — Salesforce Spends $300M on AI, Freezes Engineering Hires. ↩︎
-
AI Model & API Providers Analysis — Independent analysis of AI model costs. ↩︎