We ran the same complex task on Opus and Haiku side by side. Real tokens, real dollars, real results. Here's exactly what happened.
Task: Generate a board intelligence report on Ramp (B2B fintech, ~$750M ARR, Series D). The kind of analysis a VC or board member reviews before a quarterly meeting. Five data points, risk scoring, executive summary, recommended actions.
The question: Can Haiku (at $0.80/M tokens) do the same job as Opus (at $15/M tokens) if given proper documentation?
The answer: Yes — if the workflow is designed first by Opus and stored in AgentMemo.
We asked Opus one question: "Design this workflow so completely that a simpler model can execute it mechanically forever."
Opus documented everything. Which five data points matter for a fintech board report and why. Exact risk thresholds (ARR growth below 30% = score 4, above 50% = score 2). How to handle missing data. The exact output format. Every edge case.
That documentation was saved to AgentMemo as a workflow. This happened once.
Haiku loaded the workflow from AgentMemo and followed the blueprint. No reasoning about what metrics matter. No deciding on output format. Just execution — filling in the documented template with real data about Ramp.
The report was accurate. Haiku correctly identified Ramp's growth trajectory, financing context, and competitive position — because Opus had already defined exactly what to look for and how to evaluate it.
| Design (run 1) | $0.0920 |
| Run 2 | $0.0370 |
| Run 3–30 | $1.0730 |
| 30-day total | $1.2031 |
| Design (Opus, once) | $0.0920 |
| Run 2–30 (Haiku 3) | $0.0174 |
| - | |
| 30-day total | $0.1097 |
| Execution Model | Cost/Run | vs Opus | 30-Day Total |
|---|---|---|---|
| Opus (no AgentMemo) | $0.0370 | — | $1.2031 |
| Haiku 4 via AgentMemo | $0.0020 | 18.8x cheaper | $0.1513 (87% off) |
| Haiku 3 via AgentMemo | $0.0006 | 64x cheaper | $0.1097 (91% off) |
| Open source via AgentMemo | $0.0000 | ∞ | $0.0920 (design fee only) |
After proving the Haiku 4 result, we asked: could an even cheaper model follow the same Opus blueprint?
We ran the identical workflow through Claude Haiku 3 — a 2024 model priced at $0.25/M input (vs $0.80/M for Haiku 4). It executed in 3.4 seconds for $0.0006. It followed the workflow. It applied the scoring rubric correctly. It produced a usable board intelligence report.
The result flips the mental model: model intelligence is not the bottleneck. Workflow quality is. A detailed enough Opus blueprint makes cheap models produce expensive-model results. The smarter you make the blueprint, the further down the cost ladder you can push execution — all the way to open source at $0.00 per run.
The insight is simple but profound: intelligence and execution are different jobs.
When you run Opus on a repeated task, you're paying for intelligence you already bought. Opus re-derives the same reasoning, re-decides the same structure, re-generates the same framework — every single time.
AgentMemo separates the jobs. Opus does the hard thinking once. The result — the workflow — becomes a permanent asset. Every future run just executes the asset. Haiku is perfect for this: fast, cheap, reliable at following instructions.
As you run more executions, the average cost per report drops toward the Haiku rate. The Opus investment amortizes across every run.
Start your 7-day free trial. Build your first workflow with Opus. Run it on Haiku. Watch the savings add up in your dashboard.
Start Free Trial →