> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tokenlab.sh/llms.txt
> Use this file to discover all available pages before exploring further.

# Coding Agent Cost Optimization

> Practical strategies to reduce coding agent API costs by 60-90%

## The Cost Problem

A typical coding agent session burns through tokens fast:

| Activity        | Tokens per call | Calls per hour | Hourly tokens |
| --------------- | --------------- | -------------- | ------------- |
| Code generation | 5,000–50,000    | 10–30          | 150K–1.5M     |
| Codebase search | 2,000–20,000    | 20–50          | 100K–1M       |
| Code review     | 10,000–80,000   | 5–10           | 100K–800K     |
| Autocomplete    | 500–3,000       | 50–200         | 50K–600K      |
| **Total**       |                 |                | **400K–4M+**  |

At premium model rates, that's $3–30/hour per developer. For a team of 10, that's $500–5,000/month.

## Smart Model Selection

Not every coding task needs the most expensive model. Match the task to the right tier:

| Task                | Recommended                                   | Cost Tier        | Why                               |
| ------------------- | --------------------------------------------- | ---------------- | --------------------------------- |
| Architecture design | `claude-opus-4-6`, `gpt-5.4`                  | \$\$\$\$ Premium | Complex reasoning needed          |
| Code generation     | `claude-sonnet-4-6`, `gemini-3.1-pro-preview` | \$\$\$ Standard  | Best quality/cost balance         |
| Code review         | `claude-sonnet-4-6`, `deepseek-v4-pro`        | \$\$–\$\$\$      | Pattern matching, less creativity |
| Bug fixing          | `claude-sonnet-4-6`, `gpt-5-mini`             | \$\$–\$\$\$      | Focused, well-defined tasks       |
| Tab completion      | `gpt-5-mini`, `gemini-3.5-flash`              | \$\$ Budget      | Speed matters more than depth     |
| Boilerplate         | `deepseek-v4-flash`, `gpt-5-mini`             | \$ Economy       | Simple, repetitive patterns       |

<Tip>
  See [Model Selection Guide](/guides/coding-agent-models) for detailed model comparisons and per-tool configuration.
</Tip>

## Caching Strategies

Coding agents are ideal for caching because they repeat similar patterns constantly.

### Prompt Cache (Provider-Level)

Provider-level prompt caching is automatic through TokenLab. Long system prompts — which coding agents always include — get cached at the provider level:

| Provider  | Cache Discount | Min Tokens |
| --------- | -------------- | ---------- |
| Anthropic | 90% off reads  | 1,024      |
| OpenAI    | 50% off reads  | 1,024      |
| DeepSeek  | 90% off reads  | 64         |

Since coding agents send the same system prompt + project context on every call, prompt cache hit rates are typically **70–90%**.

### Prompt Cache Savings Example

For a request with 50,000 input tokens (typical coding agent call):

```
Direct API (no caching):
  50,000 tokens × $3.00/1M = $0.150

With prompt cache (40,000 cached + 10,000 new):
  Cached:  40,000 × $0.30/1M = $0.012
  New:     10,000 × $3.00/1M = $0.030
  Total: $0.042 (72% savings)
```

## Real Cost Comparison

Estimated costs for a typical 1-hour coding session (\~3M tokens):

| Setup                      | Hourly Cost | Monthly (160h)  |
| -------------------------- | ----------- | --------------- |
| Direct API (premium model) | \~\$15–25   | \~\$2,400–4,000 |
| TokenLab (smart routing)   | \~\$10–18   | \~\$1,600–2,900 |
| TokenLab + prompt cache    | \~\$4–8     | \~\$640–1,280   |

<Warning>
  These are illustrative estimates. Actual costs depend on your model choice, usage patterns, and cache hit rates. Check [real-time pricing](/api-reference/pricing/get-pricing) for current rates.
</Warning>

## Token Management Tips

### Set max\_tokens

Prevent runaway generation:

```json theme={null}
{
  "model": "claude-sonnet-4-6",
  "max_tokens": 4096,
  "messages": [...]
}
```

Most coding tasks need 1,000–4,000 output tokens. Setting a limit prevents the model from generating unnecessarily long responses.

### Use Auto-Compact

Most coding agents support context compaction — summarizing old conversation turns to reduce token count. Enable it:

* **Claude Code**: Built-in auto-compact triggers at context limits
* **Cursor**: Automatic context management
* **Codex CLI**: Use `--max-context` flag

### Avoid Context Bloat

* Don't paste entire files when a function is enough
* Use `.gitignore`-style patterns to exclude irrelevant files from agent context
* Clear conversation history when switching tasks

## Quick Configuration

Each tool needs just a few lines to connect through TokenLab:

<AccordionGroup>
  <Accordion title="Claude Code">
    ```bash theme={null}
    export ANTHROPIC_API_KEY="sk-your-tokenlab-key"
    export ANTHROPIC_BASE_URL="https://api.tokenlab.sh"
    ```

    [Full setup guide →](/integrations/claude-code)
  </Accordion>

  <Accordion title="Cursor">
    Settings → Models → OpenAI API Key: `sk-your-key`, Base URL: `https://api.tokenlab.sh/v1`

    [Full setup guide →](/integrations/cursor)
  </Accordion>

  <Accordion title="Codex CLI">
    ```bash theme={null}
    export OPENAI_API_KEY="sk-your-tokenlab-key"
    export OPENAI_BASE_URL="https://api.tokenlab.sh/v1"
    ```

    [Full setup guide →](/integrations/codex-cli)
  </Accordion>

  <Accordion title="Gemini CLI">
    ```bash theme={null}
    export GEMINI_API_KEY="sk-your-tokenlab-key"
    export GOOGLE_GEMINI_BASE_URL="https://api.tokenlab.sh"
    ```

    [Full setup guide →](/integrations/gemini-cli)
  </Accordion>
</AccordionGroup>