编码代理 Token 成本优化 2026：70 倍差距从哪来

2026 年 8 月的两个 benchmark 把编码代理的账单问题摆上了台面：同一个模型、同一批任务，Aider 每解决一个任务大约花 3,500 token，OpenClaw 要花 292,000 token——差了 70 倍。Composio 的企业工作流测试里，每成功一次的成本从 Pi Agent 的 $0.028 到 Claude Code 的 $0.195，而 DeepAgents 用四分之一的价格拿到了完全相同的通过率。问题不在模型，在框架。本文拆解 token 成本的真实来源，并给出可落地的优化代码。

1. 数据先说话：70 倍差距从哪来

6 月的独立 benchmark 用 12 种配置跑了同一批 12 个 Python 任务，全部走 OpenRouter，保证模型和 API 完全一致。结果是每解决一个任务消耗 3,500 到 292,000 token，而且这个排序在换成另一个模型后几乎不动——说明差距来自框架软件本身，而不是模型行为。8 月 Composio 用 8 个框架跑 30 个企业工作流（Airtable、Gmail、Google Calendar、Google Sheets、GitHub、Slack、PostHog），用程序化验证器而不是 LLM 裁判打分，还埋了诱饵数据。240 次运行中 129 次成功，每次成功成本最低 $0.028、最高 $0.195。DeepAgents 与 Claude Code 通过率完全一致，但每次成功成本只有四分之一。

// The measurement that started it all: run the same model
// through different harnesses and count tokens per solved task.
// June 2026 benchmark — 12 configs, 2 models, 12 Python tasks,
// all through OpenRouter so the model and API were identical.

import { countTokens } from "./token-counter";

const harnesses = ["aider", "claude-code", "codex", "goose",
  "hermes", "kilo", "kimi-code", "nanobot", "openclaw",
  "opencode", "qwen-code"];

for (const h of harnesses) {
  const result = await runSuite(h, {
    model: "deepseek-v4-flash",
    tasks: PYTHON_TASKS,
  });
  console.log(h, {
    tokensPerSolved: Math.round(result.tokens / result.solved),
    passRate: result.solved / result.total,
  });
}

// Output (tokens per solved task):
// aider (architect mode): ~3,500
// openclaw:             ~292,000  <- 70x gap, same model!

2. 启动税：每一轮都在重复付的固定成本

差距的核心是一个简单的测量：启动税（startup tax）。在提示词开始干活之前，框架要先带上自己的行李——系统提示词、工具描述、环境设置。Aider 的 architect 模式大约 700 token，OpenClaw 大约 26,000 token。如果这笔钱只付一次还能忍，但重发模式（resend pattern）让它每轮都付：一个带着 26,000 token 底座的框架跑 15 轮，光脚手架就吃掉大约 390,000 个输入 token。

// The startup tax: before a prompt does any work, the harness
// ships its own baggage — system prompt + tool descriptions +
// environment setup. This floor is paid on EVERY turn.

export function measureStartupTax(harness: Harness): number {
  const baseline = harness.tokenize(
    harness.systemPrompt + harness.toolDescriptions.join("")
  );
  // aider architect mode:  ~700 tokens
  // openclaw:             ~26,000 tokens
  return baseline;
}

// Why it compounds: the resend pattern. A 26,000-token floor
// through 15 turns = ~390,000 input tokens on scaffolding alone,
// before the model writes a single line of code.
const floor = 26000;
const turns = 15;
console.log(floor * turns); // 390,000

3. 为什么 DeepAgents 能用四分之一成本打平

Composio 的对比很有说服力：DeepAgents 和 Claude Code 的通过率完全相同，都是 54%，但每次成功的成本一个 $0.049、一个 $0.195。同样的任务、同样的验证标准，差距只可能来自提示词设计、工具调度和上下文管理。这说明优化杠杆不在模型而在框架——把框架的提示词地板压下来，账单直接骨折。

// Composio's August benchmark: 8 harnesses, 30 enterprise
// workflows (Airtable, Gmail, Calendar, Sheets, GitHub, Slack,
// PostHog). 900-second ceiling, programmatic verifier, decoys.

const results = [
  { harness: "pi-agent", costPerSuccess: 0.028, passRate: 0.43 },
  { harness: "deepagents", costPerSuccess: 0.049, passRate: 0.54 },
  { harness: "claude-code", costPerSuccess: 0.195, passRate: 0.54 },
  // DeepAgents matched Claude Code's pass rate exactly
  // while costing a quarter as much per success.
];

// The optimization lever is not the model — it's the harness.
// The same pass rate at 1/4 the cost means the bottleneck is
// prompt engineering of the harness, not model quality.

4. 一个 R²=0.99 的预测模型

最实用的结论来自回归分析：启动税乘以轮次数，就能以 R²=0.99 的精度预测每任务的 token 消耗——这个关系在两个不同模型上都成立。换句话说，你想优化代理账单，第一刀应该砍在提示词地板和轮次数上，而不是急着换模型或者上花哨的缓存方案。先量，再砍。

// The predictive model: startup tax × turn count explains
// tokens per solved task with R² = 0.99 across BOTH models.
// That is the actionable insight.

export function predictTokensPerTask(harness: Harness, model: Model) {
  const startupTax = measureStartupTax(harness);
  const avgTurns = harness.avgTurnsPerTask(model);
  return startupTax * avgTurns;
}

// Optimization playbook, in priority order:
// 1. Shrink the prompt floor (system prompt, tool schemas)
// 2. Reduce turn count (better planning, fewer retries)
// 3. Only then consider model swapping / caching
export const playbook = [
  "trim tool descriptions to the minimum viable schema",
  "remove rarely-used tools from the active toolset",
  "add a planning step to cut failed-attempt turns",
  "enable prompt caching for the stable prefix",
];

5. 优化行动清单

按优先级排列：先把工具描述裁到最小可用 schema，删掉不常用的工具；再加一个规划步骤减少失败重试的轮数；最后才考虑 prompt caching 这类高级手段。框架的提示词地板乘轮次数，就是你账单的骨架。

6. 别忘了给任务设预算

生产环境里给每个任务设 token 预算，超了就 fail fast，别让代理在死循环里烧钱。代码里用一个简单的定时检查器监控当前上下文 token 数，超过上限立刻终止。省下的每一分钱都是利润。

// Practical harness-side budget guard: cap spend per task
// and fail fast instead of burning tokens on loops.

export async function runWithBudget(task: Task, budget: TokenBudget) {
  const start = countTokens(await systemPrompt(task));
  let used = start;
  const guard = setInterval(() => {
    used = countTokens(await currentContext());
    if (used > budget.max) {
      console.error("budget exceeded", used, ">", budget.max);
      process.exit(1);
    }
  }, 1000);
  try {
    return await runAgent(task);
  } finally {
    clearInterval(guard);
  }
}