Skip to content

Changelog

Product updates, model releases, and platform improvements.

November 2026

Organizations, end to end

You can now create team organizations, delete one with a preview and a 7-day window to cancel, review audit events, and read usage and cost with an admin key.

If you run a team on Apertis, the jobs around it now have a place: a script reads usage and cost with an admin key, a mistaken delete can be cancelled for seven days, and every administrative change is in an audit log you can export. An account can create up to 5 team organizations.

Read usage and cost over the API

Two read-only endpoints return what your organization used and what it was charged. They keep the paths, the bucket envelope and the result objects of OpenAI's organization usage and costs API, so a script written against that API needs a new base URL and a new key, not a rewrite.

curl -G "https://api.apertis.ai/v1/organization/costs" \
  -H "Authorization: Bearer sk-admin_..." \
  -d start_time=1790035200 \
  -d group_by=project_id,api_key_id \
  -d limit=7
{
  "object": "page",
  "data": [
    {
      "object": "bucket",
      "start_time": 1790035200,
      "end_time": 1790121600,
      "results": [
        {
          "object": "organization.costs.result",
          "amount": { "value": 1.234567, "currency": "usd" },
          "project_id": "proj_12",
          "api_key_id": "key_34",
          "api_key_name": "ci",
          "charge_basis": "configured_price"
        }
      ]
    }
  ],
  "has_more": true,
  "next_page": "page_eyJ2Ijox..."
}
UsageCosts
EndpointGET /v1/organization/usage/completionsGET /v1/organization/costs
Returnsinput, output, cached and cache-write tokens, request countamount charged in USD
group_byproject_id, api_key_id, model, user_idproject_id, api_key_id, line_item
Filtersproject_ids, api_key_ids, models, user_idsproject_ids, api_key_ids
BucketsUTC days, 1 to 31 per pageUTC days, 1 to 31 per page
Rangeup to 366 days per requestup to 366 days per request
Formatsjson, csvjson, csv

Before you build on it:

  • Amounts are exact. amount.value is USD to six decimals with no float rounding, and it is the same figure the console's Usage tab sums for the same range and filters.
  • Every result says how it was priced. charge_basis is configured_price, unknown_model_fallback or not_recorded, always as separate results. A charge is never reported as configured_price unless it was one.
  • Pagination does not drift. Omit end_time and the first page fixes it; the next_page cursor carries it, so page 40 covers the same range as page 1. A cursor sent with other parameters is refused with 400 invalid_page. Results are never dropped to fit a page.
  • The end is exclusive. A row recorded at exactly end_time is not counted. The console treats its end as inclusive, so compare against end_time - 1.
  • CSV for spreadsheets. format=csv returns one row per result; the cursor is in X-Next-Page and X-Has-More. Names that would run as a formula are neutralised.
  • Rate limit: 60 requests per 60 seconds, counted against the member who created the key and shared with that member's own console queries. Create polling keys from a member who does not live in the console.

Two differences from OpenAI's API: bucket_width accepts only 1d, and limit is 1 to 31.

If a call fails, this is what you will see:

StatusCodeWhen
400invalid_request_errora bad time, a range over 366 days, an unknown group_by, limit outside 1 to 31
400range_not_availablestart_time is before the earliest usage still held; the message names that time
400invalid_pagethe cursor belongs to another organization or other parameters
401invalid_api_keya missing, unknown or revoked admin key, or an inference key
403insufficient_permissionsthe key does not allow usage.read

Create admin keys with scopes

To call those endpoints, create an admin key (sk-admin_...). It is a separate credential from an inference key: sk-... and sk-org_... keys are refused on the organization endpoints with 401. An admin key always acts inside its own organization; no parameter or cursor can name another one.

The Create Admin key dialog with Restricted chosen: Usage set to Read and Organization administration set to Write

Pick the mode that fits the job:

ModeWhat the key can do
read_only (default)read every family below
restrictedonly the families you pick, each as read or write: usage, projects, project_users, service_accounts, project_api_keys, admin_api_keys
allread and write every family
  • In the console the three modes are All, Restricted and Read only, and a new key starts as Read only. A restricted key there has two rows: Usage, and Organization administration, which covers the other five families together. The API sets each family on its own.
  • The secret is shown once, at creation. Apertis stores a hash and a masked form.
  • An admin key has no expiry. Revoke it when the automation no longer needs it.
  • A key never outlives its creator's rights. If the member who made it is disabled, leaves the organization or loses the right to manage admin keys, the key answers 403 (owner_disabled, owner_not_org_member, owner_not_authorized).

Delete an organization with a preview, and seven days to change your mind

Deleting an organization used to be a single irreversible request. Now, as an owner, you open Organization settings, choose Delete organization, and see what the deletion takes with it before you confirm: other members, pending invitations, projects, service accounts, member keys, service-account keys, admin keys and available credits, each as a count, plus anything that blocks it.

The Delete organization dialog: the date the deletion takes effect, then counts of other members, pending invitations, projects, service accounts, API keys, admin keys and available credits

Confirming schedules the deletion seven days out. During the window:

Behaviour
Reading the consoleworks
Cancelling, removing a member, leavingworks
Any other change409 org_deletion_scheduled
Organization keys, service-account keys, admin keys403 organization_deletion_scheduled

You can cancel at any point in the window and everything resumes. An organization with money in motion cannot be scheduled: a positive or negative balance, an open checkout, a card being set up, or a payment that can still be reversed. The preview names each blocker.

Review who did what, and export it

Every administrative change is recorded with the actor, the action, the result (success, blocked or failure), the target, and a field-level diff (from and to) for updates. Since 2026-10-05 each event also carries the IP address and user agent, and the actor type: a signed-in member, an admin key, a SCIM token or the system.

The Audit tab: each event with its time, actor, IP address, action, target and result, including one blocked request with its error code, under the search box and the time, category and result filters

These are the events you will find:

FamilyEvents
Organizationorg.create, org.update, org.deletion.schedule, org.deletion.cancel, org.deletion.complete
Members and invitationsorg.invite.create, accept, resend, revoke; org.member.update, remove
Keysorg.api_key.* and org.admin_key.*: create, update, revoke
Roles and groupsorg.role.*, org.group.*
Billingorg.billing.*: payment methods, portal sessions, auto-recharge, PO reference
SSOorg.sso.connection.create, org.sso.domain.add, verify, remove
Projectsproject.create, update, archive, members and service accounts

You can filter by actor, action, result, target, project and time. Export as CSV or JSON, up to 50,000 events per export; a larger range is refused with 422 audit_export_too_large instead of being cut short. Each export returns a snapshot id in X-Audit-Export-Snapshot-Id and is itself recorded as org.audit.export.

Decide who can do what

  • Owner has full control. Reader can view the organization and create their own sk-org_ key.
  • Custom roles combine eight permissions: organization.read, projects.read, projects.write, members.write, billing.write, api_keys.write, project_api_keys.write, admin_api_keys.write.
  • A team organization holds up to 50 pending invitations; resending one extends it by seven days.

Docs: Organizations

Read update

Gemini 4 Hydrogen

Gemini 4 Hydrogen

Gemini 4 Hydrogen is Google's high-efficiency multimodal model in the Gemini 4 family, optimized for high-throughput agentic workflows, coding, and multi-model routing. It delivers strong performance at lower cost and latency than Gemini 4 Carbon, supporting text, image, audio, video, and file inputs for versatile multimodal applications.

Designed for scalable production workloads and long-horizon execution, Gemini 4 Hydrogen supports up to 1M output tokens through continuation and configurable reasoning effort levels (low, medium, high, and xhigh) to balance reasoning depth, speed, and cost. It is well suited for coding assistants, autonomous agents, complex workflow automation, and large-scale AI systems.

Read update

Add Gemini 4 Carbon

Gemini 4 Carbon

Gemini 4 Carbon is Google's frontier multimodal reasoning model, designed for advanced coding, agentic workflows, and complex multi-step problem solving. It supports text, image, audio, video, and file inputs, enabling integrated reasoning and analysis across diverse modalities.

Optimized for long-horizon execution and demanding professional workloads, Gemini 4 Carbon supports up to 1M output tokens through continuation and configurable reasoning effort levels (low, medium, high) to balance reasoning depth, latency, and cost. It is well suited for autonomous agents, software engineering, multimodal analysis, and large-scale workflow automation.

Read update

Account MCP, in your agent

Point an MCP client at `npx @apertis/mcp-server`, set `APERTIS_API_KEY`, and list models, check quota, or hand bulk grunt work to a lower-cost model of your choice — without leaving the session.

An MCP server for your Apertis account

Your agent can manage your Apertis account over MCP.

Add it with npx @apertis/mcp-server and set APERTIS_API_KEY. The agent can list models, check quota, and hand bulk grunt work to a lower-cost model of your choice, without leaving the session.

This is the gateway MCP, not the docs MCP.

Actions the agent takes run under your API key and count as activity on your account. Keep APERTIS_API_KEY private.

Docs: https://docs.apertis.ai/api/sdks/mcp-server/

Read update

Docs entry points for crawlers and agents

Starting today, the docs site has machine-readable entry points: robots.txt names the sitemap, pages carry canonical and Open Graph tags, and llms.txt links an OpenAPI file for the core endpoints.

A crawler or an agent no longer has to scrape the docs to find the API contract: two requests take it from the docs index to the OpenAPI file.

To start, run the two requests:

curl -s https://docs.apertis.ai/llms.txt | tail -n 3
curl -s https://apertis.ai/docs/openapi.yaml | head -n 4
Terminal output of the last lines of llms.txt, a Reference section that links the OpenAPI specification, followed by the first four lines of that YAML file

Find every entry point

Everything below is served from docs.apertis.ai unless noted.

URLContent typeHolds
/robots.txttext/plainAn allow-all rule and the sitemap location
/sitemap.xmlapplication/xmlThe canonical URL of every page, most with lastmod
/llms.txttext/plainAn index of the pages as Markdown links, plus the OpenAPI link
/llms-full.txttext/plainEvery page's Markdown in one file
/<page>.mdtext/markdownOne page as clean Markdown
/mcpJSON-RPC over POSTThe Docs MCP server (search_docs, get_page)
https://apertis.ai/docs/openapi.yamlapplication/yamlAn OpenAPI 3.0.3 file, on the main site host

Know what the OpenAPI file covers

The OpenAPI file declares the server https://api.apertis.ai/v1, API-key auth, and three operations: POST /chat/completions, GET /models and POST /embeddings. It does not describe the Responses, Messages, image, audio or billing endpoints; use the Markdown pages in llms.txt for those.

llms.txt follows the llmstxt.org layout: an H1, a one-paragraph summary, then H2 sections of - [Title](https://docs.apertis.ai/<path>.md): description entries. Each link is the page's clean Markdown. llms-full.txt holds the same pages in one file, with a Source: <canonical URL> line before each.

Crawl with robots.txt and the sitemap

User-agent: *
Allow: /

Sitemap: https://docs.apertis.ai/sitemap.xml

Every <loc> in the sitemap is the served form with a trailing slash, for example https://docs.apertis.ai/getting-started/quick-start/. The slash-less URL answers 308 to it. lastmod is the date of the last commit that touched the page's source, as ISO 8601 with offset (2026-10-06T14:59:11+08:00). The home page and /search/ carry no lastmod.

Read the canonical and Open Graph tags

Terminal output listing the canonical link and the og:url, og:title, og:type, og:site_name and twitter tags of the Quick Start page
TagValue on /getting-started/quick-start/
link rel="canonical"https://docs.apertis.ai/getting-started/quick-start/
og:urlthe same as canonical
og:titleQuick Start | Apertis Documentation
og:typearticle (website on non-article pages)
og:site_nameApertis Documentation
twitter:cardsummary

og:description is emitted only when the page defines a description; Quick Start has none. Documentation pages have no og:image. Each page also carries <meta name="apertis-docs:markdown" content="/getting-started/quick-start.md">, which points at its Markdown twin, and JSON-LD of type TechArticle and BreadcrumbList.

Choose between the files and the Docs MCP server

The MCP server at https://docs.apertis.ai/mcp accepts POST only (GET returns 405). It has two tools: search_docs (query, optional limit from 1 to 10, default 5) returns ranked pages with title, URL and snippet, and get_page (url) returns one page as Markdown. It reads the same llms.txt and llms-full.txt, so it serves the same pages, and it does not read the OpenAPI link. Use the files for crawling and bulk ingestion, and the MCP server for search at query time.

Read update