A small "analyze the source code" task — the AI ran for nearly six minutes and burned 1.15 million tokens. The bill hides an uncomfortable truth: 99% of the cost is invisible input. Volcengine (ByteDance's cloud service) this week released a debugging plugin for DeepSeek, laying this account bare in front of developers. What we care about is not the tool itself, but the trend it exposes: in the Agent era (the paradigm where AI autonomously completes multi-step tasks), the cost black hole hides in repeatedly read files and stacked conversation histories.
What This Is
Volcengine released a debugging plugin for tracking AI Agents (AI that can autonomously call tools and complete multi-step tasks): TLS for DeepSeek Harness. It logs every AI thought, file read, and command execution — equivalent to installing a dashcam on the AI.
The tool itself is straightforward; the real story is the execution bill attached to it: asking the AI to analyze a piece of source code took nearly six minutes, consumed 1.15 million tokens (the unit for measuring processed text), of which 1.03 million were input context — input was 99 times the output.
Industry View
Supporters argue that for Agents to enter enterprise production environments, they must be observable and accountable. This is a necessary step in AI engineering.
But cooler heads are pushing back. Senior developers caution that observability tools themselves generate new complexity — a single failure can originate from the model, the context, the tools, or the environment, with none of the convenience of traditional software's "read the stack trace to locate the issue" workflow. The cost-side risk is more concrete: at DeepSeek's public pricing, a few yuan per task looks cheap, but enterprises running thousands of calls per day could see monthly bills quietly cross five figures — and no one is sounding the alarm in advance.
Impact on Regular People
For enterprise IT: a new "cost dashboard" is required. Traditional software charges per seat; AI burns money per token. Without usage monitoring, month-end reconciliation can deliver a nasty shock.
For individual professionals: people who know how to "tame" AI become more valuable. Those who can break down tasks and control context length can compress the same job from a few yuan down to a few cents.
For the consumer market: free AI assistant products may quietly shrink — either capping conversation rounds or stealthily degrading response quality.