Pricing calculator
Calculate the cheapest AI route.
Describe the workload, set the few numbers that move the bill, then compare API, batch, subscription, hybrid routing, local-model, and GPU-hosted options.
Recommended route
Hybrid router is the safest default for this workload at $222/month.
Start with Hybrid router: Route routine inference to open models and escalate only the expensive judgment calls.
$177/mo over the cheapest lane buys routing controls and a small frontier reserve.
Estimate basis
Uses public list prices and the assumptions you supplied. Treat marketplace GPU, cloud procurement, and custom enterprise plans as manual-pricing checks before production.
Recommended
$222
Hybrid router
Cheapest
$45.05
Open-model API
Yearly
$2,663
Hybrid router
Why this works
- Hybrid router gives the best balance of cost and quality for this workload.
- Routine inference moves to cheaper open-model or batch lanes instead of paying frontier rates on every request.
- The modeled cache assumption is 86%, so repeated context should reduce input cost if the app preserves cacheable prompts.
Avoid first
- Frontier API only: $748/month. Too expensive unless every output directly changes a high-value decision.
- Self-hosted H100: $692/month. Use only after utilization tests prove the GPU stays busy.
Keep these assumptions when you share this pricing run.
Compare routes
Monthly cost by strategy
Lower is better only after the quality bar is met.
Open-model API
Cheapest laneCheapest managed lane for bulk extraction, summarization, triage, and routine agent loops.
$45.05/mo
Managed cloud AI
Manual pricingAWS, Google Vertex AI, and Azure AI Foundry path for enterprise controls, region choices, and cloud procurement.
$76.07/mo
Batch API
Usable laneNon-urgent processing lane for documents, logs, offline analysis, and backfills.
$98.70/mo
Subscription reserve
Human reserveHuman-in-the-loop coding and review budget before buying heavy API volume.
$165/mo
Hybrid router
Best balanceDefault production strategy: cheap routine lane plus small frontier reserve.
$222/mo
AI Gateway router
Usable laneVercel AI Gateway control plane for provider routing, fallbacks, and observability with provider pass-through pricing.
$222/mo
Local inference and fine-tuning
Benchmark local models before you buy dedicated GPU capacity.
These gates keep tuning and self-hosting tied to eval quality, accepted outputs, and real utilization.
Local inference
Benchmark local inference before renting dedicated GPUs.
Start with an open-model API baseline so local inference has a real cost and quality target.
Benchmark before GPU rental
Fine-tuning
Tune only after routing and retrieval fail
Create a golden eval set from repeated accepted and rejected examples before changing weights.
Eval set required
Self-hosting
$1,436/mo H100 baseline
Move steady traffic to rented GPUs only when utilization, storage, egress, and operations beat API pricing.
Idle time decides the bill
Inputs and break-even checkssoftware development
Workload assumptions
Type
software development
Input
650.0M tokens/mo
Output
32.0M tokens/mo
API/tools
0/mo
Cache
86%
GPU hrs
120/mo
Storage
40 GB
Egress
12 GB
Calculated from the fields on the left. Adjust the workload and recalculate to compare another route.
Self-hosting breakpoint
$1,436/month H100 baseline
Self-hosting uses the lowest verified H100 price currently loaded (RunPod H100 serverless worker). It usually beats frontier APIs before it beats low-cost open-model APIs; the GPU has to stay busy enough to overcome idle time, storage, egress, and operations overhead.
Cost componentsRecommended and self-hosted detail
Recommended components
Hybrid router
$222
Routine open-model lane
tokens
$40.55
90% routine share routed away from frontier models.
Official prices: deepseek-v4-flash
Frontier escalation lane
tokens
$61.38
10% hard-case share kept for frontier reasoning.
Official prices: anthropic-claude-sonnet-5
ChatGPT Plus with Codex
subscription
$20.00
Monthly subscription reserve; not a production API substitute.
Official prices: openai-chatgpt-plus
Claude Max 5x
subscription
$100
Monthly subscription reserve; not a production API substitute.
Official prices: anthropic-claude-max-5x
Self-hosted cost check
Self-hosted H100
$692
RunPod H100 serverless worker
gpu
$239
120 GPU-hours/month assumption.
Official prices: runpod-h100-serverless-hour
Storage
storage
$2.80
40 GB-month assumption.
Official prices: runpod-network-storage-standard
Network egress
egress
$0
12 GB egress assumption.
Official prices: vast-ai-internet-egress
Operations overhead
operations
$450
Conservative monthly allowance for monitoring, retries, orchestration, and maintenance.