# Results

Every answered run through Recipe360 was correct: ten models from three providers, in three agent programs,
produced the same deliverables byte for byte, with one single fingerprint.

## What we repeated

We repeated our internal tests through Recipe360 itself. The agent had Recipe360's MCP server and its skill,
and nothing else from us. Every run got the same 13-step workflow, the same data and the same request.

| Agent program | Provider | Model | Correct | Median time per run | Median tokens per run | Closing word for word |
| --- | --- | --- | --- | --- | --- | --- |
| Claude Code | Anthropic | Claude Haiku 5.5 | 100% | 12 s | 143k | 70% |
| Claude Code | Anthropic | Claude Sonnet 5.5 | 100% | 12 s | 138k | 100% |
| Claude Code | Anthropic | Claude Opus 5.5 | 100% | 18 s | 178k | 90% |
| Codex | OpenAI | gpt-6.1-sol | 100% | 14 s | 89k | 100% |
| Codex | OpenAI | gpt-6-astra | 100% | 16 s | 76k | 100% |
| Codex | OpenAI | gpt-6-luna | 100% | 12 s | 87k | 100% |
| Hermes | NVIDIA | Nemotron 3.5 Lightning 30B | 100% | 23 s | 44k | 74% |
| Hermes | NVIDIA | gpt-oss-20b | 100% | 32 s | 40k | 54% |
| Hermes | NVIDIA | MiniMax M3.1 Flash | 100% | 91 s | not measured | 100% |
| Hermes | NVIDIA | Nemotron 3 Ultra 550B | 100% | 30 s | 44k | 100% |

- Every run called the recipe once and finished its 13 steps in that call.
- When a model did not repeat the closing message word for word, it reworded it; the deliverables were correct
  either way.
- Tokens follow the agent program more than the model: most of what a run processes is the agent program's own
  instructions, sent with every request. Compare models within one program, not across programs.
- Cost per run, as Claude Code estimates it: US$0.004 with Haiku 5.5, US$0.058 with Sonnet 5.5 and US$0.138 with
  Opus 5.5.

## Against working without a recipe

In our earlier internal tests, runs with and without a recipe used the same agent setup, so the comparison is
fair. With a recipe, runs processed 62% to 93% fewer tokens, for an estimated 43% to 97% lower API cost, and
finished 4 to 13 times faster. With a recipe, 100% of runs were correct; without one, 39%, and only two models
were reliable without a recipe. Without a recipe, even those two took several different paths, so no two runs
could be audited the same way.

## How we measured

- **Correct** means every deliverable is identical, byte for byte, to an independently written reference, and
  the delivery fingerprint equals the reference's.
- **Tokens** count everything the model processed, cached tokens included, as each agent program reports it.
  The cost estimate assumes cached input at 10% of the input price and output at 8 times it.
- **Time** counts the whole run, from the request to the answer.
- When a provider refused a request for its rate limit, the model never answered; such runs are left out.
- These are internal tests. The saving and the speed-up depend on the workflow, the model and the agent
  program, so each customer's will differ. Only results are published, never the data behind them.

## Check it yourself

A [public benchmark](benchmark.md) lets anyone measure the same question, with and without a recipe, on invented
data, with an independent reference and exact scoring.