Models
Keiro currently exposes four public paid-beta eb1 model IDs. If you are unsure which to use, start with eb1-preview and change only after measuring your workload.
On this page 1 of 7
Choose a public eb1 model
The current eb1 models are the first generation, named Keiro Étude. The next generation is in training.
Public models#
| Model | Choose it for | Tradeoff |
|---|---|---|
eb1-preview | General reasoning, coding, and application work | Recommended balanced default |
eb1-frontier-preview | The hardest reasoning and synthesis workloads | Prioritizes capability over latency and spend |
eb1-fast-preview | Interactive and latency-sensitive requests | Prioritizes speed over maximum quality |
eb1-efficient-preview | Repeated routine work and cost-aware throughput | Prioritizes efficiency and predictable spend |
These strings are executable API identifiers. Do not derive or invent other model names from them.
Reasoning effort#
Each eb1 model chooses how much reasoning a request needs. A reasoning effort on the request is a ceiling on that choice: eb1 never reasons above it and may reason less. Send it as reasoning.effort on Responses, reasoning_effort on Chat Completions, or output_config.effort on Messages. minimal caps at low. none and ultra set no cap: like a request without an effort, eb1 chooses the depth on its own, and none does not turn reasoning off. A Messages thinking.budget_tokens is not a ceiling.
On eb1-frontier-preview, ultra does more than remove the cap. It opts the request into ultra requests, which use more models, cost more and can take many minutes.
List available models#
Model availability is account- and key-scoped:
printf 'Keiro API key: '
IFS= read -rs KEIRO_BEARER
printf '\n'
curl -sS https://api.keirolabs.ai/v1/models \
-H @- <<<"Authorization: Bearer $KEIRO_BEARER" | python3 -m json.tool
Use only IDs returned for the key. Each entry in data carries:
id: the public model ID to send.object: always"model".created: always0.owned_by:keirofor eb1 models.descriptionandstatus: optional descriptive strings. Do not parse them.context_window: on eb1 models, a prompt of up to this many tokens, as Keiro meters them, is accepted on every request to the model. See Errors forcontext_length_exceeded.max_input_tokens: the same number under the field name Messages-format clients read.rate_limits, when present: configured Keiro limits for the requesting account and API key, not upstream provider quotas or live serving capacity.
There is no per-model path: GET /v1/models/{model} returns 404, so an SDK's model-retrieve call fails. Read the model's entry from the list instead.
Select by measurement#
Evaluate the same representative prompts against the public models you are considering. Compare output quality, end-to-end latency, token usage, and billable amount in the console. Keep safety checks and tool/image workflows in the evaluation; text-only prompts do not prove capability parity for an agentic application.
Pin the chosen public ID in production configuration. A silent fallback to a different ID hides access and compatibility errors.
Missing access or capability#
If an expected model is missing, check the selected console organization and the key's restrictions, then contact support. If an image or tool request fails with a capability error, select another account-enabled public model or ask support to confirm the contract. Do not remove required input silently.