Key facts
| Tool type | Free production cost estimator across model tiers |
| Inputs | Requests per month, input and output tokens, retries, model tier |
| Outputs | Estimated monthly spend per model and per workload |
| Cost drivers | Output tokens, agent loop iterations and retries dominate |
| Pricing comparison | Per-token estimates versus flat monthly self-serve plans |
| Companion tools | Token calculator and GPU capacity calculator |
| Models | 30+ models from free to frontier tiers |
| Product status | Live |
TL;DR
- Measure tokens before you estimate: real samples beat rules of thumb.
- Output tokens and retries are the hidden multipliers in any estimate.
- Agent loops multiply every step's tokens, so model each iteration explicitly.
- Flat monthly self-serve plans remove per-token variance on self-serve workloads.
- Re-run the numbers monthly; workloads drift, and so do estimates.
How it works, step by step
- List the workloads separately: chat, extraction, summarisation, RAG, agents, batch.
- For each, measure average input and output tokens on real samples with the token calculator.
- Estimate monthly request volume and peak concurrency from analytics or product plans.
- Open the LLM cost calculator and enter the figures per workload and model tier.
- Add realistic retry rates and agent loop iterations rather than assuming one call per task.
- Compare the per-token total with flat monthly self-serve plans on the live pricing page.
- Re-estimate monthly and after any prompt, model or architecture change.
Try it yourself
Open the LLM cost calculator →
What actually drives the bill
Four variables decide almost every estimate: volume, input tokens, output tokens and model tier. Output tokens usually cost several times more than input tokens, so a concise answer policy often saves more than a cheaper model. Retries are the quiet multiplier — timeouts, malformed JSON and rate limits all trigger extra full-price calls. Agent loops multiply everything: a five-step task sends the growing conversation five times. A useful estimate models the workload as it behaves, not as the happy path describes it.
Reading the estimate honestly
Treat the calculator output as a model, not a promise. It is most reliable when the token counts come from real samples and the retry rate comes from logs. If you have no production data yet, run a one-week pilot and instrument token usage per request before forecasting. Watch the shape of the result as well as the total: a workload whose cost scales linearly with volume is a different risk from an agent whose cost scales with steps and context growth. The shape tells you what to re-architect; the total tells you what to budget.
Choosing between per-token and flat pricing
Per-token billing is proportional but unpredictable: a viral day or a runaway agent loop shows up directly on the invoice. Flat monthly self-serve plans invert that: cost is stable and iteration is cheap, which suits teams that experiment and scale gradually. Self-hosting moves the cost to GPUs and operations, and only wins at sustained, high utilization. Run each workload through the calculator twice — once per-token, once against the live pricing page — and pick per workload, not for the whole product at once.
Honest comparison
| Pricing model | Predictability | Best for | Main risk |
|---|---|---|---|
| Per-token billing | Low, scales with usage | Sporadic or very low volume | Spikes and runaway loops |
| Flat monthly self-serve | High, fixed cost | Steady product workloads and iteration | Fair-use policy limits |
| Committed enterprise | High, negotiated | Large predictable volumes | Forecast accuracy |
| Self-hosted GPUs | Medium, fixed plus utilization | Sustained high utilization | Idle time and operations burden |
Frequently asked questions
What inputs does the cost calculator need?
Requests per month, average input and output tokens, retry rate and model tier. Use the token calculator to measure token counts from real samples first.
Why are output tokens so important?
Output tokens are usually priced higher than input tokens, and they grow with verbosity. A prompt change that shortens answers often cuts cost more than switching models.
How do I estimate retries?
Use production logs if you have them; otherwise start with a conservative rate based on timeout and JSON-parse failures and adjust after a pilot.
Does the calculator include agent loops?
You model them as additional calls per task. Each loop iteration re-sends a growing conversation, so multiply by expected steps and context growth.
Are the estimates exact?
No. They are directional models based on your inputs. Validate against real usage after launch and re-run the numbers monthly.
How does flat pricing change the picture?
Flat monthly self-serve plans make cost independent of token volume, which removes spike risk and makes iteration cheap. See the live pricing page for current plans.
Can I compare self-hosting?
Yes. Model the GPU profile with the GPU capacity calculator, add operations and utilization, and compare it with the managed total.
Is there a free way to pilot?
Yes. The free plan includes 2 free AI models (plugsky-micro and plugsky-lite) with no card, so you can collect real token data before paying.