Models & Usage

Riven exposes a small, stable set of brand model IDs. Behind each ID, Riven routes to the best available backend — including automatic failover — so your integration never chases provider model-name churn. List the live catalog anytime with GET /v1/models, or fetch names, ordering, and multipliers from GET /v1/catalog?view=tiers (no auth required).

The Riven tier family — start here

Seven intent-based tiers cover the full speed/quality range. These are the same lanes shown in the model picker in the Riven apps — the friendly name on the left is what you see there. Pick by what the work needs: riven-best (Riven Auto) when you want Riven to choose; riven-instant when latency matters most; riven-max when quality matters most.

NameModel IDToken multiplierAvailabilityDescription
Riven Autoriven-bestrouted laneAll plansAuto mode — picks the right lane for each message (Spark, Bolt, Nova, Atlas, Research, or Code). Usage is billed at whatever lane it routes to. The recommended default in the apps.
Riven Sparkriven-instant1xAll plansThe fastest lane. Sub-second responses for high-volume, latency-sensitive traffic — autocomplete, classification, quick answers. Free tier.
Riven Boltriven-fast1.5xAll plansQuick everyday chat with strong quality at low cost. A good default for user-facing assistants.
Riven Novariven-core3.75xAll plansThe balanced flagship tier — the recommended pinned lane for general workloads and production integrations.
Riven Atlasriven-pro6.7xPaid plansDeep reasoning for hard, multi-step problems. Spends more time thinking; expect higher latency and cost.
Riven Researchriven-research2.3xPaid plansWeb-grounded research with citations. Answers are backed by live sources rather than model memory alone.
Riven Apexriven-max13.3xPaid plansThe most capable model Riven offers, with a very large context window. For the hardest work, where quality matters more than cost.

Specialty tiers

NameModel IDToken multiplierAvailabilityDescription
Riven Coderiven-code1xPaid plansCode generation, review, and refactoring workloads.
Riven Visionriven-vision1.25xPaid plansMultimodal chat over image inputs.

Requesting a model outside your plan returns HTTP 429 with {"error": {"code": "rate_limit_exceeded", "type": "plan_limit", ...}} and an upgrade pointer — see Errors.

How usage is metered

Every plan comes with a monthly token allowance. Each request consumes tokens used × the tier's token multiplier from that allowance. A 1,000-token answer on Riven Spark (1x) costs 1,000 allowance tokens; the same answer on Riven Apex (13.3x) costs about 13,300. Riven Auto is billed at whatever lane it routes each message to, so easy messages stay cheap automatically.

Pay-as-you-go keys are metered the same way against a base rate of $3.00 per million tokens: effective price = base rate × tier multiplier. Direct frontier models on pay-as-you-go are billed at vendor list rates instead of multipliers.

Tiers vs. direct frontier models

The two halves of the catalog carry different stability promises. Tiers are consumer lanes: the name and intent are stable, and Riven may upgrade the backing model behind a lane at any time so it always reflects the best available option. Direct frontier models are the developer contract: exact vendor IDs (for example gpt-5.6 or claude-fable-5) at vendor list pricing, deprecated only with notice. The exact set changes as vendors ship — GET /v1/models is always the source of truth. Subscription plans use the Riven tiers above.

Embeddings & image generation

Model IDTypeAvailabilityDescription
riven-embedEmbeddingsAll plans768-dimension text embeddings for search, clustering, and RAG. Accepts a string or an array of strings.
riven-embed-fastEmbeddingsAll plansSame 768-dimension embedding space, tuned for throughput on batch workloads.
riven-imageImage generationAll plansText-to-image generation. Returns base64-encoded PNG (1024×1024 default) via POST /v1/images/generations.

Legacy model IDs

Earlier integrations used the IDs below. They remain routable so nothing breaks, but they are no longer the recommended targets — each row notes the tier to use for new work.

Model IDTypeAvailabilityDescription
riven-ultraChatAll plansLegacy flagship ID. Still served; new work should target riven-core.
riven-ultra-proChat / ReasoningAll plansLegacy capability ID with optional extended-thinking mode (see note below). New work should target riven-pro or riven-max.
riven-ultra-researchChat / ResearchAll plansLegacy research ID. New work should target riven-research.
riven-chat-smChatAll plansLegacy small chat tier. New work should target riven-instant.
riven-chat-mdChatAll plansLegacy mid chat tier. New work should target riven-fast.
riven-chat-lgChatPaid plansLegacy large chat tier. New work should target riven-core.
riven-reasonReasoningPaid plansLegacy reasoning ID. New work should target riven-pro.
riven-agentAgents / ToolsAll plansTuned routing for agentic tool-use loops and function calling.
riven-ultra-pro thinking mode: the Pro backend can spend tokens on internal reasoning before answering. With small max_tokens budgets this can consume the entire budget (empty content, finish_reason: "length"). Either allow ≥2000 tokens or disable it per-request:
{"model": "riven-ultra-pro", "chat_template_kwargs": {"enable_thinking": false}, ...}

Listing models

curl
curl https://api.rivenai.io/v1/models
response (abridged)
{
  "object": "list",
  "data": [
    { "id": "riven-instant", "object": "model", "owned_by": "riven" },
    { "id": "riven-core", "object": "model", "owned_by": "riven" },
    { "id": "riven-max", "object": "model", "owned_by": "riven" }
  ]
}

Responses to completions include the model that actually served the request and an x-riven-upstream header identifying the serving pool, so routed traffic is always attributable.