Methodology
Last updated 2026-07-14.
The cost model
Every number on this site is produced by one formula:
monthly_cost = requests × calls_per_request × (
input_tokens × [(1 − cache_hit) × input_price + cache_hit × cached_price]
+ output_tokens × output_price
) ÷ 1,000,000 × batch_factor
- calls_per_request — agent chains trigger 3–10 internal LLM calls per user request (planning, tool use, validation). A single-pass chatbot is 1.
- cache_hit — the share of input tokens served from the provider's prompt cache. Cached reads are billed at ~10% of the input price on OpenAI, Anthropic and Google, and ~1-2% on DeepSeek.
- batch_factor — 0.5 when the provider's batch API is used (OpenAI, Anthropic, Google); 1.0 otherwise. On providers that support it, the batch discount stacks with cache reads (Anthropic: 50% × 10% = 5% of list input price).
The agent cost model
The agent cost calculator extends the formula above. One attempt is a full run of the agent chain (steps × the per-call cost above); failed attempts are retried from scratch up to a maximum:
task_success = 1 − (1 − p)n
expected_attempts = task_success ÷ p
cost_per_successful_task = cost_per_attempt × expected_attempts ÷ task_success
monthly_total = tasks × expected_attempts × cost_per_attempt + escalated_tasks × human_minutes ÷ 60 × hourly_rate
- p — per-attempt success rate, measured on your own eval set; n — max attempts (1 = no retries).
- Retries re-run the whole chain. Frameworks that resume from the failed step pay less than this model predicts — treat our number as an upper bound in that case.
- Escalated tasks — the (1 − task_success) share that exhausts all attempts and goes to a human. Human fallback is priced in minutes × hourly rate and is often the largest line item.
- Same success rate across models in comparison tables — reliability does not transfer between models, so measure before switching.
The per-customer cost model
The per-customer calculator splits the average usage into power users (share s at k× regular usage) and regular users so the blended mean stays at the input value:
regular_requests = avg_requests ÷ (1 + s × (k − 1))
power_user_cost = k × regular_requests × cost_per_request
break_even_requests = subscription_price ÷ cost_per_request
- s, k defaults (10%, 7x) — a common SaaS usage skew; measure your own distribution once you have per-customer metering.
- Batch API is not applied — interactive SaaS traffic is real-time.
- This is a planning estimate, not attribution: knowing which real account consumed what requires tagging each LLM call with a customer ID and aggregating.
Use-case unit conversions
The use-case picker prices work in natural units by converting each unit to tokens, then applying the formula above. The assumptions (editable in the calculator):
- Document extraction / OCR — 1 page ≈ 1,250 input tokens (a scanned page image ≈ 1,100 tokens + 150 tokens of instructions) and 650 output tokens of structured text; 10% cache hit; batch API where available. Page images are priced at standard input-token rates — actual image tokenization varies by provider and resolution, and document/vision input support must be verified per model.
- Summarization — 1 article ≈ 3,200 input tokens (~2,200 words + instructions) and 300 output tokens (~200-word summary); 5% cache hit; batch API where available.
- RAG generation — 1 query ≈ 3,500 input tokens (system prompt 400 + retrieved chunks 2,800 + question 300) and 400 output tokens; 60% cache hit (system prompt and hot chunks repeat); real-time. Embedding and vector-store costs are excluded.
Data sources
Prices are taken exclusively from official provider pricing pages and re-verified on the date shown on each tool:
We track 13 production models across OpenAI, Anthropic, DeepSeek and Google. Preview and limited-availability models are excluded until they carry stable public pricing.
Known simplifications
The model intentionally omits charges that are workload-specific. If any of these apply to you, treat our numbers as a lower bound on the cost delta:
- Cache-write fees — Anthropic bills cache writes at 1.25× (5-minute TTL) or 2× (1-hour TTL) the input price. Cache-churn-heavy workloads pay materially more than the read-price suggests.
- Cache storage fees — Google bills context-cache storage per token-hour ($4.50/1M tokens/hour on Pro models) on top of the cache-read price.
- Long-context surcharges — Gemini 3.1 Pro doubles input pricing above 200K tokens.
- Introductory pricing — Claude Sonnet 5 is $2/$10 through 2026-08-31, then $3/$15. We list current effective prices.
- Non-token costs — hosting, vector storage and observability are out of scope; they routinely exceed token spend at production scale. Human review is modeled only in the agent cost calculator (minutes × hourly rate for escalated tasks).
Why we built this
Public answers to "how much does it cost to switch LLM providers" are dominated by wide, non-reproducible ranges. Our position: cost estimates are only useful when they take a workload profile as input — request volume, token mix, cache hit rate, agent chain depth and batch eligibility. That is exactly what the migration savings calculator does, and every worked example on this site is generated by the same open formula above.
Corrections
Found a price that moved before we did? Open an issue or email us — pricing corrections ship the same day.