TokenAtlasWhat you can model
Multi-step agents, fine-tuned models, RAG pipelines, voice transcription + LLM chains, image generation, and hybrid OpenAI/Anthropic routing.
Reusable scenarios
Save scenarios with a name and tag (e.g. 'support-bot-v2'). Re-run them when prices change to see drift.
Sensitivity analysis
Vary one input (volume, prompt length, model) and watch the cost curve update. Identifies which lever moves your bill the most.
Worked example: a multi-step agent
A three-step agent — retrieval (1,500 in / 0 out), reasoning (2,000 in / 500 out on GPT-4.1), and drafting (1,000 in / 800 out on Claude 3.5 Sonnet) — run 20,000 times/month: reasoning step is 40M in / 10M out on GPT-4.1 = $80 + $80 = $160; drafting step is 20M in / 16M out on Sonnet = $60 + $240 = $300. Modelling each step on its own model, rather than one blended rate, is what a custom estimate is for.
Cost drivers ranked in multi-model pipelines
In a chained workload, the step with the highest output volume usually dominates cost, not the step with the most calls. Rank each step by its modelled monthly total, not by how central it feels to the pipeline — a short drafting step on an expensive model can outweigh a long retrieval step on a cheap one.
Reading a custom scenario correctly
A custom estimate is a sum of per-step scenarios you defined, built from token volumes and rates you entered for each step — it is not a trace of an actual running pipeline. If the real pipeline retries a step, skips a step conditionally, or truncates output, the estimate will diverge from the eventual bill unless those behaviours are modelled explicitly as separate scenarios.
Assumptions checklist for a custom scenario
— Which model is assigned to each step, and at what input/output split — Monthly call volume per step, not just for the pipeline as a whole — Whether steps run unconditionally or are skipped/retried under some conditions — Whether output length is capped or variable, since variable output is the hardest input to estimate accurately
Translating a pipeline estimate into cost per completed run
Summing all step costs and dividing by monthly pipeline runs gives a cost-per-completion figure: the worked three-step agent on this page totals $460/month at 20,000 runs, or $0.023 per completed run. That per-run number is what belongs in a unit-economics model, since it stays stable even as the monthly run count changes, unlike the aggregate monthly total.
Failure mode: modelling the happy path only
A custom scenario built from the intended flow of a pipeline will underestimate cost if the real system retries failed steps, re-runs on validation failure, or falls back to a second model on error. Add a retry-rate assumption (e.g. 10% of runs retry one step) as a separate line in the scenario rather than folding it into the base call volume, so the two effects stay visible.
Frequently asked questions
- Can I add private models?
- Yes — define a custom rate card on the Team plan.
- Does it support multi-provider routing?
- Yes — assign a model per step in an agent chain.
- Can scenarios be shared with my team?
- Yes — workspace-shared scenarios are included on Team.