The model index: every lane on one metered API
This page is the index of the platform's whole offering: forty-six model entries across ten lanes. The lanes are chat, text to speech, speech to text, embeddings, rerank, document parsing, classification, image generation, and the two robot lanes, action and vision. All of them sit behind one OpenAI-compatible gateway, one console, and one prepaid wallet. Every entry links to its own page with the real numbers and the copy-paste request. What this index adds over a plain list is the state of each row. What is live, what is a written record beyond the list, and what was discontinued.
Last verified: 2026-10-05
This page leaves two things out on purpose. First, rates. Every listed model's rate renders on the pricing page, generated from the platform's own model data. That is the same source the API meters against, so a figure typed into prose here could only go stale. The GLM-OCR parse row and the six later-card robot rows are not on the public list; their pages are written records that carry their own numbers. Second, per-model spec tables. The listing is built by machine, and each model page owns its own working contract. The map starts with the product itself.
What an AI model API is
Concretely: an HTTP service that takes a request naming a model, runs it, and returns the output plus a usage record. For chat, the settled request shape is the one in OpenAI's chat completions reference. Name a model, send a messages array, receive a completion with a usage object. That shape has become the practical standard across hosted providers and local runtimes alike. That is why an integration usually moves between hosts by changing two strings: the endpoint base URL and the model name.
The part every index owes you and most skip: what a call is metered in, and whether the row actually serves. Token prices tell you nothing about a speech lane billed by the character or a transcription lane billed by the minute of audio. And an index that lists everything as online, when several of its rows answer nothing today, is not a map. This one names the unit for each lane below and states each row's state in a fixed vocabulary. Those two facts decide budget and build order more than any leaderboard row.
The catalog: ten lanes
Chat
The chat lane holds seventeen open-weight entries from ten publishers on one chat completions route. All of them serve; the two robot reasoning entries joined the list on October 6, 2026, and the paragraph below covers them. Qwen 3.8 27B, the dense one, takes images and video alongside text and ships with per-request thinking control, according to the official card. Sparse Qwen3.6-35B-A3B carries the throughput load for agentic coding. Muse Glimmer 30B builds failure recovery into the model itself, exactly what an unattended agent wants. MiniMax M2.7 is built for long tool-calling chains. And the Ornith 1.5 35B and 9B pair, from a family that generates its own training tasks, serves from the platform's own hardware.
Four more joined them as serving rows on September 11, 2026. GLM 5.3 Flash is a sparse multimodal build from Z.ai. DeepSeek V4 Flash 0731 and DeepSeek V4 Pro are the general-availability snapshots. Gemma 4 31B is Google's dense reasoning model. Their pricing went live with their rows.
Two more flagships joined the list on September 25, 2026. GLM 5.3 is Z.ai's flagship itself and the big sibling of the Flash build. DeepSeek V4.1 Flash is the successor snapshot to the 0731 lane. Both are live and both render on the public list. The three decision lanes from the classification family, Jev fast, Jev causal, and Jev batch, also serve on this route; the classification page carries what they are for.
Two robot reasoning entries joined the route's pages in October 2026, and both went live on October 6, 2026. Molmo 2 ER reads images and answers in plain text, an embodiment-aware reasoning build for robot work. Cosmos Reason2 2B is the compact fast-reasoning twin, built for Jetson-class edge hardware. Both are on the public list, and both serve, confirmed by the October 6, 2026 requests.
Text to speech
The speech lane is four open-weight models on one speech route, all billed by the character. Kokoro handles steady narration at volume. Chatterbox covers character voices and short-form expressive work. Pocket TTS is a compact build that runs on laptop CPUs, which also makes it the portable one. Audio8 TTS rounds out the set with multilingual reach. All four are live, and every completed call reports its own charge in characters.
Speech to text
The transcription lane holds two serving entries and one discontinued row. Hayamimi STT is the batch transcription model: recorded audio in, text out, billed by audio time. It is live. The whisper batch lane returned to the catalog on 2026-09-25, served from the platform's own hardware after its 2026-09-24 pull for a market-rate reason. It is live. VASR streaming is the live-streaming entry, built around voice activity and speaker labeling. Its row is discontinued right now. The reasons and return conditions sit in the state section below.
Embeddings
The embeddings lane turns text into vectors for semantic search, retrieval, and classification. Its engine is BGE M3, an open multilingual embedding model, metered by the token. It is live. A second engine serves alongside it: WeMM-Embedding-2B, a multimodal build that maps text, images, and video into one vector space, metered by the token.
Rerank
The rerank lane scores a retrieved shortlist against a query with a cross-encoder and returns it reordered by relevance. Its engine is BGE Reranker v2 M3, metered one search per request. It is live.
Document parsing
The parsing lane turns page images, scans, and PDFs into text a program can use, and it is the youngest lane. PP-OCR print is the volume text-out route for machine-printed pages. Archival OCR is the fidelity route for old print, with a flag that keeps original spelling and typography. GLM-OCR parse mode is the structured route, parsing documents to Markdown, priced on the token axis. Print and archival are on the public list and serve, confirmed by the October 26, 2026 requests. The parse-mode row is not on the list; its page is a written record that names its rate.
Classification
The classification lane serves three route families over open weights. IAB 2.x and IAB 3.1 handle taxonomy labels, with the 2.x-to-3.1 cross-map on the 3.1 lane. GLiNER extract does zero-shot typed extraction. The decision family, Jev fast, Jev causal, and Jev batch text, scores choices instead of generating prose; those three serve on the chat completions route. All six are live, and each page carries its exact request shape.
Image generation
The images lane generates images from text prompts on one OpenAI-shaped route, metered per image, flat in size. Flux 1 Schnell is the family's fast tier. Flux Pro 1.1 is the quality tier, carrying explicit width, height, and guidance-scale levers. Both are live.
Robot action
The robot action lane is the vision-language-action family: camera frames and an instruction go up, an action chunk for the arm comes back, with the normalization block the policy applied stated in every response. Three entries serve, on the public list since October 26, 2026 and confirmed by that day's requests: SmolVLA, the OpenVLA fine-tune, and ACT. Five sit beyond the list as written records: the pi 0.5 family with its DROID and LIBERO fine-tunes, MolmoAct2 for the SO-100 and SO-101 arms, and the compact X-VLA. The route's contract is published and buildable today.
Robot vision
The robot vision lane serves the perception side: segmentation masks and depth maps from images, on one route, metered per image. Two of its three entries serve as of October 6, 2026: SAM 2.1 base for promptable segmentation and Depth Anything 3 small for relative depth, both confirmed by the October 6, 2026 requests. SAM 2.1 large is not on the public list; its page is a written record that carries its rate and its request shape.
The state vocabulary
One word per row class, and the platform's own model data decides which word applies. The listing table on this page follows that data, so the words below describe classes, not a frozen snapshot.
| State | What it means | Where it applies |
|---|---|---|
| Serving | The row is listed on this index and the pricing page, and serves on its route | Thirty-three rows as of October 6, 2026: fifteen chat, two self-hosted Ornith, and sixteen specialty, with the two robot vision rows counted in the specialty set |
| Not listed | The model's page exists as a written record: capability, contract, numbers, and local roads. Nothing on it is an offer | Twelve pages as of October 6, 2026: the parsing set and the nine later-card robot rows across the action and vision lanes |
| Discontinued | The row was removed and returns only under named conditions | The VASR streaming entry |
Two of those classes deserve their honest detail. The not-listed pages are real and described, not vapor. Each parsing page and each robot page says what the model does and what its numbers are, as a written record. And the discontinued row carries its cause. VASR streaming: the current checkpoint emits no per-speaker output fields, misattributes in-text speaker markers, and the streaming edge caps connections at roughly 220 seconds. It returns only when all three are fixed. The fixes: a checkpoint with real per-speaker output, a passing three-speaker case with correct attribution, and that cap resolved or chunked sessions built. The whisper batch row's exit and return both carry their dates: pulled 2026-09-24 rather than priced above the market, restored 2026-09-25 on the platform's own hardware.
Why publish a vocabulary at all? Because the failure mode of model indexes is silence. Rows vanish, deprecation arrives as a surprise, and every entry lists as online regardless of truth. Buyers notice. This platform states what a row is so you can plan around it. The listing is the signal: what it carries serves, and what it omits is not on offer through this platform.
One key, one wallet, every route
The routes, at the level an index needs:
| Lane | Route | Metered in |
|---|---|---|
| Chat | POST /v1/chat/completions | Tokens: input, cached input, output |
| Text to speech | POST /v1/audio/speech | Characters of input |
| Speech to text | POST /v1/audio/transcriptions | Audio time processed |
| Embeddings | POST /v1/embeddings | Tokens |
| Rerank | POST /v1/rerank | Searches, one per request |
| Parsing | POST /v1/ocr | Pages |
| Classification | POST /v1/classification | Pages (taxonomy lanes), searches (extraction), tokens (decisions) |
| Image generation | POST /v1/images/generations | Images generated |
| Robot action | POST /v1/action | Calls, one per request |
| Robot vision | POST /v1/vision | Images in, billed per image |
One Bearer key from the console reaches all of them, and a read of the voice catalog costs nothing. Every completed call reports its own charge. A workload's cost shows up request by request instead of on an invoice. The wallet runs on prepaid credit, and the balance is the ceiling. Once it reads zero, the next call is turned away with a named error. Nothing is billed for it. Calls the gateway refuses before running never touch the wallet either. The stop reads the balance when a call arrives, not the cost of the call. The last request on a nearly empty wallet can finish slightly overdrawn. The pricing page publishes that edge.
Keys are revocation boundaries. Splitting projects across keys contains any leak to a single key. One key's revocation leaves the others and the wallet untouched. Per-key spending caps are on the roadmap and are not a shipped feature; the loaded balance is today's ceiling. The bill-shock guide named below collects the habits that keep metered budgets predictable across all ten lanes.
How to use this index
- Name the job, then follow the lane. Each lane paragraph above links its category page, which carries the deeper map for that family. Every model page ships three code tabs, curl and Python plus the third tab its lane speaks: the OpenAI SDK on the chat-family lanes, a LeRobot tab on the robot action lane, and a plain HTTP tab on the robot vision lane.
- Check the state before you commit a build. The vocabulary above is the same on every page in this catalog. If the row you need is serving, the model page's tabs work today. If it is not listed, the page is a written record: capability, numbers, and local roads. A discontinued row carries its named return conditions.
- Read rates once, on the pricing page. Rates render there for every listed row that has one, the parsing and robot pages beyond the list carry their own numbers, and a change appears there first. Counts here are dated facts: forty-six entries is the last-verified figure, and the platform, not this page, adds rows.
- Moving an existing integration. The chat lane speaks the OpenAI-compatible shape, so the port is two edited strings and a test run. The OpenAI-compatible API guide, one of the two named below, walks that port from the first tab to the last.
Getting started
- Use the console's invite flow to get in. That gate holds while the beta runs. The console holds the wallet, issues keys, and keeps the spend history, and an account created today works across every lane on the list.
- Create a key. One key per project keeps revocation surgical.
- Make the first call from a model page. Pick your lane above, open the model page that matches the job, and run its curl tab. On a listed row the tabs work today; a page beyond the list describes its model as a written record.
One source sits under all of it: the platform's own model data. It drives the listing table on this page, the rates the pricing page renders, and the units every call reports. The words on this page are written to stay true to it.