Press a switch to hold it out
Optimize utility per dollar, on every inference request.
Dispatch routes every request across frontier, open-source, bespoke and sovereign models, behind one base URL.
- Example request
- —
- Bar
- —
- Model
- —
- Cost
- —
For every bar, one model is cheapest.
Third place costs $1.34 a task and first place costs $9.15.
61 models · 10 on the frontier
Pareto frontier
Arena score against median cost per task
Chart of 61 models that publish a cost. Median cost a task runs left to right on a log scale from ten dollars down to one cent, so cost falls as you read across, and Arena's score runs bottom to top, in percentage points over the average model. 10 of them are Pareto-optimal and form the frontier, a line through Claude Fable 5.2 (Max) at $9.15 down to GPT-6 Luna (Low) at $0.01. For any bar you set, the cheapest model at or above it is the one Dispatch routes to.
- Claude Fable 5.2 (Max): $9.15 a task, +28.61%.
- Claude Opus 5.5 (Max): $5.87 a task, +18.19%.
- Gemini 4 Carbon (High): $1.34 a task, +16.45%.
- GPT-6.1 Sol (Max): $1.00 a task, +11.78%.
- GPT-6.1 Terra (Max): $0.88 a task, +11.34%.
- Gemini 4 Argon (High): $0.68 a task, +10.00%.
- GPT-6.1 Sol (Medium): $0.38 a task, +8.04%.
- GPT-6.1 Terra (Medium): $0.16 a task, +4.44%.
- Muse Ember (Max): $0.02 a task, +1.73%.
- GPT-6 Luna (Low): $0.01 a task, −7.18%.
- Frontier
- Closed
- Open weights
- Routed to
- Your bar
Pareto-optimal models
Most expensive first
- Claude Fable 5.2 (Max)Anthropic · Proprietary$9.15+28.61%
- Claude Opus 5.5 (Max)Anthropic · Proprietary$5.87+18.19%
- Gemini 4 Carbon (High)Google · Proprietary, preliminary$1.34+16.45%
- GPT-6.1 Sol (Max)OpenAI · Proprietary$1.00+11.78%
- GPT-6.1 Terra (Max)OpenAI · Proprietary$0.88+11.34%
- Gemini 4 Argon (High)Google · Proprietary$0.68+10.00%
- GPT-6.1 Sol (Medium)OpenAI · Proprietary$0.38+8.04%
- GPT-6.1 Terra (Medium)OpenAI · Proprietary$0.16+4.44%
- Muse Ember (Max)Meta · Proprietary$0.02+1.73%
- GPT-6 Luna (Low)OpenAI · Proprietary$0.01−7.18%
Routed to Gemini 4 Carbon (High) at $1.34 a task.
Source
Arena agent leaderboard, 13 Nov 2026. Public data, not a Dispatch measurement.
arena.ai/leaderboard/agentThe score
Arena’s score is each model’s improvement over the average model across Arena’s signals, in percentage points.
Coverage
2,979,271 sessions across 65 models, 61 of which publish a cost a task.
One line of code, then it routes.
- client.pyExample
- 1client = OpenAI(
- -
base_url=PROVIDER_URL, - + base_url=DISPATCH_URL,
- 3 api_key=DISPATCH_KEY,
- 4)
- 1 line changed
- Routing
01Change one line.
Point your OpenAI or Anthropic client at Dispatch.
- 02
Set your bar.
Say how good a result has to be for each kind of work.
- 03
It picks the cheapest that clears.
When a provider fails or rate-limits, the request falls back to the next.
- 04
It learns what you accept.
The results you accept and reject improve the next route.
Hosted by us, or inside your own walls.
Hosted or sovereign
Use the endpoint we host, or run Dispatch in your own environment on your own provider keys.
Any model you can call
Route across frontier and open-source models, models built for you, and models you run yourself.
- Example
The pool: one address in, four kinds of model behind it, and each request patched to the cheapest that clears its bar.
A record of every request
See which model served each request and what it cost.
requests.logExample- 1Reconcile the ledgerReconcileFrontierGPT-6.1 Sol $$$GPT-6.1 Sol $$$
- 2Extract the invoice fieldsExtractOpen sourceKimi K3 $ Kimi K3 $
- 3
Mark up the indemnityMark upBespokecontracts-v2 contracts-v2 No reply - →Mark up the indemnityMark upFrontierGPT-6.1 Sol $$$GPT-6.1 Sol $$$Rerouted
- 4Answer from patient notesAnswerSovereignKimi K3.1 Pro $$ Kimi K3.1 Pro $$
Integrity, if you want it
Turn on Integrity and the model Dispatch picks is steered while it works.
Stop paying frontier prices for work a cheaper model does as well.
Tell us the kind of work you send and the bar it has to clear.
Get in touch