跪拜 Guibai
← All articles
AI Coding · AIGC

The AI Coding Hangover: When Token Bills Outpace Payroll Savings

By 李剑一 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The assumption that AI coding tools are a straightforward cost reduction is breaking down as token pricing rises across all major platforms. Teams that over-committed to AI-generated code now carry both the original staffing costs and a growing, unpredictable infrastructure bill, with no clear ceiling in sight.

Summary

Internal memos from Tencent, Alibaba, Microsoft, and Uber reveal a pattern: aggressive AI coding rollouts followed by abrupt pullbacks. Free token programs get suspended, budgets evaporate in months, and a single overnight test run can rack up a 2 million yuan bill. The promised efficiency gains collide with the reality that AI's capacity is bounded by context windows and token limits, while its operational costs keep climbing.

The core tension is straightforward. Two engineers plus AI rarely produce the output of five, and when they do, the token consumption can exceed a junior developer's monthly salary. When something breaks, no AI model will take responsibility, leaving companies exposed to client blowback and unrecoverable losses.

With every domestic coding plan raising prices simultaneously, the industry faces an uncomfortable truth: AI hasn't eliminated the largest cost center. It has simply relocated it from payroll to a metered API bill that only goes up.

Takeaways
Tencent and Alibaba launched free internal AI coding token programs, then suspended them months later and switched to role-based dynamic allocation.
Microsoft encouraged Claude Code adoption in January, then cut off access and mandated Copilot by June.
Uber burned through its entire 2026 AI budget of roughly $3.4 billion in four months.
Meta consumed 60 trillion tokens in 30 days; miHoYo reportedly ran a 2 million yuan token bill in a single night.
All domestic coding plan prices have risen, not just a few providers.
AI cannot take responsibility when generated code fails, leaving companies exposed to client trust and financial risk.
Two engineers with AI rarely deliver the output of five, and token costs can exceed a junior developer's salary.
Conclusions

The pattern of enthusiastic AI adoption followed by cost-driven rollbacks suggests that current pricing models are unsustainable for large-scale, always-on usage.

Token consumption is not linear with productivity; complex, long-context tasks can produce bills that dwarf the labor cost they were meant to replace.

The simultaneous price hike across all domestic coding plans points to an industry-wide margin squeeze rather than competitive differentiation, which limits buyer power.

Framing AI as a labor replacement ignores that accountability still sits with humans, creating a risk asymmetry where cost savings are realized by the company but liability stays with the remaining staff.

Concepts & terms
Token
The unit of text that an AI model processes or generates, used as the basis for billing in most coding assistant APIs. Longer context windows and complex requests consume more tokens, directly increasing cost.
Context window
The maximum amount of text (measured in tokens) that an AI model can consider at once when generating a response. Larger codebases and longer conversations push against this limit, requiring truncation or summarization that can degrade output quality.
Coding Plan
A subscription or usage-based pricing tier offered by AI coding tool vendors, typically bundling a set number of tokens, model access, and features for a recurring fee.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗