GUIDE

GPT-5.6 Luna Agentic Harness Benchmark

Public agent leaderboards are noisy. We ran a small, owned harness: GPT-5.6 Luna must call tools to check price and inventory, then reorder under a written policy — measured on Hetzner with Chat Completions + function tools.

7–9 min read

Key takeaways
  • Harness: local tool executor + OpenAI Chat Completions (reasoning_effort: none)
  • Task: if ABC-9 inventory < 10 and price < 50 → submit_reorder(ABC-9, 25) → return confirmation JSON
  • Measured 2026-07-31: PASS · 3 tool calls · 1043 in / 99 out · $0.000327 · 3508 ms
  • Why it matters: proves Luna agent loops on the same stack we host — not a screenshot of someone else’s Terminal-Bench
  • Scope: one policy task, one run — extend the script for bake-offs; see also benchmark inflation

Why a tiny owned harness

METR and the SWE-bench Pro audit (see our inflation piece) are reminders that agent scores are harness-dependent. We already publish scheduling cadence stats from real CloudAxis duties (scheduling benchmark). This post adds a tool-loop microbench we can re-run after model or SDK changes.

Task and tools

Hidden world state: ABC-9 costs 42.5 with inventory 4. Tools: get_price, get_inventory, submit_reorder. The model must not invent numbers — only tool results count. Final answer must be JSON with the confirmation id.

Result

On 2026-07-31, gpt-5.6-luna completed the policy correctly:

Log: sites/cloudaxis/content/gpt-5-6-luna-agentic-harness-measurement.mjs. Runner: gpt-5-6-luna-agentic-harness.mjs (needs OPENAI_API_KEY; we executed inside the CloudyBot API container on Hetzner).

What this is not

Frequently asked questions

Is this a full agentic leaderboard?

No. It is a lean, owned tool-loop task we can re-run cheaply after model changes. Extend the script for broader bake-offs.

Did you use the Responses API?

This pass used Chat Completions with function tools and reasoning_effort set to none — the path our lean agent forces for GPT-5.6 + tools.

How much did the run cost?

About $0.000327 at Luna list rates for 1043/99 tokens.

Can I reproduce it?

Yes. Run sites/cloudaxis/content/gpt-5-6-luna-agentic-harness.mjs with OPENAI_API_KEY on a host that can reach api.openai.com.

Related reading in this series
Long-context recall test · Luna landing · Scheduling benchmark