One Tokaroo key
Authenticate from a server or trusted harness. Permanent keys never belong in browser or mobile bundles.
Tokaroo product
One API key. Five intelligent models. Every major provider.
Your application asks Tokaroo for Max, Pro, Auto, Fast, or Basic. Tokaroo makes the optimized routing decision, handles compatible fallback, meters the result, and returns the response in the format your software expects.
The product boundary
Tokaroo routes intelligence. Your software stays in control.
Keep your application, agent harness, tools, files, and execution environment. Replace separate model accounts and provider-specific routing code with one managed connection.
Five model names
Choose the result you need instead of maintaining a catalog of vendor model IDs.
Optimized routing
Tokaroo weighs compatibility, measured quality, price, speed, health, and availability across its managed provider network.
Clear usage
See the requested Tokaroo mode, token counts, latency, charge, balance, and aggregate benchmark savings.
Five virtual models
Choose the operating mode, not the vendor.
Tokaroo uses fresh task-specific evaluations internally, then optimizes each request for the operating mode you choose.
Max
The hardest and highest-stakes work where capability is the first requirement.
Optimizes for the strongest measured capability available.
Pro
Serious daily technical, business, production, and customer-facing work.
Optimizes for dependable production-grade results.
Auto
The default for applications that want Tokaroo to make the whole routing decision.
Optimizes the full quality, speed, reliability, and price tradeoff.
Fast
Rapid answers, edits, extraction, and interactive back-and-forth.
Optimizes for low latency while preserving task suitability.
Basic
Everyday questions and routine tasks where you need an answer, not a project.
Optimizes for the lowest practical cost that can complete the work.
Each mode describes the result Tokaroo should optimize while routing adapts to the request.
Classify the work
Quality is measured separately for reasoning, code, math, reading, writing, tools, multilingual, and spatial work.
Measure current performance
Fresh task-specific evaluations establish which candidates produce dependable results.
Build compatible choices
Only candidates meeting internal quality, format, context, and tool checks remain eligible.
Optimize the route
Tokaroo then weighs speed, price, reliability, context, tools, and availability inside that basket.
Math and quantitative reasoning
Sep 4, 2026 | Fresh Tokaroo evaluation evidence
7 freshly tested2 decision-grade4 provider families
Max2eligible candidatesCapability first
Pro7eligible candidatesProduction quality
Auto7eligible candidatesBalanced automatically
Fast7eligible candidatesLatency first
Basic7eligible candidatesLowest practical cost
Reading and synthesis
Aug 28, 2026 | Fresh Tokaroo evaluation evidence
2 freshly tested2 decision-grade1 provider families
Max2eligible candidatesCapability first
Pro2eligible candidatesProduction quality
Auto2eligible candidatesBalanced automatically
Fast2eligible candidatesLatency first
Basic2eligible candidatesLowest practical cost
Writing and editing
Aug 28, 2026 | Fresh Tokaroo evaluation evidence
3 freshly tested3 decision-grade1 provider families
Max2eligible candidatesCapability first
Pro3eligible candidatesProduction quality
Auto3eligible candidatesBalanced automatically
Fast3eligible candidatesLatency first
Basic3eligible candidatesLowest practical cost
Structured data and tool decisions
Aug 31, 2026 | Fresh Tokaroo evaluation evidence
6 freshly tested6 decision-grade3 provider families
Max1eligible candidatesCapability first
Pro5eligible candidatesProduction quality
Auto6eligible candidatesBalanced automatically
Fast6eligible candidatesLatency first
Basic6eligible candidatesLowest practical cost
Multilingual work
Aug 30, 2026 | Fresh Tokaroo evaluation evidence
4 freshly tested2 decision-grade3 provider families
Max1eligible candidatesCapability first
Pro3eligible candidatesProduction quality
Auto4eligible candidatesBalanced automatically
Fast4eligible candidatesLatency first
Basic4eligible candidatesLowest practical cost
3 additional task families are still accumulating fresh decision-grade evidence.
Evaluation evidence is limited to successful Tokaroo tests in the latest 30-day window; imported zero-sample signals are excluded. The concrete serving model, serving provider, and private settlement cost remain internal.
Compatible by design
Four wire formats cover modern applications and harnesses.
Use Tokaroo directly or point an existing provider client at the matching base URL. LangChain and the Vercel AI SDK are optional client channels, not gateway dependencies.
Direct REST
Native HTTP, JSON, server-sent events, and WebSocket routes.
OpenAI-compatible
Use existing OpenAI clients with a base URL and key change.
Anthropic-compatible
Messages, tool use, streaming, and token counting for Claude-oriented clients.
Gemini-compatible
GenerateContent, streamGenerateContent, and countTokens.
Current model and provider families
GPT / OpenAIClaude / AnthropicGemini / GoogleGroqKimi / MoonshotDeepSeekQwen / AlibabaGrok / xAIOpenRouterMistralLlamaCohereAugure AIInterfaze AIRaspy multimodal
The researched catalog changes as availability, verified pricing, and compatibility change. Your five Tokaroo model names remain stable.
Production routing
The decision and the accounting travel together.
A useful router must do more than proxy a response. Tokaroo owns the route decision, compatible retry path, and request ledger.
Compatibility before scoring
Candidates must support the request's wire, tools, context, modality, and output constraints before they can be ranked.
Health-aware fallback
Unhealthy or incompatible routes are excluded, and eligible provider fallback is handled behind the same response format.
Fresh measurements
Quality, latency, reliability, and verified price inputs inform optimized routing instead of a static vendor list.
One usage ledger
Tokaroo mode, charge, tokens, latency, and aggregate benchmark savings are recorded against the key and workspace.
Stable model contract
Applications ask for one of five Tokaroo modes while the underlying provider catalog can improve without application changes.
Harness-aware affinity
Native model families can be preferred for proven tool loops, but only inside the selected quality band and without removing compatible fallback.
Proven in real systems
Gwen demonstrates orchestration. Raspy extends the route beyond text.
These products clarify Tokaroo's role without turning Tokaroo into a hosted harness platform.
Gwen reference application
Gwen selects and runs agent harnesses, then sends their model traffic through Tokaroo. Gwen owns orchestration and execution; Tokaroo owns optimized model routing, fallback, and metering.
Raspy multimodal rail
Tokaroo routes speech, transcription, realtime audio, and avatars through Raspy. Images and video use compatible routes from the broader model network.
One key. Five intelligent models. Every major provider.
Start with Auto and let Tokaroo make the routing decision.