Orbi

Blog · 2026-09-24

DeepSeek coding agent cost per merged PR

See how Orbi uses DeepSeek as an unattended coding agent, what each merged PR costs, why cache reads matter, and how to connect your own API key in practice.

Orbi turns a GitHub Issue into a reviewed, merged PR and a tagged release. From September 22 to 24, 2026, Orbi's own repository merged 20 PRs using deepseek-flash for the whole loop: writing code, independent review, fixes, merge, and release. You can inspect issue #1328, issue #1332, and issue #1336.

This is a cost report, not a model leaderboard. The unit here is one merged PR. The numbers come from the same dataset as the site's full cost table.

What one merged PR cost

Across those runs, the median was 10.6M tokens per merged PR. At DeepSeek's official prices, checked on September 24, 2026, that works out to:

The figures use DeepSeek's official pricing. Peak prices are double the off-peak rates during weekdays 01:00–04:00 and 06:00–10:00 UTC. For an unattended coding agent, scheduling around those windows is a practical cost control. The auto-merge AI PR guide explains the delivery loop; this post puts a price on its merged-PR unit.

Why the number stays low

Of the tokens in these runs, 97.1% were cache reads. A cache hit costs $0.003 per one million tokens; a cache miss costs $0.15 per one million. That is 1/50 of the uncached price. Orbi's session structure repeatedly reuses the same context, so the context needed for an independent review and its fixes is not paid for as a fresh read each time.

One caveat matters: when the cache-hit rate drops below 90%, cost rises quickly. The median is useful for planning, but your own context shape and review rounds determine the result.

Connect DeepSeek to a self-hosted Orbi

Orbi runs on Pi. Its engine can use an OpenAI-compatible API or a Codex subscription; it does not support Claude Code. Orbi is open source under AGPL-3.0.

The provider template in templates/pi-providers/deepseek.json is:

{
  "providers": {
    "deepseek": {
      "baseUrl": "https://api.deepseek.com/v1",
      "api": "openai-completions",
      "apiKey": "$DEEPSEEK_API_KEY",
      "models": [
        {"id": "deepseek-flash", "name": "DeepSeek Flash", "contextWindow": 131072, "maxTokens": 16384}
      ]
    }
  }
}

Set the key before starting the agent:

export DEEPSEEK_API_KEY="your-key"

Use the provider documentation for the rest of the self-hosted setup. Never put the key in a checked-in config file.

What to take away

Measure cost per merged PR, keep the cache-hit rate above 90%, and move unattended work outside peak hours when that fits your queue. The important unit is not a model call in isolation: it is the complete path from GitHub Issue to PR, independent review, merge, and tagged release. Orbi's own 20-PR sample makes that path inspectable, and the cost page keeps the full table in one place.

Related

Read the cost table, Cloud, and auto-merge AI PR guide.