> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orbitrage.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Models
> The model catalog — tiers, providers, pricing, vision, and context windows.
Orbitrage's catalog has two access modes, and every chat model is in exactly one
of them:
| Access mode | Which models | Who bills you |
| -------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------- |
| **Served by Orbitrage** (pooled) | Every open-weight model — the Amazon Bedrock catalog, Fireworks, Cerebras, and the open Azure AI Foundry deployments (MiniMax / Kimi / DeepSeek) | Orbitrage credits, at the upstream price + a **2.5%** infra fee |
| **BYOK-only** | Every closed-weight frontier chat model: `claude-*`, `gpt-*` (except `gpt-oss-*`), `gemini-*`, `grok-*` | **Your provider**, directly, at your rate. Orbitrage charges **\$0** |
BYOK models are still fully first-class — same routing, tracing, cost analytics,
guardrails and Tools Gateway — they just need an enabled provider key. Calling one
without a key returns `403 byok_key_required`; there is no silent pooled fallback.
See [BYOK](/concepts/byok).
Provider **brand** does not determine access mode. OpenAI's `gpt-oss-*` models
(including `gpt-oss-safeguard-*`) and Google's `gemma-*` models are open-weight —
Orbitrage serves and bills those normally. Only the closed frontier lines listed
above are BYOK.
Let the router pick (`model: "auto"`) or pin any model by id. Auto routing only
ever selects models your organization can actually reach, so it never returns
`byok_key_required`.
Open-source models are addressed by their **bare id** — e.g. `gemma-3-27b-it`,
not `google.gemma-3-27b-it`. Orbitrage maps it to the upstream id for you.
The catalog below is representative. The **live, authoritative list** for your
account is always `GET https://api.orbitrage.ai/v1/models` — see
[List models](/api-reference/list-models). Each entry carries a `byok_required`
flag so you can tell the two modes apart programmatically.
## Pricing & tiers
Prices are USD per **1M tokens** (input / output).
* For **served** models, that's the upstream price; your credits are debited at
that price plus the 2.5% infra fee.
* For **BYOK** models, that's the **provider's own list rate** — shown for
comparison only. Your provider bills you (possibly less, at your negotiated
rate); Orbitrage debits nothing.
### Frontier — BYOK only
Requires an enabled **Anthropic** key.
| Model | Provider | Input | Output | Vision | Context |
| ------------------- | --------- | ------- | ------- | ------ | ------- |
| `claude-fable-5` | Anthropic | \$10.00 | \$50.00 | ✓ | 200K |
| `claude-opus-4-8` | Anthropic | \$5.00 | \$25.00 | ✓ | 200K |
| `claude-opus-4-7` | Anthropic | \$5.00 | \$25.00 | ✓ | 200K |
| `claude-sonnet-5` | Anthropic | \$2.00 | \$10.00 | ✓ | 1M |
| `claude-sonnet-4-6` | Anthropic | \$3.00 | \$15.00 | ✓ | 200K |
Requires an enabled **OpenAI** key.
| Model | Provider | Input | Output | Vision | Context |
| -------------- | -------- | ------- | ------- | ------ | ------- |
| `gpt-5.5` | OpenAI | \$5.00 | \$20.00 | ✓ | 400K |
| `gpt-5.4` | OpenAI | \$1.25 | \$5.00 | ✓ | 400K |
| `gpt-5.3-chat` | OpenAI | \$1.50 | \$6.00 | ✓ | 400K |
| `gpt-5.4-nano` | OpenAI | \$0.20 | \$1.25 | ✓ | 256K |
| `gpt-5.4-mini` | OpenAI | \$0.10 | \$0.40 | ✓ | 256K |
| `gpt-5-nano` | OpenAI | \$0.10 | \$0.40 | ✓ | — |
| `gpt-4o` | OpenAI | \$2.50 | \$10.00 | ✓ | 128K |
| `gpt-4o-mini` | OpenAI | \$0.165 | \$0.66 | ✓ | — |
The full BYOK list also includes `gpt-5.5-pro`, `gpt-5.4-pro`, `gpt-5.3`,
`gpt-5.3-codex`, `gpt-5.2`, `gpt-5.2-codex`, `gpt-5.1`, `gpt-5`, `gpt-5-mini`,
`gpt-chat-latest`, `gpt-4-turbo`, `gpt-4`, `gpt-3.5-turbo`, `o1`, `o1-mini`, `o3`,
`o3-mini` and `o4-mini`.
Requires an enabled **Google** key: the `gemini-*` family.
Requires an enabled **xAI** key: the `grok-*` family (`grok-4`, `grok-4.3`,
`grok-4-fast`, `grok-3` variants, …).
`claude-sonnet-5` is priced at its **introductory** rate of $2 / $10 per 1M tokens
through Aug 31, 2026 (standard $3 / $15 after). Like Opus 4.7+/Fable 5 it uses
adaptive thinking and ignores `temperature`/`top_p` — Orbitrage strips them for you.
`claude-fable-5` is Anthropic's top-tier model (above Opus) — the highest-quality
option in the catalog, used for the hardest reasoning and agentic tasks.
### High — served by Orbitrage
| Model | Provider | Input | Output | Vision | Context |
| --------------- | ------------- | ------ | ------ | ------ | ------- |
| `Kimi-K2.6` | Azure Foundry | \$0.95 | \$4.00 | ✓ | 200K |
| `DeepSeek-V3.2` | Azure Foundry | \$0.40 | \$1.10 | — | 131K |
| `glm-5.2` | Fireworks | \$1.40 | \$4.40 | — | 1M |
### Mid — served by Orbitrage
| Model | Provider | Input | Output | Vision | Context |
| ------------------- | ------------- | ------ | ------ | ------ | ------- |
| `FW-MiniMax-M2.5` | Azure Foundry | \$0.30 | \$1.20 | — | 262K |
| `DeepSeek-V4-Flash` | Azure Foundry | \$0.14 | \$0.28 | — | 131K |
| `minimax-m3` | Fireworks | \$0.30 | \$1.20 | ✓ | 512K |
### Basic — served by Orbitrage
| Model | Provider | Input | Output | Vision | Context |
| --------------------- | -------- | ------ | ------ | ------ | ------- |
| `gpt-oss-20b` | Bedrock | \$0.07 | \$0.30 | — | 128K |
| `nemotron-nano-9b-v2` | Bedrock | \$0.06 | \$0.23 | — | 128K |
| `glm-4.7-flash` | Bedrock | \$0.07 | \$0.40 | — | 203K |
| `gemma-3-4b-it` | Bedrock | \$0.04 | \$0.08 | ✓ | 128K |
### Open source (Amazon Bedrock, served by Orbitrage)
A broad open-weights catalog, billed to your Orbitrage credits with the same
2.5% infra fee. Address each by its **bare id** (the engine adds the vendor prefix
upstream). Vision models accept image content blocks.
| Model | Vendor | Input | Output | Vision | Context |
| -------------------------------- | -------- | ------ | ------ | ------ | ------- |
| `gemma-3-27b-it` | Google | \$0.23 | \$0.38 | ✓ | 128K |
| `gemma-3-12b-it` | Google | \$0.09 | \$0.29 | ✓ | 128K |
| `gemma-3-4b-it` | Google | \$0.04 | \$0.08 | ✓ | 128K |
| `qwen3-vl-235b-a22b-instruct` | Qwen | \$0.53 | \$2.66 | ✓ | 256K |
| `qwen3-coder-next` | Qwen | \$0.50 | \$1.20 | — | 256K |
| `qwen3-coder-480b-a35b-instruct` | Qwen | \$0.45 | \$1.80 | — | 128K |
| `qwen3-coder-30b-a3b-instruct` | Qwen | \$0.15 | \$0.60 | — | 256K |
| `qwen3-next-80b-a3b-instruct` | Qwen | \$0.14 | \$1.20 | — | 256K |
| `qwen3-235b-a22b-2507` | Qwen | \$0.22 | \$0.88 | — | 256K |
| `qwen3-32b` | Qwen | \$0.15 | \$0.60 | — | 32K |
| `deepseek-v3.2` | DeepSeek | \$0.62 | \$1.85 | — | 164K |
| `deepseek-v3.1` | DeepSeek | \$0.58 | \$1.68 | — | 128K |
| `glm-5.2` | Z.AI | \$1.40 | \$4.40 | — | 1M |
| `glm-5` | Z.AI | \$1.00 | \$3.20 | — | 200K |
| `glm-4.7` | Z.AI | \$0.60 | \$2.20 | — | 203K |
| `glm-4.7-flash` | Z.AI | \$0.07 | \$0.40 | — | 203K |
| `glm-4.6` | Z.AI | \$0.60 | \$2.20 | — | 203K |
| `kimi-k2.7-code` | Moonshot | \$0.95 | \$4.00 | ✓ | 262K |
| `kimi-k2.5` | Moonshot | \$0.60 | \$3.00 | ✓ | 256K |
| `kimi-k2-thinking` | Moonshot | \$0.60 | \$2.50 | — | 256K |
| `minimax-m3` | MiniMax | \$0.30 | \$1.20 | ✓ | 512K |
| `minimax-m2.5` | MiniMax | \$0.30 | \$1.20 | — | 196K |
| `minimax-m2.1` | MiniMax | \$0.30 | \$1.20 | — | 196K |
| `minimax-m2` | MiniMax | \$0.30 | \$1.20 | — | 1M |
| `nemotron-super-3-120b` | NVIDIA | \$0.15 | \$0.65 | — | 256K |
| `nemotron-nano-3-30b` | NVIDIA | \$0.06 | \$0.24 | — | 256K |
| `nemotron-nano-12b-v2` | NVIDIA | \$0.20 | \$0.60 | ✓ | 128K |
| `nemotron-nano-9b-v2` | NVIDIA | \$0.06 | \$0.23 | — | 128K |
| `mistral-large-3-675b-instruct` | Mistral | \$0.50 | \$1.50 | ✓ | 256K |
| `devstral-2-123b` | Mistral | \$0.40 | \$2.00 | — | 256K |
| `magistral-small-2509` | Mistral | \$0.50 | \$1.50 | ✓ | 128K |
| `ministral-3-14b-instruct` | Mistral | \$0.20 | \$0.20 | ✓ | 128K |
| `ministral-3-8b-instruct` | Mistral | \$0.15 | \$0.15 | ✓ | 128K |
| `ministral-3-3b-instruct` | Mistral | \$0.10 | \$0.10 | ✓ | 128K |
| `voxtral-small-24b-2507` | Mistral | \$0.10 | \$0.30 | — | 32K |
| `voxtral-mini-3b-2507` | Mistral | \$0.04 | \$0.04 | — | 32K |
| `palmyra-vision-7b` | Writer | \$0.15 | \$0.60 | ✓ | 4K |
| `gpt-oss-120b` | OpenAI | \$0.15 | \$0.60 | — | 128K |
| `gpt-oss-20b` | OpenAI | \$0.07 | \$0.30 | — | 128K |
| `gpt-oss-safeguard-120b` | OpenAI | \$0.15 | \$0.60 | — | 128K |
| `gpt-oss-safeguard-20b` | OpenAI | \$0.07 | \$0.20 | — | 128K |
### Ultra-fast (Cerebras — Pro plan only)
Cerebras runs open models on wafer-scale hardware at **\~1,000–3,000 tokens/sec** —
near-instant responses. These are addressed by a distinct **`-cerebras`** id so
they never collide with the standard (Bedrock) versions.
Cerebras models require the **Pro plan**. A free-plan request for one returns
`403 pro_plan_required`; on the [Models page](https://app.orbitrage.ai/models)
they're marked **Pro only** and can't be enabled. They're also explicit-pin
only — `model: "auto"` never routes to them.
| Model | Backing model | Speed | Input | Output | Vision | Context |
| ----------------------- | ------------- | ------------ | ------ | ------ | ------ | ------- |
| `gpt-oss-120b-cerebras` | GPT-OSS 120B | \~3000 tok/s | \$0.50 | \$1.38 | — | 128K |
| `glm-4.7-cerebras` | Z.AI GLM 4.7 | \~1000 tok/s | \$1.20 | \$4.40 | — | 128K |
| `gemma-4-31b-cerebras` | Gemma 4 31B | \~1850 tok/s | \$0.70 | \$1.10 | ✓ | 128K |
### Image
| Model | Provider | Billing |
| ------------- | -------- | --------- |
| `gpt-image-2` | Azure | Per image |
### Audio (managed, included)
Speech-to-text and text-to-speech are offered as a managed service on
**Deepgram** — no BYOK needed. See [Audio](/examples/audio).
| Model | Provider | Type | Billing |
| --------------------- | -------- | -------------- | -------------------- |
| `nova-3` | Deepgram | Speech-to-text | Per minute of audio |
| `nova-3-multilingual` | Deepgram | Speech-to-text | Per minute of audio |
| `nova-2` | Deepgram | Speech-to-text | Per minute of audio |
| `aura-2-thalia-en` | Deepgram | Text-to-speech | Per 1,000 characters |
The full catalog also includes the DeepSeek R1 reasoning model and the whole
open-source Bedrock catalog above (served), plus the remaining GPT-5, Gemini and
Grok variants (BYOK). Query `/v1/models` for everything enabled on your account —
each entry's `byok_required` flag tells you which mode it's in.
## Cost, savings, and the infra fee
Every call records four cost figures, so you can see exactly where your money goes:
| Field | Meaning |
| ----------------------- | -------------------------------------------------------------------------------------------------------- |
| **Cost** (`cost_usd`) | What Orbitrage bills you — upstream price plus the 2.5% infra fee. Always `0` on a BYOK call. |
| **Provider cost** | The raw upstream price Orbitrage paid (internal). `0` on a BYOK call — your provider billed you, not us. |
| **Baseline cost** | What the same call would have cost on a frontier baseline (`claude-sonnet-4-6`). |
| **Saved** (`saved_usd`) | `baseline − cost` — the routing savings, never negative. |
The baseline lets the dashboard answer "how much did routing save me?" by
comparing every routed call against a single frontier model.
`cost_usd = 0` covers the **LLM tokens** only. If the turn invoked the
[managed Tools Gateway](/concepts/tools-gateway), those tool calls ran on
Orbitrage's pooled tool keys and are billed normally — so a BYOK turn's cost is
exactly its tool cost.
## Vision & multimodal
Vision-capable models accept image content blocks in the standard OpenAI format,
and the router prefers one automatically when your prompt contains images. The
engine also exposes image generation/editing and audio (transcription,
translation, speech) endpoints — see
[the API reference](/api-reference/introduction).
## Pinning vs. aliases
* **Pin** a model by passing its exact id — no scoring, straight to that model.
* **Route** by passing `auto` (or `router` / `default` / `orbitrage`).
* Model ids match flexibly — `claude-sonnet-4.6` and `claude-sonnet-4-6` resolve
to the same model, and family names (e.g. `claude-opus`, `grok-4`) resolve to
the appropriate provider and tier.
* Pinning a **BYOK** model without an enabled key for its vendor fails fast with
`403 byok_key_required` rather than quietly routing somewhere else. Pin an
open-weight id, or use `auto`, if you'd rather never see that error.