Cheaper Inference API

Create a chat completion

OpenAI-compatible. The request is optimized, served from the exact-match cache when possible, and billed against the wallet. Two opt-in request headers form the client contract Keak Code uses: `x-ci-model-lock` (serve exactly the named model) and `x-ci-stream-receipt` (receipt frames in the stream).

Query parameters

Request body ChatCompletionRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response ChatCompletionResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Create a text completion

OpenAI-compatible legacy completions.

Query parameters

Request body CompletionRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response CompletionResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Generate images

Generate one or more images. Streaming is not supported on this endpoint.

Request body ImageGenerationRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response ImageResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Edit images

Edit up to 10 source images with a 20 MiB decoded aggregate limit across the sources and optional mask. The OpenAI-compatible multipart form accepts `image` uploads and an optional PNG `mask`. The JSON extension accepts `input_references` containing HTTP(S) URLs or image data URLs; external URLs are passed only to compatible providers and are never fetched or stored by CheaperInference. Models are eligible only after a live edit canary; discover them through `capabilities.image_edit` on GET /v1/models.

Request body ImageEditJsonRequest

Response ImageResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Create a message

Anthropic Messages. `max_tokens` is required. Errors use Anthropic's envelope, which carries no `code` or `param`.

Query parameters

Request body AnthropicMessagesRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response AnthropicMessagesResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Count input tokens

Estimate the input tokens for a Messages request. Makes no upstream call and moves no money, so `max_tokens` is not required.

Request body CountTokensRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response CountTokensResponse

Successful response

Create a response

OpenAI Responses, the surface Codex uses. Stateless: `store` must be `false`, and `previous_response_id` and `conversation` are refused. Replay prior output items in `input` instead. Namespace tools may mix function and custom/freeform entries, including inside `input[].additional_tools`; namespace and call metadata are preserved in buffered and streamed output. Gemini 3 requests using `reasoning_effort: none` fail with code `unsupported_reasoning_effort`. To use Pro, send a POST request to /v1/responses with store: false and reasoning.mode: "pro". Set reasoning.effort separately to control how much reasoning the model does. Pro requires the Responses endpoint. Chat Completions returns HTTP 400 for Pro requests. The models gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol and gpt-6-astra support Pro mode. Set stream: true to receive output as it arrives, or stream: false to receive one complete response. These models use standard mode when reasoning.mode is omitted. For gpt-5.5-pro, the model ID already selects Pro. Omit reasoning.mode and set stream: false. This model does not support streaming. Your price, provider, region and data-retention settings still apply. The provider must also support the features in your request. If no compatible Pro provider is available, the API returns HTTP 503 without charging you. A provider response that does not confirm Pro is rejected without a charge. Other OpenAI Pro model IDs, model IDs with a date suffix, and Pro reasoning replay are not yet supported. This includes requesting reasoning.encrypted_content. Pro can take longer and use more tokens than standard mode. Billing includes the total input and output tokens used across its model work, so input_tokens can exceed your original prompt's token count. The gpt-5.5-pro model has its own pricing. For streaming requests, read the final event as well as the HTTP status. A stream can start with HTTP 200 and later report an error. Pro streams are charged only when a valid final response confirms Pro and includes token usage.

Query parameters

Request body ResponsesRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response ResponsesResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Ask a decision model typed questions

Ask predicate, choice and score questions about one input and receive probabilities instead of text. OpenAI Decisions API shape; the OpenAI SDKs work with only a base URL change. Billed on input tokens only. Streaming is not supported.

Request body DecisionRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response DecisionResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

List models

Every model this platform serves, with capabilities and prices. `pricing_version` is a SHA-256 over the list that was returned, so a filtered response validates its own cache and not the unfiltered one.

Query parameters

Response ModelList

Successful response

List models (public)

The advertised catalog, with live prices, and NO API KEY. This is what the marketing pages and any other surface that cannot hold a key read. Prices are resolved by the same rule that bills a request, so a quote here and the charge cannot disagree. `/v1/models` covers the same catalog for authenticated clients and adds marketplace-only models.

Response PublicModelCatalog

The catalog.

Get one model (public)

One catalog entry. The id may contain slashes (`google/gemini-3.5-flash-lite`) and is matched exactly, including case.

Query parameters

Response PublicModel

The model.

Price changes

Price moves at or after `since`, oldest first, so a client can detect changes without re-diffing the catalogue. Store the `changed_at` of the last change you processed and pass it as `since` on the next poll, then follow `next_cursor` to the end. Requires the `usage:read` scope.

Query parameters

Response PricingChangeList

Successful response

Count supply under a price ceiling

How many supply sources would currently serve a chat request carrying `min_discount_percent` and/or `zdr=true` for this model, using the same filter the request path applies. At least one of the two is required. Candidates are priced under the ceiling now; health and fill are not guaranteed.

Query parameters

Response ModelSupply

Successful response

Price a workload across models

What a given number of prompt and completion tokens would cost on each of up to ten models, cheapest first. Needs usage:read only. It discloses catalog pricing, nothing about the account.

Query parameters

Response ModelComparison

Successful response

List requests

Per-request history for the calling workspace, newest first. Requires the `usage:read` scope. `cursor` is opaque and comes from the previous page's `next_cursor`; do not parse it. With `request_ids`, the named requests are returned instead of a page: `has_more` is false, `next_cursor` is null and `missing_request_ids` lists the ids that did not come back.

Query parameters

Response UsageRequestList

Successful response

Daily spend

Daily spend for the calling workspace, defaulting to the last 30 days and zero-filled across the range so a chart does not close gaps. Requires the `usage:read` scope. Only settled requests contribute to spend.

Query parameters

Response UsageDaily

Successful response

Team Savings

What the calling key's workspace spent, against the list price of the same model (`baseline.id: same_model_list_price_v1`), in total and by member, project or model, over inclusive UTC days (at most 92). Requires `usage:read`; `scope=team` also requires `account:read`. Includes every key in the workspace, not only Keak Code; requests without a Keak Code project tag are Unattributed. Spend: a settled request at its charge, a failed request that was charged at its charge, a cache hit at $0; in-flight requests are counted as pending with no amount. Requests with no known list price are left out of the baseline and of the comparable spend, and `coverage` says how much. Savings may be negative. Amounts are 6-decimal strings rounded half-even one by one, so a figure derived from two others can differ by $0.000001; groups sum to the totals exactly before rounding. At most 500 groups are returned (`groups_truncated`); totals are always complete. A member is the owner of the key that sent the request.

Query parameters

Response TeamSavings

Successful response

Team Savings request records

The request records behind a Team Savings figure, newest first, with the same scope, filters and permissions as `GET /v1/usage/team-savings`. Records only: no prompt, response or request body. The charged pages of a period sum to the report's `actual_usd`. `cursor` is opaque and comes from the previous page's `next_cursor`.

Query parameters

Response TeamSavingsRequestList

Successful response

Score a request

Rate a completed request 1-5. Not present on the legacy service.

Request body FeedbackRequest

Response FeedbackResponse

Successful response

Get the account

Email, workspace and role, how many keys are live and when one was last used, and how many optimization techniques are actually on.

Response Account

Successful response

Get the wallet balance

Balance, what is reserved against requests in flight, and auto-recharge state. Spend against `available_usd`, not `balance_usd`.

Response AccountBalance

Successful response

Summarize usage over a window

Requests, billed, saved versus list price and the top models. The rolled-up counterpart to /v1/usage/daily's per-day series.

Query parameters

Response AccountUsage

Successful response

Get the savings snapshot

Savings over 24 hours, 7 days and 30 days, every experiment's per-request economics, and the technique catalog with this account's state on each.

Response AccountSavings

Successful response

Read the workspace admin policy

The admin policy of the calling key's workspace, and `can_edit` (owner or admin). Five settings bind every member, owners and admins included: a maximum task limit that clamps every task budget, an advisory default task limit, whether Keak Code contract requests need an open task budget, techniques locked off, and a prompt-cache rule. Requires `account:read`; any member may read it. Free: runs no inference.

Response WorkspaceAdminPolicyView

Successful response

Replace the workspace admin policy

Replace the whole policy of the calling key's workspace; the workspace is always the key's own. Send `expected_version` as read: a policy saved in between is 409 with the current one. Lowering `task_limit_max_usd` lowers every open task budget above it in the same transaction (`clamped_task_budget_count`); raising or clearing it changes no budget. Every save is audited. Requires `account:write` and a workspace owner or admin.

Request body WorkspaceAdminPolicyInput

Response WorkspaceAdminPolicySaved

Successful response

List the workspace's task budgets

Every member's task budgets in the calling key's workspace, newest first, each with its member and key. Requires `account:read` and a workspace owner or admin.

Query parameters

Response AdminTaskBudgetList

Successful response

Read one of the workspace's task budgets

One budget in the calling key's workspace. With `include=requests`, up to 500 of the requests it held, oldest first, and `requests_truncated`; that also needs the caller to see the member's requests in Team reports. Requires `account:read` and a workspace owner or admin.

Query parameters

Response AdminTaskBudget

Successful response

Close a member's task budget

Close a budget in the calling key's workspace now: `close_reason: admin` and `closed_by_user_id`, audited. Requests it already held settle into it later. Idempotent: a closed or expired budget is answered as it is. The member may open a new budget. Body `{}` or none. Requires `account:write` and a workspace owner or admin.

Query parameters

Response AdminTaskBudget

Successful response

List experiments

Every experiment on the account with its cost, savings and feedback.

Response ExperimentList

Successful response

Get an experiment

One experiment broken down per variant, so arms can be compared.

Query parameters

Response Experiment

Successful response

List optimization techniques

The catalog with this account's state on each entry. Preview and unimplemented techniques are included and flagged, not hidden.

Response TechniqueList

Successful response

Enable or disable a technique

Requires account:write. A technique that is not implemented, or is deprecated, is refused with a 404. Turning on a technique the workspace admin policy locks off is refused with 409 technique_locked_by_workspace; turning it off is always allowed.

Query parameters

Request body TechniqueUpdateRequest

Response TechniqueUpdateResponse

Successful response

List API keys

Prefixes only. Saved full keys can be copied by their owner in the dashboard.

Query parameters

Response ApiKeyList

Successful response

Create an API key

Requires account:write, and refuses a caller authenticated through a chat integration: a short-lived internal token must not be able to mint a permanent credential. The plaintext key is returned in the response. Its owner can also copy the saved key in the dashboard.

Request body ApiKeyCreateRequest

Response ApiKeyCreated

Successful response

Revoke an API key

Requires account:write. Revoking the key you are authenticating with is allowed. The next request simply fails to authenticate.

Query parameters

Response ApiKeyRevoked

Successful response

List connected chat channels

Every Slack, Discord, Telegram and email channel this account has connected, plus the event catalog PATCH accepts.

Response ChannelList

Successful response

Get a channel

One channel's subscriptions and digest schedule.

Query parameters

Response Channel

Successful response

Change a channel's schedule or subscriptions

Requires account:write. An events list containing an unknown id is refused whole rather than partially applied.

Query parameters

Request body ChannelUpdateRequest

Response Channel

Successful response

Post the savings digest now

Requires account:write. Throttled to one send per channel per minute, and does not consume the next scheduled digest.

Query parameters

Response ChannelReportResponse

Successful response

Record a refused request against the calling account

Reported by the MCP server when it refuses a call. `injection`, `escalation` and `policy` describe hostile arguments and trigger a permanent ban on the first occurrence, across every surface, not only the account API. `unauthorized` and `forbidden` describe a credential rather than an intent (a rotated key, a missing scope), so they are accepted and have no consequence. There is no rolling window, no warning tier and no escalation ladder. No scope is required. Returns the account's standing, and files a support ticket on the account's behalf so a block has somewhere to be contested.

Request body SecurityStrikeRequest

Response SecurityStanding

Successful response

Get the calling account's standing

Whether the account is blocked, and where to appeal. A block is a ban: it has no expiry, so `blocked_until` is always null and `permanent` is true whenever `tier` is `blocked`. The counting fields (`strikes`, `attempts`, `remaining`, `warn_after`, `block_after`, `window_hours`) are retained for compatibility with deployed MCP clients and now describe a one-strike policy rather than a tally. No scope is required. A banned account is refused at authentication, so this route answers for accounts that are not blocked, which lets a client notice a lift.

Response SecurityStanding

Successful response

Start a wallet top-up

Requires account:write. Creates a Stripe Checkout session and returns its link. Nothing is charged here. The customer completes the payment on Stripe's own page, with the amount in front of them.

Request body TopupRequest

Response TopupResponse

Successful response

Search the documentation

With `q`, the matching sections in full, best match first. Without it, the table of contents includes every section's id, title, category and link without bodies. A client can see what exists before choosing what to read. Needs a valid key and no scope: this is the same text served at /docs.

Query parameters

Response DocSectionList

Successful response

Attribute your requests to a task

Tag requests made with the calling key with a client task, an opaque project id and the client name. Requires the `inference` scope (describing your own traffic is part of sending it); only an API key authenticates here. All or nothing: every listed request must belong to the calling key and its actor, otherwise nothing is written. A tag never changes once written: a different task, project or client for an already tagged request is a 409, and repeating an identical post is a no-op that counts the ids as `unchanged`. The tag is read back as `attribution` on `GET /v1/usage/requests`.

Request body UsageAttributionRequest

Response UsageAttribution

Successful response

Report a completed task

Record that a task the calling key already attributed (`POST /v1/usage/attributions`) was completed. Requires the `inference` scope; only an API key authenticates here. `completed` is the only outcome recorded: a Keak Code task can be resumed after it failed or was cancelled, so those are not final. Idempotent: repeating it answers 200 with `created: false`. The workspace is copied from the task's tags and the member from the key. Team Savings reads it as "completed tasks, as reported by Keak Code".

Request body UsageTaskOutcomeRequest

Response UsageTaskOutcome

Successful response

Read the calling key's technique profile

What each Cheaper Inference technique will do for requests made with this key, derived from the same resolver that serves requests: the key's own override, the workspace setting, the effective state and where it comes from. Also the key's id, prefix, scopes, monthly spend limit and `prompt_cache_mode`. On a workspace key also the workspace admin policy (`admin_policy`): a technique it locks off reads `locked_by_workspace: true` and `effective: false`, and `prompt_cache_mode_effective` applies its cache rule. Requires the `inference` scope. Only `self` is accepted as `key_id` in this release.

Query parameters

Response KeyTechniqueProfile

Successful response

Narrow the calling key's technique profile

Move the key's own techniques toward conservative and set its prompt-cache default. `profile: "coding-safe-v1"` writes an explicit off override for every implemented technique; entries in `techniques` then apply on top of it. `off` is always allowed; `inherit` only while the workspace has that technique off; `on` never (403 `technique_change_requires_manager`, turned on by the key's manager in the dashboard). Unknown fields are refused. Requires the `inference` scope. Only `self` is accepted as `key_id`. Answers with the profile read back after the change. Under a workspace `prompt_cache_rule`, `prompt_cache_mode` is still stored as the key's default and the answer is 200: `prompt_cache_mode` shows it and `prompt_cache_mode_effective` the rule. `inherit` on a technique the workspace locks off is refused like any `inherit` while the workspace has it on, because it would turn on when the lock is lifted.

Query parameters

Request body KeyTechniqueProfileUpdate

Response KeyTechniqueProfile

Successful response

List the calling key's task budgets

The key's open task budget (null when none is open and unexpired) and its 20 newest budgets, newest first. A task budget is a spending cap Cheaper Inference enforces on this key while it is open: every inference request the key admits holds its worst-case charge against the budget before any wallet hold, and one that does not fit is refused with 400 `code: task_budget_exhausted`. Enforcement is per key, not per task: every request on the key counts, whichever client sent it, and only requests that reach Cheaper Inference are held. Also the workspace admin policy (`admin_policy`). Requires the `inference` scope; only `self` is accepted as `key_id`. Free: runs no inference.

Query parameters

Response TaskBudgetList

Successful response

Open a task budget on the calling key

Open the key's task budget for one client task, with a limit and a lease. Until it is closed or the lease expires, each request on the key holds its worst-case charge against the limit and settles to what it was billed; a request that does not fit is refused before anything is held. At expiry enforcement ends and the key falls back to its normal limits, so renew the lease with PATCH while the task runs. One open budget per key: repeating the same task and limit replays it (200, `replayed: true`); the same task with another limit is 409 `task_budget_exists` (raise it with PATCH instead); another task's open budget is 409 `task_budget_conflict` with that budget in `error.active`. On a workspace key, a limit above the workspace admin policy's maximum is clamped to it, never refused: `limit_usd` is the maximum and `clamped_to_max` is true. A replay compares the clamped limit.

Query parameters

Request body TaskBudgetOpenRequest

Response TaskBudget

Already open for this task with this limit (`replayed: true`).

Read one of the calling key's task budgets

The budget as it is now. `status` is the effective one: an open budget past its lease reads `expired`. `read_at` is the server's clock, for lease arithmetic. With `include=requests`, up to 500 of the requests it held, oldest first, and `requests_truncated`.

Query parameters

Response TaskBudget

Successful response

Raise a task budget or renew its lease

Raise the limit, renew the lease from now, or both. The limit asked for never goes down. Only the budget of the task named in the body changes, and only while it is open. On a workspace key a raise above the workspace admin policy's maximum lands at it (`clamped_to_max: true`), and a renewal alone of a budget above the maximum lowers it to the maximum.

Query parameters

Request body TaskBudgetUpdateRequest

Response TaskBudget

Successful response

Close a task budget

Close the budget now; requests on the key are no longer held by it. Requests it already held stay counted and settle into it later. Idempotent: a closed or expired budget is answered as it is. Closing another task's budget needs `force: true`.

Query parameters

Request body TaskBudgetCloseRequest

Response TaskBudget

Successful response

Generate a video

Generate a video and wait for the completed result. Streaming is not supported.

Request body VideoGenerationRequest

Additional properties are forwarded to the serving provider rather than rejected.

Response VideoResponse

Successful response

• x-ci-zdr — Whether zero data retention applied to this request. Absent when no policy was recorded.
• x-ci-discount-percent — Quoted customer rate discount versus list, not the final effective savings after usage and optimizations. See usage history for the settled effective discount.
• X-Cheaper-Inference-Request-Id — Request identifier. Legacy name, still emitted; identical to x-ci-request-id.
• x-ci-request-id — Request identifier. Canonical going forward.
• x-ci-tokens-saved — Input tokens removed by the optimizer.
• x-ci-saved-usd — Total saved against the provider's list price, in USD.
• x-ci-techniques — Comma-separated ids of the optimization techniques applied. Empty when none were.
• x-ci-cache — `hit` or `miss` for the exact-match cache.
• x-ci-prompt-cache-affinity — Sticky prompt-cache outcome (KEA-5164). One of `hit`, `moved`, `new`, `stale`. `hit` means the request was served on the route its conversation was warm on; `moved` means a warm route was remembered but another route served this turn, so its cache was not reused. Present only when the request carried a cacheable prefix and the affinity mode was not off; the exact-match `cache` header covers the separate response cache.
• x-ci-experiment-variant — Name of the assigned experiment variant. Present only when the request included X-CI-Experiment and matched a running experiment.
• x-ci-routing-overhead-ms — Elapsed milliseconds from proxy request start to dispatch of the successful provider attempt, including routing and prior attempts. On an exact-match cache hit, measures elapsed proxy processing time before returning the cached response. Present on successful buffered, streaming, and cached responses.
• x-ci-model-response-ms — Elapsed milliseconds from dispatch of the successful provider attempt to its complete response body. Present on successful non-streaming responses; absent on streaming responses and when no provider was called.
• x-ci-warning — Present when the request was served with a parameter value other than the one sent, because the model does not accept the value sent. Names the parameter, the value sent and the value used. Example: GPT-6 reasoning_effort "minimal" is served as "low". A request pinned to a provider is refused with 400 unsupported_parameter instead.
• x-ci-stage-timing — Optional diagnostic timing breakdown, emitted when PROXY_STAGE_TIMING=1 and a request timer is active. Comma-separated stage=milliseconds entries with total first; zero-duration stages are omitted. Stage names depend on the request path.

Upload a temporary image

Store an image for a short time and receive a `file_id`. Send that id in a chat message and it is swapped for a signed URL on the way to the provider. At most 4 MB per image; the whole request body may not exceed 4.5 MB. Rejected for an account with zero data retention.

Delete a temporary image

Remove an upload before it expires. An id belonging to another account answers 404, exactly as an unknown one does.

Query parameters

OpenAIError

Every `/v1/*` path except `/v1/messages*`.

AnthropicError

`/v1/messages` and `/v1/messages/count_tokens`. No `code`, no `param`.

ChatCompletionRequest

Additional properties are forwarded rather than rejected.

ChatCompletionResponse

Additional properties are forwarded rather than rejected.

CompletionRequest

Additional properties are forwarded rather than rejected.

CompletionResponse

Additional properties are forwarded rather than rejected.

VideoGenerationRequest

Additional properties are forwarded rather than rejected.

VideoResponse

Additional properties are forwarded rather than rejected.

DecisionRequest

Additional properties are forwarded rather than rejected.

DecisionPredicateQuestion

A yes or no question, answered with the probability of yes.

Additional properties are forwarded rather than rejected.

DecisionChoiceQuestion

A pick from the listed choices, answered with the chosen value and a probability per choice.

Additional properties are forwarded rather than rejected.

DecisionScoreQuestion

A rating on the listed levels, answered with a probability-weighted level index and a probability per level.

Additional properties are forwarded rather than rejected.

DecisionResponse

Additional properties are forwarded rather than rejected.

DecisionPredicateAnswer

Additional properties are forwarded rather than rejected.

DecisionChoiceAnswer

Additional properties are forwarded rather than rejected.

DecisionScoreAnswer

Additional properties are forwarded rather than rejected.

DecisionRefusalAnswer

The model declined to answer this question. Other questions in the same request can still receive answers.

Additional properties are forwarded rather than rejected.

DecisionUsage

Billed on input tokens; a decision has no cache or reasoning tokens, so the detail counters are zero.

Additional properties are forwarded rather than rejected.

ImageGenerationRequest

Additional properties are forwarded rather than rejected.

ImageEditJsonRequest

CheaperInference JSON extension. URL references are never fetched or stored by CheaperInference and are eligible only on provider routes that accept URLs.

ImageEditMultipartRequest

ImageResponse

Additional properties are forwarded rather than rejected.

AnthropicMessagesRequest

Additional properties are forwarded rather than rejected.

AnthropicMessagesResponse

CountTokensRequest

Additional properties are forwarded rather than rejected.

CountTokensResponse

ResponsesRequest

Additional properties are forwarded rather than rejected.

ResponsesFunctionTool

Additional properties are forwarded rather than rejected.

ResponsesCustomTool

A freeform/custom tool. Providers without native custom tools receive a function wrapper whose sole input is the raw custom-tool string.

Additional properties are forwarded rather than rejected.

ResponsesNamespaceTool

Responses Lite namespace. Function and custom entries may be mixed in one namespace.

Additional properties are forwarded rather than rejected.

ResponsesHostedTool

An optional hosted tool. Offering one is tolerated, but selecting it with tool_choice is rejected because CheaperInference cannot execute it.

Additional properties are forwarded rather than rejected.

ResponsesTool

A passthrough object: the API does not constrain its shape.

ResponsesAdditionalToolsItem

Additional properties are forwarded rather than rejected.

ResponsesInputItem

A message, reasoning item, replayed function/custom call or output, agent message, or Responses Lite additional-tools declaration.

A passthrough object: the API does not constrain its shape.

ResponsesResponse

Additional properties are forwarded rather than rejected.

ModelList

Model

PublicModelCatalog

PublicModel

DetailError

The error body the /public routes answer with.

UsageRequestList

UsageRequest

UsageRequestBilling

The request's billing state for a client reconciling its own ledger. Derived from `status`: settled carries the settled charge; an incomplete request whose delivered partial stream was billed is also settled, with that charge, matching the stream receipt; cached, and failed with no charge, say `0.000000`; validating and forwarding (in flight, or a settlement that has not landed) are `pending` with a null amount, because the amount is not known.

UsageRequestTechnique

UsageRequestAttribution

UsageAttributionRequest

Strict: any other field is refused with 400.

UsageAttribution

UsageAttributionError

The OpenAI error envelope with the ids the caller must act on inside `error`, never at the top level.

UsageTaskOutcomeRequest

Strict: any other field is refused with 400.

UsageTaskOutcome

TeamSavingsSummary

Amounts are USD strings with 6 decimal places, rounded half-even, and may be negative. Rates are fractions with 6 decimal places, or null when there is nothing to divide by.

TeamSavingsGroup

A summary for one member, project or model.

TeamSavings

TeamSavingsRequest

A request record. No prompt, response or request body.

TeamSavingsRequestList

KeyTechniqueProfile

KeyTechniqueProfileEntry

KeyTechniqueProfileUpdate

Strict and non-empty: send at least one field; any other field is refused.

TaskBudget

A task budget on one API key. Money is USD as strings with exactly 6 decimal places. `limit_usd - reserved_usd - settled_usd` is what is left (`remaining_usd`, never below 0). A held request counts its worst-case charge in `reserved_usd` until it settles, then its billed amount in `settled_usd`; a request that was not billed (a failure, or no usage reported) settles at 0. On a workspace key the workspace admin policy's maximum caps `limit_usd`: an open, raise or renewal above it lands at it, and lowering the maximum lowers every open budget above it (`limit_clamped_at`).

TaskBudgetHold

One request held by the budget. Every request on the key counts, whichever client sent it; compare the ids with your own to find requests you did not send.

TaskBudgetList