Spend less on AI without gambling on quality
Compare AI models and provider prices, estimate your batch cost, and pay only for the work that runs.
- No signup to get a price
- No subscriptions or minimums
- No provider lock-in
Where the money goes
Illustrative · per $100
- Standard real-time spendThe same workload at the reference rate.
- $100.00
- Provider batch pricingAssume a 50% lower batch rate.
- $50.00−50%
- A lower eligible provider costAssume a 12% lower base price.
- $44.00−12%
- Native-provider discount10% off the provider cost.
- $39.60−10%
- Batchrouter service fee5.90% on the $44 prediscount base.
- +$2.60+5.90%
- Illustrative token total
- $42.20
- OpenAI and Anthropic batch rates vs. standard, same models
- ≈50%
- Further off native OpenAI and Anthropic lanes
- −10%
- Estimates reserve; the receipt charges what ran
- Actual usage
- Service fee refunded if we miss the window
- Standard 24h
Where the savings come from
Batch pricing, provider choice and actual-usage settlement. Understand the levers before you move your workload.
Batch what doesn't need an answer this second
Nightly enrichment, document pipelines, backfills and labelling can run as batches. Use published batch pricing when your workload can wait.
Compare providers on every run, not once
Provider prices vary, and the cheaper eligible lane can change from run to run. Wire up a single provider and you are locked to whatever it charges. We re-check every batch.
Only count savings that survive your quality bar
A cheaper model that gets it wrong is not cheaper. Compare candidates on your own data before you move production work. Automated evaluation sweeps are coming soon.
Pay for what actually ran
You see the cost before work starts, and we settle against real usage at the end. Anything reserved and unspent goes straight back. No subscriptions and no minimums.
OpenAI and Anthropic publish batch prices at 50% below their standard API prices. Model eligibility and current rates still matter.
Work that can wait a few hours, and should cost less for it
If nobody is watching a spinner for the answer, you can trade immediate delivery for a batch price. Explore a worked example, then price your own volume from the published catalogue.
Illustrative case · E-commerce
Nightly catalog enrichment
A marketplace receives thousands of new listings a day with thin titles, no attributes, and inconsistent categories.
- The job
- Rewrite titles, extract size, colour and material, and assign a category for every new listing.
- Why it can wait
- Listings go live the next morning. Nobody is waiting on a single row, so there is no reason to pay for instant answers.
- Example quality target
- Category matches a human reviewer on at least 95% of a labelled sample.
- Example volume
- ≈120k listings per night
These are example targets to test on your own data, not measured results. Automated evaluation sweeps are coming soon.
Illustrative monthly cost
E-commerce- Standard real-time spend
- $9,600.00
- Batched and routed spend
- $4,050.00
Where each $100 goes in the worked example
Hypothetical amounts, not customer results or a quote. Assumes a 50% provider batch reduction, a hypothetical 12% lower eligible provider base cost, a 10% native-provider discount and a 5.90% service fee charged on provider cost before that discount. Lane fees and tool charges are excluded. Actual costs depend on the model, volume and current published rates.
The $100 worked example
- Standard real-time spend
- $100.00
- Provider batch cost
- $50.00
- Hypothetical lower eligible provider cost
- $44.00
- After native-provider discount
- $39.60
- Service fee on provider cost before discount
- $2.60
- Illustrative token total
- $42.20
The test is simple
A good batch workload has three things in common:
- Many items, the same instructions on each one.
- The result is read later: a table, an index, a report, a queue.
- You can say what a good answer looks like.
Keep these on Realtime
Use immediate responses when someone is waiting:
- A chat reply a person is waiting on
- Autocomplete or in-product suggestions
- A voice or support agent mid-conversation
- Anything that must answer in seconds
The Realtime tier answers through POST /v1/chat/completions, best effort, at the lane’s direct price.
Cheap only counts if it is good enough
Public benchmarks are run on someone else's data. Score the candidates on yours, set the bar where your business needs it, and let price decide among the models that cleared it.
- Bring a sample of your own work, and what a good answer looks like.
- Every candidate provider runs it at once — one job, one price, one settlement.
- Grading happens on our side and costs you nothing. You pay only for the inference.
- A model has to prove it clears your bar, not merely average out at it.
Scorecard
Your bar: 0.92| Model | Score | Cost / 1k | Verdict |
|---|---|---|---|
| Sonnet-class | 0.96 | $2.41 | Passes — 2.8× the price |
| Haiku-class | 0.94 | $0.86 | Cheapest passing — routed here |
| Mini-class | 0.91 | $0.71 | Below your bar |
| Open-weights 70B | 0.78 | $0.34 | Below your bar |
The cheapest model on the board is not the one we pick. The cheapest one that cleared your bar is.
Illustrative scorecard — your models, scores, and costs come from your own run.
Start with a price, not a migration
Size the savings before you talk to anyone, and prove the quality before you move a production job.
Price it
Put your real volume into the calculator and see what the work would cost across providers. No account, no card.
Prove it (coming soon)
Sweeps, which run one sample of your own data on many models at once, are coming soon. Until then, compare providers on price and speed in the rankings.
Route it
Choose your model and routing requirements. Each batch goes to an eligible provider, delivered within the Standard tier's 24 hours. Automated quality evaluation sweeps are coming soon.
Wiring it into your own systems? The documentation covers the API, file formats, webhooks and moving over from an existing batch setup.
Read the documentationDirect chat completions
The standard OpenAI request body, answered right away. Set "stream": true to get the reply as server-sent events.
POST /v1/chat/completions
One request, queued as a batch
The same request, run as a one-item batch. You get 202 and a job id right away, then read the result from GET /v1/chat/completions/{id} or a webhook.
POST /v1/chat/completions with "batch": true
Bulk batches
Many requests at once, from the native API or a JSONL file. An existing OpenAI Batch client works with a base-URL swap.
POST /v1/batches, or /v1/openai/v1 for OpenAI Batch clients
Published tiers are Realtime (best effort), Priority (4h), Standard (24h), Relaxed (48h), Flex (72h) and Economy (7 days). The API reports which this environment offers.
Current fees and SLA availabilitySee what is available today
Browse published models and compare their provider options. Open a model for pricing, capabilities and lane status.
Find out what you would actually save
Price your real volume in under a minute. No account, no card, no call.