mergisi Model Cost Compare
Given a task description (and optionally a rough prompt / input size), estimate the cost of running it on each available model tier and recommend the cheapest one that can actually do the job.
When to use
- "Which model should I use for X?"
- "Is it worth running this on Opus or will Sonnet do?"
- "Can I offload this to a local model?"
- "/model-cost-compare — classify 10k support tickets"
Pricing table (indicative — always flag as "check provider docs")
Use these rough figures. They are not exact; confirm before quoting real numbers to the user.
| Model | Tier | Input ($/1M tok) | Output ($/1M tok) | Context | Strengths |
|---|---|---|---|---|---|
| Opus 4.6 (1M) | Frontier | ~$15 | ~$75 | 1M | Agentic, long-context, hard reasoning |
| Sonnet 4.6 | Mid | ~$3 | ~$15 | 400k | Everyday coding, agents, drafting |
| GLM-5.1 | Budget hosted | ~$0.60 | ~$2.20 | 256k | Cheap bulk work, decent reasoning |
| Minimax M2.7 | Budget hosted | ~$0.40 | ~$1.80 | 256k | Very cheap, OK for templated output |
| Gemma 4 (local) | On device | $0 marginal | $0 marginal | 32k | Free but slow, weak at multi-step logic |
Indicative pricing as of OpenClaw 2026.4.11. Check the provider docs before billing decisions.
Instructions
- Parse the user's task. Extract:
- Task type: reasoning, extraction, classification, drafting, translation, agentic tool use, long-context synthesis.
- Input size estimate: in tokens. If the user says "10k tickets averaging 500 tokens", that's 5M input tokens. If unknown, ask for a rough size.
- Output size estimate: short label? full essay? JSON record?
- Volume: one-off or batch?
- Rule out incapable models. Use this capability floor:
- Agentic multi-tool flows with long reasoning → Opus or Sonnet only.
- Structured extraction / classification with clear schema → any tier, including Gemma 4 local.
- Long-context synthesis (>400k tokens) → Opus only.
- Privacy-sensitive data that cannot leave the machine → Gemma 4 local only.
- For each surviving model, compute:
Multiply by volume. Show your arithmetic so the user can sanity-check.cost = (input_tokens / 1_000_000) * input_price + (output_tokens / 1_000_000) * output_price - Print the comparison as a Markdown table sorted cheapest first. Bold the recommended row.
- End with a one-line recommendation:
Recommended: <model> — <1-sentence reason>.
Output example
Input: "Classify 10,000 customer support emails into 5 categories. Avg 400 input tokens, 20 output tokens."
Total tokens: 4M input, 200k output
| Model | Input cost | Output cost | Total | Capable? |
|-------------|-----------:|------------:|--------:|---------:|
| **Gemma 4** | $0.00 | $0.00 | $0.00 | yes |
| Minimax M2.7| $1.60 | $0.36 | $1.96 | yes |
| GLM-5.1 | $2.40 | $0.44 | $2.84 | yes |
| Sonnet 4.6 | $12.00 | $3.00 | $15.00 | yes |
| Opus 4.6 | $60.00 | $15.00 | $75.00 | overkill |
Recommended: Gemma 4 local — classification with a fixed 5-label schema is trivial for on-device models and costs nothing.
Anti-patterns
- Don't recommend Opus by default "just to be safe". If Sonnet can do it, say so.
- Don't recommend a local model for agentic tool-use loops — they spiral.
- Always flag prices as approximate.
Example invocations
/model-cost-compare classify 10k support tickets into 5 buckets- "What's the cheapest model that can draft 200 SEO meta descriptions a day?"
- "Should I use Opus or Sonnet for this 800k-token codebase review?"