PWR Auto
The right model for each run, with reasons
PWR Auto is a routing layer, not a model. For each run it picks a catalogued third-party model based on the job, your policy and expected cost, and records why. Credits, estimates and budgets keep spend visible before and after the run.
Teams that want sensible model defaults and spend controls without choosing a model by hand for every task.
How it works
PWR Auto, step by step
- 01
Leave the model on Auto
Agents can use PWR Auto as their model, or as their planning or review model. A single task can override it with an explicit model.
- 02
Filters apply first
Routing only considers providers you can reach and have approved, models with the required capabilities, enough context window, and an expected cost under your per-run cap.
- 03
A deterministic score picks the model
Remaining candidates are scored on a quality proxy, expected cost and context headroom, weighted by the purpose of the run and adjusted by recent reliability.
- 04
The decision is recorded
Every routed or explicit selection is stored with a readable reason. The run is metered and charged in credits like any other.
- Available
- Implemented and covered by automated tests in the current platform build.
- Beta
- Implemented; verification against the live external service is still pending.
- Upcoming
- Planned or in progress. Not usable yet.
Routing
PWR Auto orchestrates models from third-party providers. PWR does not train a foundation model.
Purpose-aware model selection
BetaPlanning, review and security runs weight quality and context. Implementation is balanced. Chat favors lower cost. An economy, balanced or quality tier shifts the weighting.
Routing logic is covered by automated tests using provider simulators; no live multi-provider run has been verified yet.
Organization model policy
BetaOwners set approved providers, the default tier, a maximum cost per run and whether fallback is allowed. The policy applies to explicit model choices too.
Routing logic is covered by automated tests using provider simulators; no live multi-provider run has been verified yet.
Routing decision log
BetaEach decision records the chosen model, its component scores, the expected cost, how many models were eligible or excluded, and the policy in force.
Routing logic is covered by automated tests using provider simulators; no live multi-provider run has been verified yet.
Fallback between attempts
BetaIf a provider refuses a run, for example with a rate limit or an outage, the next attempt avoids that provider for a cooldown period when your policy allows fallback.
Routing logic is covered by automated tests using provider simulators; no live multi-provider run has been verified yet.
Per-role and per-task model overrides
BetaSet separate implementation, planning and review models on an agent, or pin a model on a single task.
Routing logic is covered by automated tests using provider simulators; no live multi-provider run has been verified yet.
Measured quality signals
UpcomingRouting on benchmark or outcome data. Today the quality signal is a proxy built from price rank and model features.
Credits, estimates and budgets
Usage is charged in credits against an append-only ledger. You can see a range before you start and a hard limit applies while you run.
Append-only credit ledger
AvailableGrants, purchases, holds and charges are ledger entries that cannot be edited or deleted. Each entry is idempotent and the balance can be recomputed from the ledger.
Reserve, then settle
AvailableA managed run places a hold before it starts and settles to actual usage when it ends, like a card authorization. Concurrent runs cannot overdraw the balance.
Estimates before you launch
BetaRequest a low and high credit range for a task, a goal, a coverage run or a chat, including the model the run would be routed to and the assumptions used.
Estimates use starting token profiles and catalog prices that have not been tuned against real usage yet.
Budgets and notifications
AvailableSet a monthly hard limit and a low-balance threshold per organization. Managed runs that would exceed the limit are refused before they start.
Low-balance and budget notifications are recorded in the dashboard. Email delivery is not built yet.
Plans, subscriptions and top-ups
BetaSubscription plans with included credits and one-off credit packages purchased through Stripe.
Verified against a simulated Stripe; live Stripe verification is pending and plan pricing is not final.
Providers and keys
Anthropic Agent SDK runtime
AvailableThe primary agent runtime, with tool allowlists and workspace path confinement.
Verified with real model runs. The path that brokers a customer key through the credential broker is covered by simulators only.
OpenAI, Google Gemini and xAI adapters
BetaTool-calling adapters with streaming and normalized usage reporting, plus the Anthropic Messages API.
Verified against in-repo provider simulators; verification against the live provider is pending.
Bring your own key
BetaBind your provider key to an agent. A run with a bound key never falls back to a PWR-managed key, and routing stays within that provider.
Verified against in-repo provider simulators; verification against the live provider is pending.
Credential broker
BetaProvider keys stay in the worker. The sandbox receives a placeholder and a local proxy injects the real key on the way out.
Verified against in-repo provider simulators; verification against the live provider is pending.
Versioned model catalog
BetaEach model has versioned capabilities and pricing. A model cannot be published while its pricing is unverified.
Catalog pricing has not yet been confirmed against provider price pages.
Planned controls
Where model selection is going. None of the items below can be used today, and none changes what a plan costs until it is announced with its own pricing.
Auto Economy, Balanced and Best Quality
UpcomingThree settings for PWR Auto that trade cost against capability for the whole project.
Choose the model yourself
UpcomingPick a specific Claude, OpenAI GPT or xAI Grok model for a project instead of leaving it on Auto.
Reasoning-effort control
UpcomingSet how much reasoning a model spends, on the models that support it.
Per-model credit consumption
UpcomingSee how many credits each model uses before you choose it.
Questions
Is PWR Auto its own AI model?
No. PWR Auto is a routing and orchestration layer. It selects among catalogued models from third-party providers and then runs the job through the same adapters, metering and pricing as an explicit model choice.
How does PWR Auto judge model quality?
Today it uses a proxy: a model's price rank among the candidates plus bonuses for features such as reasoning controls and structured output. It is not based on benchmarks. The score also includes expected cost, context headroom and the recent success rate of each model on your runs.
Can I see why a model was chosen?
Yes. Every decision is stored with a reason that names the winning model, its scores, the expected cost, the weights used and the policy in force. Administrators can list decisions per run or per agent.
Will PWR Auto switch providers in the middle of a run?
No. Fallback happens between attempts, never inside a run. If you use your own key, routing never crosses to a PWR-managed key.
How accurate are credit estimates?
They are ranges, not quotes. An estimate prices a typical token profile for the kind of run and reports a low and a high figure with its assumptions. The profiles are starting values and will be tuned as real usage accumulates.
Put an AI development team on your idea.
Describe what you want to build. PWR plans, builds, tests and reviews it, and you approve each step that matters.