Lumen
Docs Start free
Start free

Model family

Five models. One API. No renegotiation.

Swap models by changing a string. Prices are per million tokens, billed to the token — no rounding up to the nearest thousand, no minimum spend, and cached input costs 90% less.

The line-up

lumen-flash

Fastest

High-volume work where latency is the product: classification, routing, drafting, extraction. Roughly a quarter the price of pro at 85% of its quality on our internal evals.

Context128K
Input€0.80 / M
Output€2.40 / M

lumen-pro

Recommended

The default for agents. Best-in-family tool calling, a 400K window that swallows an entire codebase or contract set, and the steadiest instruction-following we ship.

Context400K
Input€2.60 / M
Output€10.40 / M

lumen-max

Deepest

Extended reasoning for problems worth waiting on — migration plans, multi-document synthesis, adversarial review. Slower and dearer on purpose; use it selectively.

Context1M
Input€9.00 / M
Output€36.00 / M

lumen-embed

Retrieval

1,024-dimension embeddings for search, clustering and deduplication. Matryoshka-truncatable to 256 dimensions when index size matters more than the last point of recall. €0.06 / M input.

lumen-guard

Safety

A small classifier for policy screening on the way in and on the way out. Returns a label, a confidence and the matched policy clause — never free text. €0.20 / M input.

Specifications

Context, latency and price

Latency is measured p50 time-to-first-token from Frankfurt over the last 30 days, at a 4K-token prompt.

Lumen model specifications, updated 1 August 2026. Prices in EUR per million tokens.
Model Context Max output Input / M Output / M Cached input / M p50 TTFT Best for
lumen-flash 128K 8K €0.80 €2.40 €0.08 180 ms Classification, routing, extraction, first-draft replies
lumen-pro 400K 32K €2.60 €10.40 €0.26 210 ms Agents, tool use, long documents, production RAG
lumen-max 1M 64K €9.00 €36.00 €0.90 640 ms Extended reasoning, migrations, adversarial review
lumen-embed 8K — €0.06 — — 40 ms Vector search, clustering, near-duplicate detection
lumen-guard 32K 1K €0.20 €0.60 — 90 ms Policy screening, PII detection, jailbreak defence

Cached input applies to any prefix repeated within five minutes — system prompts, tool definitions and retrieved context all qualify. Most agent workloads land between 60% and 80% cache hit rate.

Capability matrix

What each model can actually do

Capabilities are enforced server-side. Calling one a model does not support returns 400 unsupported_capability rather than degrading silently.

Feature support by model.
Model Tool calling Parallel tools Vision JSON schema Streaming Prompt caching Fine-tuning
lumen-flash Yes Yes Yes Yes Yes Yes Yes
lumen-pro Yes Yes Yes Yes Yes Yes Yes
lumen-max Yes Yes Yes Yes Yes Yes No
lumen-embed No No No No No No Yes
lumen-guard No No Yes Yes No No Yes

Choosing

Start on pro. Move down, not up.

Build against lumen-pro until the behaviour is right, then try lumen-flash on the same evaluation set. Most teams find half their traffic runs fine on flash — that is where the 31% average saving comes from.

  1. Write the eval first

    Fifty real examples with the answer you want. Without it, a model change is a coin flip.

  2. Route by difficulty

    Send the easy 60% to flash and escalate only when confidence drops below your threshold.

  3. Reserve max for the rare case

    Deep reasoning on 2% of traffic barely moves the bill and noticeably moves quality.

router.py 
# Escalate only when the cheap model is unsure.
def answer(question):
    draft = client.run(
        model="lumen-flash",
        input=question,
        confidence=True,
    )

    if draft.confidence >= 0.82:
        return draft

    # ~9% of traffic reaches this line.
    return client.run(
        model="lumen-pro",
        input=question,
        context=draft.reasoning,
    )

Rate limits

Limits by plan

Limits are per workspace, not per key. Hitting one returns 429 with a Retry-After header — the SDKs back off for you.

Requests per minute, tokens per minute and concurrency by plan.
Plan Requests / min Tokens / min Concurrent runs Batch queue Support
Free 60 40,000 4 — Community
Starter 600 400,000 25 10K jobs Email, 1 business day
Scale 4,000 3,000,000 200 250K jobs Shared Slack, 4 hours
Enterprise Custom Custom Custom Unlimited Named engineer, 1 hour

Need more? Limits are soft above Starter — write to us with a traffic estimate and we will raise them, usually the same day.

Batch runs cost 50% less Anything that can wait an hour — nightly enrichment, backfills, evaluation sweeps — belongs in the batch queue at half the per-token price.

Try every model on the free tier

Five million tokens a month across the whole family, including max. No card, no quota negotiation.

Get an API key Estimate your bill