GPT-5.6 Cost per Completed Task: Luna vs Terra vs Sol

Most comparisons stop at $/1M tokens. Operators care about cost per finished job — a monitor that returns structured output, a classifier that ships JSON, a duty that completes. This post is D4 (tier cost-per-task) plus E3 (we actually measured). Sibling context: Sol vs Terra vs Luna · Luna guide.

Methodology

Date: 2026-07-31. Endpoint: OpenAI /v1/chat/completions. Parameter: reasoning_effort: "none" on all three (required for fair Chat Completions parity with tool-capable stacks). List prices used for cost: Luna $0.20/$1.20 · Terra $2/$12 · Sol $5/$30 per 1M.

• Task A — competitor_cheapest_json: three named prices → JSON {cheapest, spread_usd}
• Task B — sku_monitor_40_rows: 40-row SKU snapshot + specialist system prompt → JSON with cheapest SKU, vendor, price, spread, n
• Success: parseable JSON matching the ground-truth cheapest row / spread
• Not in scope: browser tools, multi-round agents, long-context needle tests

Raw run log lives in-repo as sites/cloudaxis/content/gpt-5-6-cost-per-task-measurement.mjs so others can reproduce or challenge the arithmetic.

Results

Task B (SKU monitor) — 668 input / 34 output tokens on every model:

• Luna — $0.000174 · 1403 ms
• Terra — $0.001744 · 807 ms · 10.0× Luna
• Sol — $0.004360 · 1263 ms · 25.0× Luna

Task A (tiny JSON) — 47 / 15 tokens: Luna $0.0000274 · Terra $0.000274 · Sol $0.000685. Same 10× / 25× pattern.

All six runs (3 models × 2 tasks) returned the correct answer. On these tasks, paying for Terra or Sol bought latency differences, not accuracy.

Scale it to a duty

If a scheduled monitor looks like Task B and runs 30×/day (~900/month) with the same token shape:

• Luna: ~$0.16/mo API list
• Terra: ~$1.57/mo
• Sol: ~$3.92/mo

Real CloudAxis duties add system prompts, tool schemas, and browser rounds — often multi-thousand prompt tokens. The ratio still tracks list prices when token counts stay similar. Where Sol finishes in one try and Luna needs three, Sol can win on cost-per-success — measure retries, not just first-pass tokens.

What this means for Auto Mode

CloudAxis Auto Mode is fixed to GPT-5.6 Luna on every plan including Free (landing). For high-volume completed tasks that Luna already solves correctly, that is the rational default on list economics. Escalate off-platform (or wait for product capacity features) when your evals show Luna failing tasks Terra/Sol clear — not because the $/1M table felt scary.

Limitations (read before citing)

Be precise when you link this:

• Two synthetic completed tasks — not a full agent harness with Browserbase
• Single run per model/task (no variance bands)
• reasoning_effort none — higher effort changes tokens and quality
• List prices, not your negotiated or cached-effective rates

Frequently asked questions

How much does a GPT-5.6 Luna completed task cost vs Terra?

On our 2026-07-31 SKU-monitor microbench with identical tokens, Luna was about $0.000174 and Terra about $0.001744 — roughly 10×, matching list price ratios.

Did Sol get the answer more often?

Not on these two tasks. All three models returned the same correct JSON. Sol cost ~25× Luna for the same tokens.

Does this include browser tool calls?

No. This microbench is Chat Completions only. Tool-heavy agent runs use more tokens; re-measure on your harness.

Can I reproduce the numbers?

Yes. Use the same prompts with reasoning_effort none, record usage.prompt_tokens and completion_tokens, multiply by list $/1M. Our run log is in the CloudAxis content repo measurement module dated 2026-07-31.