Skip to main content
Agents and coding sessions can run on models from four providers — Anthropic, OpenAI, Google, and Kimi. Multi-model is available to all organizations with no special enrollment.

​
Capability tiers

End users pick a tier, not a specific model — so prompts and agent configs stay portable when the model behind a tier changes.

Smartest

Deepest reasoning for complex, multi-step work.

Balanced

The best everyday default — strong quality at moderate cost.

Fastest

Quick and cheap for simple, high-volume tasks.
By default, each tier resolves to the Claude model in that slot — Opus 4.8 (Smartest), Sonnet 5 (Balanced), Haiku 4.5 (Fastest). Organizations can customize the different tiers of models for their end users.

​
Model matrix

Each provider fills the same three tiers. The number next to each model is its relative cost — a multiple of Claude Opus 4.8 (1×), blended across input and output tokens (assuming a ~3:1 input-to-output mix), so every model sits on one scale. Figures are public list prices as of July 2026; actual billing is usage-based.
TierAnthropicOpenAIGoogleKimi
SmartestClaude Opus 4.8 · 1×GPT-5.5 · 1.1×Gemini 3.5 Flash · 0.34×—
BalancedClaude Sonnet 5 · 0.6×GPT-5.4 · 0.56×Gemini 2.5 Flash · 0.09×Kimi K2.7 Code · 0.17×
FastestClaude Haiku 4.5 · 0.2×GPT-5.4 mini · 0.17×Gemini 3.1 Flash-Lite · 0.06×Kimi K2.6 · 0.17×

Using models with agents

How a configured agent locks its chats to a single model.