Skip to content
FoundationAI teamsMaestroZB CloudApply for access

Model Router

The best model for every task.

No AI model is best at everything. One writes the best plan, another writes the best code, a third checks the work. The Model Router reads each task as it starts, gives it to the model that does it best at the lowest price, checks what comes back, and hands the result to the next model with its context intact. That choice is Maestro’s. Maestro is the model at the heart of ZB Stack. It knows every model, and the cloud each one runs on.

Apply for access See how one run works

300+ AI model endpoints across 14 leading model providers.

What makes it different

A router asks where to send a request. Maestro asks the best way to get the job done.

Most routers send one request to one model and stop there. Maestro looks at the whole job. Which model for each task. Which cloud. Whether the result passed. What to do when it did not. Then it learns from the answer, so the next job goes better.

How one run works

Six steps, the right model for each.

Every run your AI team starts goes through the same six steps. Plain code runs them, checks the work and keeps the budget, with no model needed for that part. Each step goes to the model that suits it: the top model plans and signs off, low-cost models do the bulk of the work side by side, and a mid-price model takes the harder fixes and reviews each one. Only passing attempts are billed.

How one run works. Plain code sits at the center: the workflow runner, the Model Router, the micro-router and a $15 safety stop. Once the founder approves the plan and the balance is OK, six steps go round it. Plan: claude-fable-5-1, the high tier, writes the workflow script, with a schema check of up to 5 tries. Scan: gpt-4.1-mini or deepseek-v4-flash, the low tier, many at once, findings in fixed JSON. Fix: gpt-4.1-mini first; tries 2 and 3 move up to glm-5.2 or deepseek-v4-pro. Check: tests, secrets, file limits and cost stop, in plain code; after 3 tries a fix is marked failed and kept on record. Review: glm-5.2 or deepseek-v4-pro, a different model from the one that wrote the fix. Sign off: claude-fable-5-1 reads only the summary, then ships to ZB Cloud with a preview link for the founder, or holds for the founder. Output price per 1M tokens: high $52.63, medium $4.17 to $4.40, low $1.26 to $1.68. Only passing attempts are billed.
How one run worksTap to open full size
  1. Step 1 · High

    Plan the run.

    The top model, claude-fable-5-1, writes the workflow script for the whole run: what to scan, what to fix, and which checks each fix must pass. Plain code checks the script’s shape and asks for a new one when needed, up to five tries.

  2. Step 2 · Low

    Scan, many at once.

    Low-cost models, gpt-4.1-mini or deepseek-v4-flash, read the work side by side. Each one returns its findings in the same fixed JSON shape, so the next step can act on them directly.

  3. Step 3 · Low, then Medium

    Fix, one copy per fix.

    gpt-4.1-mini writes the first try at each fix, in its own isolated copy of the code. Tries two and three move up to glm-5.2 or deepseek-v4-pro. The higher price applies to that one fix, not the whole run.

  4. Step 4 · Plain code

    Check every fix.

    Plain code runs the tests and checks for secrets, file limits and cost. A fix that passes moves on to review. After three tries, a fix is marked failed and kept in the record.

  5. Step 5 · Medium

    Review with a second model.

    glm-5.2 or deepseek-v4-pro reads each passing fix. The reviewer is always a different model from the one that wrote the fix. Approved fixes are merged.

  6. Step 6 · High

    Sign off and ship.

    claude-fable-5-1 reads a short summary: what passed, what failed, and the cost at each price tier. The run ships to ZB Cloud and the founder gets a preview link. Every step, model and cost goes in the record.

Example

One task, followed through the Router.

An abandoned-cart email for Northwick Candle Co., our example store, which sells in two currencies. Each step shows its model and its cost. One test fails. Only the fix moves to a stronger model.

Example
StepModelWhy this modelCost
Write the task specFableBest at specs$0.22
Plan the filesGPTDrafting and planning$0.03
Write the code, four filesDeepSeekStrong code at a low price$0.04
Write the testsSonnetTesting$0.05
One test failed. Re-routed to the next model up.OpusStronger model for the fix$0.08
Review the changeOpusReview by a model that did not write the code$0.07
Write the email copy and subject linesGPTCopy$0.01
Merge and deploy a previewZB CloudHosting$0.00
Total$0.52

The same task with everything on one model: $1.45. List prices on 10 October 2026. The deploy step is part of hosting, charged separately; the 5% fee is separate too.

At scale

The same build. Three prices.

Northwick Candle Co. is our example store: a storefront, an AI shopping assistant and the eighteen systems behind them. 3,559 AI tasks across 12 models.

Fable, one model for everything$720.00
Opus, one model for everything$290.97
ZB Stack, routed by the Model Router, with the 5% fee$29.10
90% less than Opus · 96% less than Fable

Seventy-nine tasks failed a check and were re-routed; their retries are inside the $29.10. Example build, not a customer result.

The board

Twelve models on the board today.

The board is the set of models the Router uses most for each kind of work. It changes when a better model ships or a price moves; your AI team gets the change on the next task.

Kind of workModel today
Kind of workModel today
SpecsFable
Drafting, planningGPT
Hard planningAstra
Code, summariesDeepSeek
TestingSonnet
ReviewOpus
ResearchGrok
ImagesGemini
TranslationGLM
ExtractionQwen
CleanupLlama
TaggingMistral

Behind the board: 300+ AI model endpoints across 14 leading model providers. From OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, GLM, Kimi, MiniMax, Mistral, NVIDIA, Cohere, Perplexity and more.

Key ideas

Built on five ideas.

Everything the Model Router does comes from five simple pieces.

Task
One piece of work with one result: a spec, a file of code, an image, a reply to a customer.
Rule
Where a task may go and how much it may spend. Set once, applied to every task it covers.
Route
The Router’s choice of model for a task, made as the task starts.
Check
A test, a review or a rule that the result must pass, run by something that did not produce it.
Record
Every task with its model, the reason, the check and the cost.

For developers

Rules, written as a short list.

Developers can set rules for a product or a whole workspace. Route work by cost, speed, quality and your own rules. You can pin a model on any task.

Examplerules.yaml
# Example. A routing rule for one product.
rule: northwick-store
prefer: sponsored            # use sponsored models when they do the task well
data_location: [us, ca]      # providers that keep data in the US or Canada
check:
  code: tests + review       # tests must pass, a second model reviews
  copy: review
on_fail: next_model_up       # re-route only the failed task
max_cost_per_task: auto      # set by the balance
pin:
  images: gemini             # always use this model for images

Model Router

Every task, on the best model for it.

Tell us what you want to build. Sponsored models are free.

Apply for access

How it works →