Model Provider
How Federation registers and manages AI model providers
The Federation AIService abstracts over multiple AI model providers. A model provider is any backend that can serve model requests — OpenAI, Anthropic, Google, or a local inference server.
Registration
Providers register with City by exposing:
- Available model IDs
- Capability flags (streaming, tool use, vision, etc.)
- Authentication requirements
Unified API
Regardless of the backend, the Federation AIService exposes a unified API to Agents. This means:
- Agents do not need provider-specific code
- Switching providers is a configuration change, not a code change
- New providers can be added without modifying Agents
Capability discovery
When an Agent binds to a model ID, City reports:
- Whether the model supports streaming
- Whether the model supports tool use
- Maximum context length
- Pricing per token
Official cloud models
The main updated models are listed below. Supplier base prices were checked on November 4, 2026, in USD per million tokens, before the official service's billing multiplier.
| Model ID | Supplier | Channel input limit | Base input / output price |
|---|---|---|---|
gpt-6.1-sol | APIMart | 922K | 1.6 / 8 |
gpt-6.1-terra | APIMart | 922K | 0.8 / 4 |
gpt-6-astra | APIMart | 922K | 8 / 40 |
gpt-6-sol | APIMart | 922K | 1.6 / 8 |
gpt-6-luna | APIMart | 922K | 0.08 / 0.4 |
claude-opus-5-5 | APIMart | 128K | 3.2 / 16 |
claude-sonnet-5-5 | APIMart | 1M | 1.6 / 8 |
claude-haiku-5-5 | APIMart | 1M | 0.08 / 0.4 |
claude-fable-5.1 | APIMart | 1M | 8 / 40 |
grok-4.7 | APIMart | 500K | 1.6 / 4.8 |
gemini-3.8-flash | APIMart | 1,048,576 | 0.6 / 3 |
deepseek-v4.1-flash | 302.AI | 1M | 0.15 / 0.6 |
deepseek-v4-pro | 302.AI | 1M | 0.66 / 1.98 |
glm-5.3 | 302.AI | 1M | 1.4 / 4.4 |
kimi-k3.1-pro | 302.AI | 1,048,576 | 2.4 / 12 |
qwen3.8-max | 302.AI | 1M | 2.16 / 6.36 |
DeepSeek base costs are multiplied by 2; other models use 1.2. Credits are rounded up. DeepSeek now uses 302.AI's listed fixed rates; the direct provider's peak/off-peak schedule no longer applies. GPT requests above 272K input tokens use twice the input and cache rates and 1.5 times the output rate for the whole request. Grok 4.7 doubles its rates from 200K input tokens. Haiku 5.5 uses five times the listed rates above 100K input tokens.
DeepSeek V4.1 Flash handles image input directly and routes other files to gpt-6-luna. DeepSeek V4 Pro continues to route files to Luna. Requests are billed for the model that handles them. Update old deepseek-v4-flash configurations to deepseek-v4.1-flash.
New per-image entries:
| Model ID | Supplier | Base price (USD per image) |
|---|---|---|
apimart-gpt-image-2.5-ext | APIMart | 1K: 0.0085; 2K: 0.03; 4K: 0.04 |
apimart-nano-banana-2.1-ext | APIMart | 1K: 0.012; 2K: 0.014; 4K: 0.018 |
apimart-seedream-5-0-flash | APIMart | 0.0137144 |
GPT Image 2.5 Ext defaults to Flare; set version: "sunburst" in APIMart provider options to select Sunburst at the same current prices. Nano Banana 2.1 Ext accepts one image per request and disables fallback to the token-billed official model. Seedream 5.0 Flash generates one image per request. Existing image entries remain available.
The official catalog no longer includes gpt-5.4, gpt-5.5, claude-opus-4-7, or gemini-3-flash-preview. Refer to the connected service's model catalog for available IDs, current prices, and input limits.
Continue with:
Jev decision models
The official Edge Worker provides two Jev models with 64K context windows:
| Model ID | Provider | Global / Env credential |
|---|---|---|
teamorouter-jev-1.14.0 | TeamoRouter | TEAMOROUTER_API_KEY |
jev-latest | B.AI | B_API_KEY |
Administrators configure each credential in the Federation's Global / Env. TeamoRouter uses https://api.teamorouter.ai/v1/systemone with upstream model jev-1.14.0; B.AI uses https://api.b.ai/v1/decisions with upstream model jev-latest. Availability depends on the connected service catalog and configured credentials.
These models expose the decision action for Boolean, Choice, and Score results. Select them in Desktop's decision model list or Group dispatch model selector. They do not provide chat streaming. Choice accepts 2–255 options; Score ranges must map to 2–10 levels (round(max - min) + 1). Boolean results use a 0.5 probability threshold. Empty JSON objects and arrays are rejected as decision state.
Decision models are listed directly from the registered models; the persistent relay catalog manages streaming language models only. Provider token usage is forwarded when available. Neither model declares a Credits price in the official catalog.
Modelflare Gemini models
The official service provides gemini-4-argon through Modelflare's OpenAI-compatible API, with streaming text, tool calls, and replay of previous tool results.
| Model ID | Base input / cache read / output price (USD per million tokens) |
|---|---|
gemini-4-argon | 0.6 / 0.03 / 3 |
Prices use Modelflare's gemini-stable group. The catalog's pricing contains base rates. Billing multiplies the base cost by 1.2, converts at 1 USD = 1,000,000 Credits, and rounds up. Cache reads use the cached token usage received by the service. Bills identify the modelflare-openai channel.
Administrators configure MODELFLARE_API_KEY with access to the gemini-stable group in the Federation's Global / Env. Self-hosted services use https://modelflare.dev/v1 as the API base URL. The catalog declares context_window: 1048576, using Google's unified input and output limit. The Agent reserves output tokens and a safety margin within this window. This value does not represent a measured maximum request size through Modelflare. Reasoning efforts remain unspecified until verified for this channel. Check the connected service catalog for availability.
Krill GPT models
The official service provides separate Krill model entries:
| Model ID | Context window | Base input / cache read / output price (USD per million tokens) |
|---|---|---|
krill-gpt-5.6-terra | 372K | 1.6 / 0.16 / 9.6 |
krill-gpt-5.6-sol | 372K | 3.2 / 0.32 / 16 |
krill-gpt-5.6-luna | 372K | 0.16 / 0.016 / 0.96 |
krill-gpt-6-astra | 1M | 8 / 0.8 / 40 |
krill-gpt-6-sol | 1M | 1.6 / 0.16 / 8 |
krill-gpt-6.1-sol | 1M | 1.6 / 0.08 / 8 |
These entries support streaming text, tool calls, and low, medium, high, xhigh, and max reasoning efforts.
Each Krill model uses the APIMart prices for the same model ID without the krill- prefix, including input tiers, cache reads, and cache writes. The table shows base rates for requests with at most 272K input tokens. Total input includes cached tokens; above 272K, the entire request uses twice the input and cache rates and 1.5 times the output rate. Standard or 5-minute cache writes cost 1.25 times the input rate for the selected tier; 1-hour cache writes cost 2.5 times that rate. Usage is based on the metering received by the service.
The catalog's structured pricing contains tiered base rates and cache rates before markup. Billing multiplies the base cost by 1.2, converts at 1 USD = 1,000,000 Credits, and rounds up. Bills identify the Krill channel and the full krill- model ID.
Check the connected service catalog for availability.