Skip to main content
AddedChangedFixedAPIAudioVideo
User provisioning from Okta (SCIM); choose whether removing a member suspends their keys; your request logs in your own Grafana; video routes per input type and a dedicated edits endpoint; web search follows the key's tool policy; `weight_profile` on router select; long-context 1-hour cache writes priced on their own; WebSocket keys out of the URL; top-level speech `speed`; MiniMax `vol` and `emotion`; 12-hour prompt-cache writes priced on their own; agent run tracing with `X-Mesh-Run-Id`; a key can describe its own org, scopes, model access and limits

User provisioning from Okta

Added. Okta can now add and remove the members of your Mesh organization over SCIM 2.0. In Settings → Single sign-on, click Set up provisioning under User provisioning (SCIM), then paste the base URL and token into Okta. Assign someone in Okta and they join as a Member; unassign them and they leave, with their API keys suspended. Push an Okta group, pick it under Group roles on the same card, and set it to Admin to make its members admins. Someone who already has a Mesh account is linked, not duplicated. The owner is never removed, and removing someone in Mesh sticks. Enterprise, alongside SSO. See User Provisioning.

Choose whether removing a member suspends their keys

Added. In Settings → People, click Remove on a member: the dialog now lists the API keys they created and suspends them by default. Untick Suspend these keys to keep them working under your organization’s limits. Set the default with the Removing a member → Suspend their API keys switch on the same page. Admin keys can do the same through the Admin API. See Removing members.

A key can describe itself

Added. GET /v1/keys/{id} reports what an API key is and may do — its org, status, scopes, model access and resolved limits — and never returns the key secret or its hash. Unlike the other key-management routes, an rsk_ key may call it for itself: the {id} must be the key’s own id, and any other value is 404, so it is no way to probe for other keys. A mak_ admin key holding keys:read can describe any key in its reach. A suspended key, or one blocked by its IP allow-list, cannot reach it. See API Keys.

Your request logs in your own Grafana

Added. Mesh serves a Loki-compatible read API over your organization’s request logs. In Grafana, add a Loki data source with the URL https://api.meshapi.ai/v1/logs and an Authorization: Bearer mak_... header, then query the logs in Explore and Logs Drilldown. The key must be an admin key holding logs:read at org scope; an rsk_ key gets 401. Every query reads only the calling key’s organization.Each authenticated HTTP request has one http_request record, kept 7 days, with OpenTelemetry field names: http_request_method, http_route, url_path, http_response_status_code, client_address, user_agent_original, request_id, key_id, org_id, owner, and trace_id / span_id. latency_ms is the time to the response headers (time to first byte on a stream) and total_duration_ms the time until the body finished. An unusual method is recorded as _OTHER, with the raw value in http_request_method_original. The labels are service_name and deployment_environment_name. Request and response bodies are not in these records, and your own calls to /v1/logs are not recorded.The API serves the read paths Grafana uses under /v1/logs/loki/api/v1/: query, query_range, labels, label/{name}/values, series, detected_labels, detected_fields, detected_field/{name}/values, patterns, format_query, index/stats, index/volume, index/volume_range and status/buildinfo. Live tail, push, OTLP, delete and rules paths return 404.
  • 600 requests a minute per organization, across all your keys and dashboards. Past it the answer is 429 with Retry-After.
  • 32 MiB per answer. A larger one is refused with 422 log_result_too_large, never cut short; narrow the time range or the query.
  • Nothing before 2026-11-12 11:00 UTC. A time range that starts earlier is moved forward. An instant query that reaches before that time, a query using offset, or a time or query parameter given twice returns 400.
Added. GET /v1/logs/bodies returns the request and response bodies your body-log policy exports, newest first. Filter with request_id, key_id, since and until (at most 7 days apart; the last hour by default) and limit (1 to 50, default 20); it takes these typed filters, not LogQL. Page with the returned next_cursor, sending the same window and filters each time. A page can hold fewer records than limit, or none, while next_cursor is still set; the last page has no next_cursor.
  • Every read is written to your audit log as log_bodies.read before anything is fetched; if the audit store is down the answer is 503.
  • With zero data retention on, the answer is 403, even for bodies stored before it was turned on. If the setting cannot be confirmed, 503 zdr_state_unavailable.
  • The body-log policy is read again on every read: turn a key’s policy off, or narrow it to one direction, and its bodies stop being served, including ones already stored.
Added. Two org permissions. logs:read covers the Loki paths and is held by the owner, admin and viewer roles. logs:read_bodies covers GET /v1/logs/bodies, is held by owners and admins only, cannot be carried by a custom role, and is never implied by logs:read. Both can be granted on an admin key.

Video: a route per input type, and a dedicated edits endpoint

Added. POST /v1/video/t2v, POST /v1/video/i2v and POST /v1/video/v2v take the same body as POST /v1/video/generations; the path names the modality the serving model must support. A model without it, or any fallback target without it, returns 422 before any provider call. Jobs stay asynchronous: submit, then poll the task. GET /v1/models and GET /v1/models/{id}/providers now return supports_t2v, supports_i2v, supports_v2v and video_options on each model and provider, so you can pick a model for a route before you submit.Added. POST /v1/video/edits takes video edits.Changed (API version 2026-11-12). On requests pinned to 2026-11-12, POST /v1/video/generations takes a text prompt and reference images (image_url items with role: "reference_image"); frames, video, audio, previous_task_id and other edit inputs go to POST /v1/video/edits instead and are a 422 on generations. A resolution, ratio or duration the model does not support is refused. Requests on earlier versions, including unpinned keys, keep the combined body. See API versioning.

Web search follows the key’s tool policy

Changed. POST /v1/web/search now checks the key’s tool_policy, as a model request’s tools already did. A key whose policy blocks web_search, or whose allowed list does not name it, gets 403 tool_not_allowed with details.tool: "web_search". A video submit with web_search: true is refused the same way. Keys with no tool policy are unaffected; to keep access on a key with an allowed list, add web_search to it.

weight_profile on /v1/router/select

Added. POST /v1/router/select accepts weight_profile in the body (or the X-Mesh-Weight-Profile header; the body wins) and resolves the profile the same way POST /v1/chat/completions does: the request, then the key, then its team or organization template, then the default. It used to use the global default every time. An unrecognised name falls back to balanced instead of failing the request. See Routing Algorithms for the precedence, and Auto-routing for which requests a profile can affect.

Long-context 1-hour cache writes are priced on their own

Added. GET /v1/models returns long_context_cache_write_1h_input_usd_per_1m in each model’s pricing: the price per 1M tokens of a 1-hour cache write on a prompt over the model’s long-context threshold. It is null on models without a separate rate.Changed. Those writes now bill at that rate where a model has one (today, models served directly by Anthropic) instead of the long-context 5-minute rate. Models where the field is null bill as before.Releases v1.0.377 to v1.0.412.

WebSocket keys out of the URL

Fixed. The realtime, live and audio WebSockets take your key as the subprotocol openai-insecure-api-key.<rsk_...> (the form OpenAI’s browser client sends) or as an Authorization: Bearer rsk_... header. The Bearer rsk_... subprotocol these docs showed before never connected: a subprotocol can’t contain a space, so the handshake failed with a 400. ?api_key= keeps working unchanged, though a URL ends up in logs, so prefer the two forms above. If a request carries more than one, the subprotocol wins, then the header, then ?api_key=. See Realtime audio authentication.Changed. Mesh echoes back the first subprotocol you offer that is not the key, and none when you offer none. /v1/realtime and /v1/live used to answer openai-realtime every time, which a client that offered realtime (OpenAI’s browser client) or nothing rejected.

Speech speed control

Added. You can set a top-level speed (0.25 to 4.0) on POST /v1/audio/speech; it used to be dropped without an error. Mesh forwards speed as sent and never clamps it, so each provider applies its own range; a value outside 0.25 to 4.0 is a 422. It takes effect on models served by DeepInfra, SiliconFlow, Venice, and MiniMax on Novita; other models accept the value and ignore it. voice_settings.speed still works unchanged and is the field ElevenLabs reads.

MiniMax speech: volume and emotion

Added. POST /v1/audio/speech accepts an optional vol field (greater than 0, up to 10) and an optional emotion field on MiniMax speech models; other speech providers ignore emotion. MiniMax speech requests now pass pitch, vol and emotion through to the provider instead of dropping them.Changed. A MiniMax speech request returns a 422 before reaching the provider, with nothing billed, when pitch is not a whole number from -12 to 12, or when emotion is fluent or whisper on a model other than speech-2.6.

12-hour prompt-cache writes are priced on their own

Added. usage.prompt_tokens_details.cache_write_12h_tokens reports the prompt tokens written to the 12-hour prompt cache. GET /v1/models returns cache_write_12h_input_usd_per_1m and long_context_cache_write_12h_input_usd_per_1m, and their flex counterparts flex_cache_write_12h_input_usd_per_1m and flex_long_context_cache_write_12h_input_usd_per_1m, in each model’s pricing object.Changed. 12-hour cache writes now bill at their own rate on standard, long-context and flex pricing instead of the 30-minute write rate. Long-context pricing on the affected OpenAI models now applies above 272,000 input tokens, matching OpenAI’s published boundary.Release v1.0.384.

Trace an agent run

Added. Send the same X-Mesh-Run-Id header on every request an agent run makes, and GET /v1/usage/runs/{run_id} reports the run as one unit: every hop oldest first, retried attempts included and flagged, and one total_charged_usd for the whole run. It is a header, so setting it once on your client’s default headers carries it across every request in the run.Added. run_id is reported on each usage event and is a filter on POST /v1/usage/events and POST /v1/usage/rejections.An unusable run id — longer than 128 characters, containing /, or shaped like an API key — is dropped, and the request is served exactly as if no header had been sent. No request is ever refused over a run id. The accepted id is echoed back on the response as X-Mesh-Run-Id, so read that header to confirm what was recorded.Run Tracing is the full reference.

API

  • You can now set layer_decomposition on an image generation request to split an input image into a base image plus up to 99 PNG layers.
  • You can now control decomposition with layer_decomposition_options, which accepts max_images and text_recognition.
  • ImageOptions now accepts the same layer_decomposition and layer_decomposition_options fields.
  • Each item in data[] now includes provider-reported metadata when available: size, output_format, background, z_index, bounding_box, name, description and text_recognition.
  • Layer decomposition works only on models that support it, and a decomposition request sent to any other model returns a 400.
  • An explicit layer_decomposition: false is accepted on any model and has no effect.
  • The balance hold for a decomposition request is sized from layer_decomposition_options.max_images, so set it to avoid reserving balance for up to 100 images.
  • Every returned image, including each layer, is billed at the tier of the requested size.

API

  • You can now read a key’s details with GET /v1/keys/{key_id}, including its org, scopes, allowed_models, blocked_models, model_limits and effective limits.
  • The response never includes the key’s secret or hash.
  • An rsk_ key can read only itself, and any other key_id returns 404.
  • A mak_ admin key or a user session with the keys:read permission can read any key within its scope, and a key outside that scope returns 403.
  • Effective limits use the same minimum-wins calculation as GET /v1/keys/{key_id}/limits.