INTELLIGENCE, WELL DIRECTED.

Less choosing.
More doing.

Every task deserves
its own intelligence.

We train Aster to recognise what your task needs—and choose which model takes it on.

Meet aster-work & aster-code ↓

ASTER / DECISION FIELD

aster-code / INCOMING TASKReview this pull request

Matched to Claude↗

Some tasks need a frontier model. Others don’t. You shouldn’t have to work out which. Using a frontier model for everything gets expensive. Always choosing the cheapest can compromise the result. The Aster Model uses intelligent algorithms trained to consider each task and its context, then select the model for that step.

We train
the decision.

Aster Router is our purpose-trained model for choosing which intelligence takes on your task.

A routing policy trained with reinforcement learning—using the task, its context and the cost of the next move.

A TASK ARRIVES

Code reviewFind the bug across these files

Tools & skills Repository + tool results Task requirements

RL TRAINING

Aster RouterTASK-AWARE ROUTING POLICY

Outcomes inform the next decision

THE RIGHT INTELLIGENCE

Claude

GPT

Gemini

Muse

Mistral

Nemotron

Matched to Claude

TRAIN ON THE WORK

Real task data + model outcomes.
Our own model, trained for model choice.

ADAPT TO WHAT’S NEXT

New models. More feedback.
Real-time adaptation is our research direction.

<500ms

Routing latency target

The cost of getting
the task done.

Estimated costs as a coding conversation grows to 100 turns, compared across five options.

Fable 5.1
GPT-6 Astra
Gemini
Open-source mix
Aster

Cost over 100 coding turns

Estimated total API cost · USD

Conversation turns

Task success rate

Unverified estimates · at 100 turns per task

Successfully completed / attempted tasks

At 100 conversation turns Adjusted cost estimates · task success is unverified

Claude Fable 5.1
Total cost
$117.00
Task success*
94%
GPT-6 Astra
Total cost
$92.00
Task success*
93%
Gemini 3.8 Flash
Total cost
$15.00
Task success*
79%
Mix OpenSource Models
Total cost
$20.00
Task success*
73%
Aster
Total cost
$9.12
Task success*
94%

Adjustable total API cost estimates in USD, by coding conversation length
Conversation turns	Claude Fable 5.1	GPT-6 Astra	Gemini 3.8 Flash	Mix OpenSource Models	Aster
10	$2.70	$2.12	$0.35	$0.44	$0.19
20	$7.40	$5.82	$0.95	$1.23	$0.54
30	$14.10	$11.09	$1.81	$2.36	$1.06
40	$22.80	$17.93	$2.92	$3.84	$1.73
50	$33.50	$26.34	$4.29	$5.67	$2.56
60	$46.20	$36.33	$5.92	$7.84	$3.55
70	$60.90	$47.89	$7.81	$10.36	$4.70
80	$77.60	$61.02	$9.95	$13.23	$6.02
90	$96.30	$75.72	$12.35	$16.44	$7.49
100	$117.00	$92.00	$15.00	$20.00	$9.12

Why the cost grows
with each turn.

You’re paying for more than your latest message. Each follow-up can carry the conversation, instructions, tools, skills and earlier results. A short request can arrive with a very long history.

Aster Context Method

Our method combines reinforcement learning algorithms with task analysis, caching and context management to shape each request and model choice.

Caching
Reusable prompt prefixes, with cache handling that respects provider and model boundaries.
Task-aware routing
Task intent, complexity and tool requirements inform model selection.
Context-aware selection
Relevant history, instructions and dependencies shape the context sent to the model.
RL training & real-time feedback
Reward signals from task outcomes and real-time feedback inform routing-policy refinement.

ACCUMULATED CONTEXT

System instructions

Tools & skills

Relevant history

Stale tool output

Unrelated history

Current task

SELECTED CONTEXT

System instructions

Tools & skills

Relevant history

Current task

A focused request, ready for the selected model.

TASK-RELEVANT CONTEXT PER TURN

TURN 1Instructions + new input

TURN 2Select relevant history

TURN 3Select relevant history

Frontier models are
part of Aster.

A difficult step can still go to Astra, Fable or Gemini. The context system is designed to reduce what that model needs to process, while the router selects a suitable model for each step.

FROM OUR CURATED COLLECTION

AstraOpenAI
FableAnthropic
GeminiGoogle

And more.

Model availability depends on the product and enabled route.

Your work.
One model name.

Aster’s curated collection, ready to use.
Choose your product. We handle model selection.

YOUR NEXT REQUESTJSON

{
  "model": "aster-code",
  "messages": [{
    "role": "user",
    "content":
      "Find and fix the failing test in this repository."
  }]
}

Read the integration guide

{
  "model": "aster-code",
  "messages": [
    {
      "role": "user",
      "content": "Find and fix the failing test in this repository."
    }
  ]
}

20curated models.
And growing.