The longer you can wait, the less you pay
Per million tokens read and written, by size and by deadline.
Two prices per size. Nothing hidden.
One price for what you send and one for what you get back, and one choice: how long you can wait.
Repeated input is free
When a prompt starts with text we have just processed, that part is served from cache and costs nothing. You pay for new input and for output, and nothing else.
No line items to decode
No surcharge for long prompts, no cache line items. What you send and what you get back, per million tokens, plus $0.002 per call when a workflow calls your own endpoint.
The deadline is the only dial
Each deadline is one multiple of Standard, so the whole card fits on one screen. There is no separate batch product to learn: every request is priced as batch work.
The rate card
US dollars per million tokens, input / output. Each row is a deadline; the crossed-out row is the model each size is compared with, at its regular price.
| Tier | Small otium-small | Medium otium-medium | Large Planned otium-large |
|---|---|---|---|
Compared with their regular rate | 0.10 / 0.50 gpt-6-luna | 2.00 / 10.00 gpt-6.1-sol | 4.00 / 20.00 Claude Opus 5.5 |
Remnant < 7 days | 0.0408 / 0.204 | 0.765 / 3.825 | 1.70 / 8.50 |
Standard < 24 hours | 0.048 / 0.24 | 0.90 / 4.50 | 2.00 / 10.00 |
Priority Planned < 6 hours | 0.0672 / 0.336 | 1.26 / 6.30 | 2.80 / 14.00 |
Rush Planned < 1 hour | 0.096 / 0.48 | 1.80 / 9.00 | 4.00 / 20.00 |
- Remnant is 15% under Standard, Priority 40% over it, and Rush twice it.
- Tool calls: $0.002 per call your endpoint answers, at every deadline; a call that fails costs nothing. Details.
- A different model, or volume pricing: Ask in your portal.
Enrich 100,000 product listings
800 input + 400 output tokens each: 80M in, 40M out. Small class.List prices on both sides, before prompt caching, which lowers both bills: it is free here and a tenth of the input price at OpenAI.
Pick a deadline
A deadline is a promise about when your results arrive, not a place in a queue. Remnant and Standard are on now; Priority and Rush have prices and switch on once enough work flows to keep machines in reserve.
Remnant
LiveBacklogs, re-processing, anything with no fixed date on it.
Standard
LiveNightly jobs: the everyday choice.
Priority
PlannedSame-business-day batches.
Rush
Planned“I need it this hour.”
What each deadline does when it is tight, and how to pick one: deadlines explained and choosing a deadline.
How a workflow is billed
Each model step is charged like a single request, at the rate for its size and your deadline, per answer it gets back; a tool step per call your endpoint answers. A skipped step costs nothing.
How this differs from OpenAI’s batch service
OpenAI and Anthropic both sell a batch service: send work, get it back within a day, at half price. A good product, and a different one.
| Aspect | AI company batch services | Otium |
|---|---|---|
| Deadline | Within 24 hours, and it runs then whether prices are good or not. | Up to 7 days. The extra time is exactly what buys the lower price. |
| Multi-step work | Not offered. One request at a time, one deadline per step: a seven-step job is seven submissions, seven waits, and the code between them is yours. | The whole pipeline is one workflow with one deadline. Otium works out the graph, runs steps that don’t depend on each other side by side, and checks each output before the next step sees it. |
| What you build and run | The orchestrator between steps: a queue, a retry loop, schema checks, scheduling, and a service to run them, written and operated by your team. | A document and its tests. The orchestration is ours, so a small team manages prompts, not pipeline infrastructure. |
| Tool calls | A request may ask for a tool call, but the batch ends there: you run the tool yourself and submit a new batch with the answer. | A tool step calls your own endpoint mid-pipeline, signed, and the answer feeds the next step. |
| Repeated input | Billed at a tenth of the input price, and on some cards there is a separate charge to write the cache. | Free. |
| Models | A specific model name, their own, with a licence fee built into every token. | A size: small or medium today, large to come, each on an open model anyone can download. No licence fee, and pooling everyone onto a short list is part of the discount. |
Compare for yourself
Cost your own workload, see every model on the card’s compare row at its own prices, and check the computers underneath at the live GPU prices and the monthly report.
What your work would cost
That is $1,461 a year against gpt-6-luna’s regular price.
Assumes 2 input tokens per 1 output token, as in the costed job above. Repeated input is priced as each side bills it: free here, a tenth of the input price there, with their charge for writing the cache left out.
Every model we compare against
| Model | Vendor | Standard in / out | Batch in / out | Batch, blended |
|---|---|---|---|---|
| gpt-6-luna anchors Small | OpenAI | 0.10 / 0.50 | 0.05 / 0.25 | $0.117 |
| Claude Haiku 5.5 | Anthropic | 0.10 / 0.50 | 0.05 / 0.25 | $0.117 |
| gpt-6.1-sol anchors Medium | OpenAI | 2.00 / 10.00 | 1.00 / 5.00 | $2.333 |
| Claude Sonnet 5.5 | Anthropic | 2.00 / 10.00 | 1.00 / 5.00 | $2.333 |
| Claude Opus 5.5 anchors Large | Anthropic | 4.00 / 20.00 | 2.00 / 10.00 | $4.667 |
| gpt-6-astra | OpenAI | 10.00 / 50.00 | 5.00 / 25.00 | $11.667 |
We compare against where this work runs today, OpenAI’s models for Small and Medium and Claude Opus 5.5 for Large, not against whichever model is cheapest: until a test on your task says otherwise, a cheaper model that gets the job wrong is not cheaper.
Already own GPUs? Put their spare hours to work
GPUs bought for peak load are rarely busy around the clock, and how much room they have depends on your teams, their timezones and what else shares the hardware. Filling it takes a batch platform, which is a project of its own. Otium is that platform: point it at your GPUs and they work through the backlog whenever they have room.
What you would otherwise build
-
A batch API your code already speaks
OpenAI-compatible batch files, the same SDK and a base-URL change, plus a command-line tool.
-
A scheduler that knows your hours
Work runs only in the hours you give the machines, nights and weekends, and waits for the next window rather than crowding the day.
-
Deadlines, not a place in a queue
Each batch carries a finish time, and the scheduler plans the hours ahead around it.
-
Retries that lose nothing
A request that fails, or a machine that restarts mid-batch, is retried; each job is answered once.
-
Pipelines, not just prompts
Multi-step workflows with tool calls to your own systems and steps a person signs off.
-
Somewhere to see it all
Every batch, run and result in one place, with usage by account and model.
Two ways to run it: self-hosted workers are in beta and you can start today, and on-prem is being built with design partners.
Self-hosted workers
BetaYour GPUs run the work; the platform runs in our cloud: the API, the scheduler, storage and the run views.
Includes 2 billion tokens a month, then $0.02 per million tokens, input and output alike, whatever the model.
First 30 days free, no credit card needed.
-
One small program on your side
The worker sits beside your own inference server (vLLM, llama.cpp, anything OpenAI-compatible) and takes work from the queue.
-
Pay for the months it runs
A month is 30 days from the first job your workers take, so a month they sit idle costs nothing.
-
Tool calls inside your network
A tool runner on your own network is free, so a workflow can call systems that never face the internet.
-
What still leaves your network
Prompts and results pass through our control plane, encrypted at rest. If they must never leave, that is on-prem.
On-prem
In developmentThe whole platform, installed and run in your own data center or cloud account, with your GPUs. Nothing runs on ours.
An annual license with support, set with you for the GPUs you run it on.
-
Nothing leaves your site
The API, scheduler, consoles and workers run on your cluster, against your own storage, keys and sign-in.
-
Batch work in the idle hours
Your teams submit batches all day; they run on your GPUs nights and weekends, each with a finish time you can plan around.
-
Kubernetes or Slurm
Install it on your cluster, or run the workers as jobs under the scheduler you already have.
Give us a deadline and see what it costs.
Otium is in beta. Leave your email and we’ll send you a link to create your account. Then point your OpenAI-compatible client at our base URL and send your first batch.
Free to try on our demo model. No credit card, no spam: we email you a link to finish.
Otium is in beta and open to everyone. Already have an account? Sign in.