{
  "fixes": {
    "slug": "cartesia-ink-stt",
    "name": "Cartesia Ink",
    "listing": "https://www.anchorterminal.com/tools/cartesia-ink-stt",
    "markdown": "# Fix list: Cartesia Ink\n\nFrom Anchor Terminal's listing at https://www.anchorterminal.com/tools/cartesia-ink-stt, the October 2026 research run, assessed 9 October 2026. Grade B, 67.9 out of 100.\n\nThis is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.\n\nFor a coding agent working on Cartesia Ink: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.\n\n## 1. Security \u0026 auth, 52 out of 100, up to 8.4 more on the total\n\nWhy it scored 52: Model reading. Standard keys (`sk_car_...`) and admin keys are separate types that are not interchangeable, and a server can mint an access token with an `stt` grant that lasts at most one hour. That token travels in the `access_token` query parameter on WebSockets, the API key itself goes in a header, and no key limited to speech-to-text was found (26, less 5 for the token in the URL, 21 of 30). The terms, last revised 14 June 2024, say inputs, outputs and interactions may be used to train models unless otherwise agreed, with an opt-out form in the Playground's data controls (5 of 20). Zero data retention covers STT audio and transcripts and is limited to the Enterprise plan. No retention period for other plans was found, and the DPA gives 180 days for deletion after termination (5 of 15). Credit usage and key metadata are readable through admin endpoints. No audit log was found (7 of 15). security.txt is valid until 3 October 2027 with a contact and a policy link to the trust centre. The docs and pricing page state SOC 2 Type II, HIPAA and PCI-DSS. The trust centre is drawn by script and was not read, and no bug bounty was found (14 of 20).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-security):\n\n- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.\n- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.\n- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.\n- 0 to 15, audit logs or per-call visibility for the operator.\n- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.\n\nModels are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.\n\n## 2. Payments \u0026 pricing, 35 out of 100, up to 8.1 more on the total\n\nWhy it scored 35: No x402, MPP or L402 (0). The docs pricing page gives 3 credits a second for `ink-2` and the public pricing page gives plan prices and overage rates of $65, $45 and $38 per million credits, all read without a login (20). The Free plan is $0 with 20,000 credits a month, about 1 hour 51 minutes of `ink-2` audio. Neither page says whether sign-up asks for a card, and we did not sign up, so 15 of 20. A person creates the key in the Playground in a browser, and the hosted MCP server also needs a browser sign-in (0).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):\n\nThe published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).\n\n- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.\n- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for \"contact sales\" or prices behind a login.\n- 20, a free tier or trial that doesn't need a card.\n- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).\n\nPayment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.\n\nOpen-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.\n\n## 3. Reliability, 73 out of 100, up to 5.4 more on the total\n\nWhy it scored 73: Hosted reading. The status page at status.cartesia.ai (incident.io) lists Speech to Text as its own component in four regions (20). For July to October 2026 it showed 99.97% for the US, 99.994% for the EU and APAC and 100% for Australia. The history from 17 July to 8 October names no speech-to-text incident and lists five API-wide ones (1 August, 2 and 3 September, 7 and 8 October) whose durations were not shown to us, read as minor incidents only (20). Concurrency limits are published per plan, 8, 12, 20 and 60 concurrent STT requests, with a 3-minute idle timeout on sockets (15). Exceeding the limit returns 429 with the `concurrency_limited` code, and the SDKs retry 429 and 5xx twice with backoff on HTTP calls. No `Retry-After` header, backoff guidance or reconnection advice for a dropped socket was found (8 of 15). No SLA was found for any plan (0). `ink-2` is marked Stable, with `ink-preview` as the preview line (10).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):\n\nHosted APIs, MCP servers, models and platforms.\n\n- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).\n- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.\n- 15, rate limits documented with numbers.\n- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.\n- 10, an SLA published for any paid tier.\n- 10, the surface agents use is generally available, not beta or preview.\n\nLocal packages, SDKs, frameworks and stdio MCP servers.\n\n- 20, installs from an official package with supported runtimes stated.\n- 25, a public CI and test suite, passing on the default branch.\n- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).\n- 15, semver discipline and breaking changes called out in a changelog.\n- 15, version 1.0 or later, or declared stable.\n\nProtocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.\n\n## 4. Agent ergonomics, 74 out of 100, up to 4.2 more on the total\n\nWhy it scored 74: API reading, as for the other speech-to-text listings. Turn events carry a cumulative transcript that is never revised, so an agent reads one `turn.end` a turn, and word timestamps are opt-in on the manual and batch endpoints (20 of 25). There is nothing to page through. Batch takes a file of any length without client chunking, keyterms steer vocabulary, and there is no diarisation or subtitle output (12 of 20). Errors are structured JSON with `error_code`, `title`, `message`, `request_id` and an optional `doc_url` on API version 2026-03-01 and later (17 of 20). Transcription holds no state to duplicate and the SDKs retry failed HTTP calls. A realtime session cannot be resumed, and nothing on reconnecting was found (12 of 20). Realtime needs `model`, `encoding`, `sample_rate` and the version, turn settings have documented defaults, and there are official Python and TypeScript SDKs. Realtime accepts raw PCM only (13 of 15).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):\n\n- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).\n- 20, pagination, filtering and output-size controls.\n- 20, actionable, documented error responses, codes and messages an agent can recover from.\n- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.\n- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.\n\nModels are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.\n\n## 5. Transparency \u0026 trust, 67 out of 100, up to 2.9 more on the total\n\nMade of editorial 49, provenance 84.\n\nWhy it scored 67: The service is closed and the SDKs are Apache-2.0. The Terms of Service cover the website, its subdomains and APIs, were last revised on 14 June 2024, describe a platform for synthesised sound and never mention transcription (12 of 30). The terms and privacy policy agree that content may be used for training with an opt-out. The DPA says Cartesia will not retain or use customer personal data beyond performing the services, which sits uneasily with the training licence. No retention period is given outside zero data retention, and the privacy policy says the services are designed for users in the United States only while regional endpoints are sold in four other regions (12 of 30). A deprecations page with dated sunsets and a stated promise to preserve fields for a pinned `Cartesia-Version` (15 of 20). The DPA promises 30 days' notice of sub-processor changes to subscribers and points to a list in the trust centre, which we could not read. Regional deployments are named for Enterprise only (10 of 20).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):\n\n- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.\n- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).\n- 0 to 20, a deprecation policy or notices with dates.\n- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).\n\nThe other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.\n\nProvenance checks not met in full (half of this category, computed from checked facts):\n\n- Domain age: cartesia.ai, registered 2023-05-10 (3 years) (7 of 15)\n- Terms of service: read, states 6 of the 7 things a reader expects, and has 2 clauses that cost points (5.1 of 10)\n- Privacy policy: read, states 4 of the 8 things a reader expects (7 of 10)\n\n## 6. Schema \u0026 documentation, 89 out of 100, up to 1.8 more on the total\n\nWhy it scored 89: Model reading. The docs index links an OpenAPI file (`latest.yml`) and AsyncAPI files for the WebSocket endpoints, and each reference page embeds its part of them (25). llms.txt with a Markdown twin of every page (10). The comparison page says which of the three endpoints to use and when, and the turn detection guide explains each event and threshold (18 of 20). The AsyncAPI definition has enums for `encoding` and `cartesia_version`, minimum and maximum values on the four turn settings and required fields marked. `model` is a plain string on the auto endpoint (13 of 15). Example scripts for each endpoint in Python and TypeScript and an errors page with nine codes. `model_sunsetted` appears on the deprecations page and not in that table, and the docs say a wrong `encoding` or `sample_rate` might not return an error (11 of 15). Dated API versions through the `Cartesia-Version` header and a changelog grouped by month with no day dates (12 of 15).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):\n\nAPIs and MCP servers.\n\n- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).\n- 10, llms.txt or Markdown docs served for agents.\n- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.\n- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.\n- 0 to 15, examples and documented error responses.\n- 15, versioning and a public changelog.\n\nModels are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.\n\n## 7. Maintenance \u0026 community, 85 out of 100, up to 1.3 more on the total\n\nWhy it scored 85: Model reading, as for the other hosted speech-to-text listings. The model page dates the current `ink-2` and `ink-preview` 17 September 2026, 22 days before the check (30). The deprecations page gives dated sunsets, and the July 2026 changelog announced three text-to-speech sunsets for 20 October 2026. No minimum notice period is stated and no Ink model has been retired (8 of 12). No speech-to-text model was retired or renamed in the last 90 days (8 of 8). The changelog has speech-to-text entries in May, June, July and September 2026, dated by month only (11 of 15). The Python SDK repository had 4 open issues on 9 October. Replies were not sampled (5 of 10). Official `cartesia` and `@cartesia/cartesia-js` SDKs both reached 4.2.0 on 2 September 2026, generated for API version 2026-08-14, and the MCP server's v0.26.1 is dated 8 October (15). CI workflows, release automation and lock files in both SDK repositories (8 of 10).\n\nThe checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):\n\n- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.\n- 20, at least three releases or dated changelog entries in the last 90 days.\n- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.\n- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).\n- 10, package health, current dependencies and CI.\n\nModels are read for deprecation notice periods and model churn rather than release counts.\n\n## What we couldn't check\n\nWhat we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.\n\n- unchecked: trust.cartesia.ai is drawn by script, so the certifications, the sub-processor list and any disclosure policy or bug bounty were not read\n- unchecked: the duration and scope of the five API-wide incidents on the status history (1 August, 2 and 3 September, 7 and 8 October 2026). Only titles and dates were shown to us\n- unchecked: whether the Free plan asks for a card at sign-up, and how keys are rotated or revoked in the Playground, which is behind a sign-in\n- unchecked: replies on the SDK issue trackers, and weekly download counts on npm and PyPI\n- unchecked: the OpenAPI file at `/latest.yml` was not fetched whole. The definitions were read as embedded in the reference pages' Markdown twins\n- The Terms of Service forbid using 'automation software (bots)' and 'any robot, spider, crawlers, scraper, or other automatic device' that accesses the Services to monitor or collect data, and forbid using the Services to develop competing products. No benchmarking clause was found. Recorded as a fact with no deduction. It matters before any probe is run\n- docs.cartesia.ai's robots.txt carries `Content-Signal: ai-train=yes, search=yes, ai-input=yes`. www.cartesia.ai allows every path. status.cartesia.ai and trust.cartesia.ai serve no robots.txt\n- The lead described a REST batch API for Ink 2. The docs say batch `/stt` accepts `ink-whisper` only and mark `ink-2` batch as 'Not available yet'. The listing is named Cartesia Ink and covers both models\n- The model page dates `ink-2` 17 September 2026 while the changelog announced Ink 2 in May 2026. The September date is read as the multilingual release, and `lastRelease` uses it\n- Cartesia's pricing FAQ quotes $0.39 an hour for Ink on the Scale plan. Plan arithmetic (10,800 credits an hour at $299 for 8 million) gives $0.40\n- The privacy policy says the services are designed for users in the United States only, while the docs sell regional deployments in the EU, UK, India and Australia\n- The access token page lists an OpenAI-compatible `/audio/transcriptions` endpoint under the STT grant. No reference page for it was found in the docs index\n- The `cartesia-mcp` repository has no licence file in its root. The SDK repositories are Apache-2.0\n- Both SDK repositories returned 134 stars and 4 open issues from the GitHub API on 9 October. The match was not investigated\n- Whether this listing should stand apart from the Cartesia text-to-speech listing is the editor's call. It shares the API host, keys, plans, terms and status page, and has its own models, endpoints, prices and status components\n\n## Weaknesses\n\n- The terms let Cartesia train models on inputs and outputs unless otherwise agreed. Opting out is a form in the Playground's data controls\n- Zero data retention is an Enterprise plan setting, and no retention period for other plans was found in the terms, privacy policy or docs\n- `ink-2` supports English, French, Hindi, Japanese and Spanish only and has no batch endpoint. `POST /stt` accepts `ink-whisper` only\n- No diarisation or speaker labels in the reviewed documentation, and realtime input is raw mono PCM with `encoding` and `sample_rate` required\n- The terms of 14 June 2024 forbid automation software (bots) and any robot or scraper that accesses the Services to collect data, which matters before any probe is run\n- No SLA was found, and a 429 for exceeding the concurrency limit is documented without a `Retry-After` header or backoff guidance\n\n## What costs an agent a turn today\n\nThe notes we give agents before they call it. Each one is a workaround an agent shouldn't need.\n\n- Use `wss://api.cartesia.ai/stt/turns/websocket` with `model=ink-2`, `encoding`, `sample_rate` and `cartesia_version=2026-08-14`. All four are required.\n- Send raw mono audio in chunks of about 100 ms at the speed it was spoken. Pushing a whole file into the socket can return an internal server error.\n- Read the final text from `turn.end` only. `transcript` is cumulative within a turn, so joining `turn.update` events duplicates text.\n- Send `{\"type\": \"close\"}` after the last audio and keep reading until the server closes the socket, or the buffered tail is lost.\n- Check `encoding` and `sample_rate` against the source before sending. The docs say the server might not return an error when they are wrong.\n\n## When it's done\n\nSend what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `\"kind\": \"dispute\"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.\n",
    "grade": "B",
    "score": 67.9,
    "assessed": "2026-10-09",
    "run": "October 2026 research run",
    "categories": [
      {
        "key": "security",
        "name": "Security \u0026 auth",
        "score": 52,
        "maxGain": 8.4,
        "reason": "Model reading. Standard keys (`sk_car_...`) and admin keys are separate types that are not interchangeable, and a server can mint an access token with an `stt` grant that lasts at most one hour. That token travels in the `access_token` query parameter on WebSockets, the API key itself goes in a header, and no key limited to speech-to-text was found (26, less 5 for the token in the URL, 21 of 30). The terms, last revised 14 June 2024, say inputs, outputs and interactions may be used to train models unless otherwise agreed, with an opt-out form in the Playground's data controls (5 of 20). Zero data retention covers STT audio and transcripts and is limited to the Enterprise plan. No retention period for other plans was found, and the DPA gives 180 days for deletion after termination (5 of 15). Credit usage and key metadata are readable through admin endpoints. No audit log was found (7 of 15). security.txt is valid until 3 October 2027 with a contact and a policy link to the trust centre. The docs and pricing page state SOC 2 Type II, HIPAA and PCI-DSS. The trust centre is drawn by script and was not read, and no bug bounty was found (14 of 20).",
        "checklist": [
          "- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.\n- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.\n- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.\n- 0 to 15, audit logs or per-call visibility for the operator.\n- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.",
          "Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-security"
      },
      {
        "key": "payments",
        "name": "Payments \u0026 pricing",
        "score": 35,
        "maxGain": 8.1,
        "reason": "No x402, MPP or L402 (0). The docs pricing page gives 3 credits a second for `ink-2` and the public pricing page gives plan prices and overage rates of $65, $45 and $38 per million credits, all read without a login (20). The Free plan is $0 with 20,000 credits a month, about 1 hour 51 minutes of `ink-2` audio. Neither page says whether sign-up asks for a card, and we did not sign up, so 15 of 20. A person creates the key in the Playground in a browser, and the hosted MCP server also needs a browser sign-in (0).",
        "checklist": [
          "The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).",
          "- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.\n- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for \"contact sales\" or prices behind a login.\n- 20, a free tier or trial that doesn't need a card.\n- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).",
          "Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.",
          "Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-payments"
      },
      {
        "key": "reliability",
        "name": "Reliability",
        "score": 73,
        "maxGain": 5.4,
        "reason": "Hosted reading. The status page at status.cartesia.ai (incident.io) lists Speech to Text as its own component in four regions (20). For July to October 2026 it showed 99.97% for the US, 99.994% for the EU and APAC and 100% for Australia. The history from 17 July to 8 October names no speech-to-text incident and lists five API-wide ones (1 August, 2 and 3 September, 7 and 8 October) whose durations were not shown to us, read as minor incidents only (20). Concurrency limits are published per plan, 8, 12, 20 and 60 concurrent STT requests, with a 3-minute idle timeout on sockets (15). Exceeding the limit returns 429 with the `concurrency_limited` code, and the SDKs retry 429 and 5xx twice with backoff on HTTP calls. No `Retry-After` header, backoff guidance or reconnection advice for a dropped socket was found (8 of 15). No SLA was found for any plan (0). `ink-2` is marked Stable, with `ink-preview` as the preview line (10).",
        "checklist": [
          "Hosted APIs, MCP servers, models and platforms.",
          "- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).\n- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.\n- 15, rate limits documented with numbers.\n- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.\n- 10, an SLA published for any paid tier.\n- 10, the surface agents use is generally available, not beta or preview.",
          "Local packages, SDKs, frameworks and stdio MCP servers.",
          "- 20, installs from an official package with supported runtimes stated.\n- 25, a public CI and test suite, passing on the default branch.\n- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).\n- 15, semver discipline and breaking changes called out in a changelog.\n- 15, version 1.0 or later, or declared stable.",
          "Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-reliability"
      },
      {
        "key": "ergonomics",
        "name": "Agent ergonomics",
        "score": 74,
        "maxGain": 4.2,
        "reason": "API reading, as for the other speech-to-text listings. Turn events carry a cumulative transcript that is never revised, so an agent reads one `turn.end` a turn, and word timestamps are opt-in on the manual and batch endpoints (20 of 25). There is nothing to page through. Batch takes a file of any length without client chunking, keyterms steer vocabulary, and there is no diarisation or subtitle output (12 of 20). Errors are structured JSON with `error_code`, `title`, `message`, `request_id` and an optional `doc_url` on API version 2026-03-01 and later (17 of 20). Transcription holds no state to duplicate and the SDKs retry failed HTTP calls. A realtime session cannot be resumed, and nothing on reconnecting was found (12 of 20). Realtime needs `model`, `encoding`, `sample_rate` and the version, turn settings have documented defaults, and there are official Python and TypeScript SDKs. Realtime accepts raw PCM only (13 of 15).",
        "checklist": [
          "- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).\n- 20, pagination, filtering and output-size controls.\n- 20, actionable, documented error responses, codes and messages an agent can recover from.\n- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.\n- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.",
          "Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-ergonomics"
      },
      {
        "key": "transparency",
        "name": "Transparency \u0026 trust",
        "score": 67,
        "maxGain": 2.9,
        "reason": "The service is closed and the SDKs are Apache-2.0. The Terms of Service cover the website, its subdomains and APIs, were last revised on 14 June 2024, describe a platform for synthesised sound and never mention transcription (12 of 30). The terms and privacy policy agree that content may be used for training with an opt-out. The DPA says Cartesia will not retain or use customer personal data beyond performing the services, which sits uneasily with the training licence. No retention period is given outside zero data retention, and the privacy policy says the services are designed for users in the United States only while regional endpoints are sold in four other regions (12 of 30). A deprecations page with dated sunsets and a stated promise to preserve fields for a pinned `Cartesia-Version` (15 of 20). The DPA promises 30 days' notice of sub-processor changes to subscribers and points to a list in the trust centre, which we could not read. Regional deployments are named for Enterprise only (10 of 20).",
        "blend": "editorial 49, provenance 84",
        "checklist": [
          "- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.\n- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).\n- 0 to 20, a deprecation policy or notices with dates.\n- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).",
          "The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-transparency"
      },
      {
        "key": "schema",
        "name": "Schema \u0026 documentation",
        "score": 89,
        "maxGain": 1.8,
        "reason": "Model reading. The docs index links an OpenAPI file (`latest.yml`) and AsyncAPI files for the WebSocket endpoints, and each reference page embeds its part of them (25). llms.txt with a Markdown twin of every page (10). The comparison page says which of the three endpoints to use and when, and the turn detection guide explains each event and threshold (18 of 20). The AsyncAPI definition has enums for `encoding` and `cartesia_version`, minimum and maximum values on the four turn settings and required fields marked. `model` is a plain string on the auto endpoint (13 of 15). Example scripts for each endpoint in Python and TypeScript and an errors page with nine codes. `model_sunsetted` appears on the deprecations page and not in that table, and the docs say a wrong `encoding` or `sample_rate` might not return an error (11 of 15). Dated API versions through the `Cartesia-Version` header and a changelog grouped by month with no day dates (12 of 15).",
        "checklist": [
          "APIs and MCP servers.",
          "- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).\n- 10, llms.txt or Markdown docs served for agents.\n- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.\n- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.\n- 0 to 15, examples and documented error responses.\n- 15, versioning and a public changelog.",
          "Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-schema"
      },
      {
        "key": "maintenance",
        "name": "Maintenance \u0026 community",
        "score": 85,
        "maxGain": 1.3,
        "reason": "Model reading, as for the other hosted speech-to-text listings. The model page dates the current `ink-2` and `ink-preview` 17 September 2026, 22 days before the check (30). The deprecations page gives dated sunsets, and the July 2026 changelog announced three text-to-speech sunsets for 20 October 2026. No minimum notice period is stated and no Ink model has been retired (8 of 12). No speech-to-text model was retired or renamed in the last 90 days (8 of 8). The changelog has speech-to-text entries in May, June, July and September 2026, dated by month only (11 of 15). The Python SDK repository had 4 open issues on 9 October. Replies were not sampled (5 of 10). Official `cartesia` and `@cartesia/cartesia-js` SDKs both reached 4.2.0 on 2 September 2026, generated for API version 2026-08-14, and the MCP server's v0.26.1 is dated 8 October (15). CI workflows, release automation and lock files in both SDK repositories (8 of 10).",
        "checklist": [
          "- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.\n- 20, at least three releases or dated changelog entries in the last 90 days.\n- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.\n- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).\n- 10, package health, current dependencies and CI.",
          "Models are read for deprecation notice periods and model churn rather than release counts."
        ],
        "checklistUrl": "https://www.anchorterminal.com/benchmark/#checklist-maintenance"
      }
    ],
    "provenance": [
      {
        "label": "Domain age",
        "value": "cartesia.ai, registered 2023-05-10 (3 years)",
        "points": 7,
        "max": 15
      },
      {
        "label": "Terms of service",
        "value": "read, states 6 of the 7 things a reader expects, and has 2 clauses that cost points",
        "points": 5.1,
        "max": 10
      },
      {
        "label": "Privacy policy",
        "value": "read, states 4 of the 8 things a reader expects",
        "points": 7,
        "max": 10
      }
    ],
    "unchecked": [
      "unchecked: trust.cartesia.ai is drawn by script, so the certifications, the sub-processor list and any disclosure policy or bug bounty were not read",
      "unchecked: the duration and scope of the five API-wide incidents on the status history (1 August, 2 and 3 September, 7 and 8 October 2026). Only titles and dates were shown to us",
      "unchecked: whether the Free plan asks for a card at sign-up, and how keys are rotated or revoked in the Playground, which is behind a sign-in",
      "unchecked: replies on the SDK issue trackers, and weekly download counts on npm and PyPI",
      "unchecked: the OpenAPI file at `/latest.yml` was not fetched whole. The definitions were read as embedded in the reference pages' Markdown twins",
      "The Terms of Service forbid using 'automation software (bots)' and 'any robot, spider, crawlers, scraper, or other automatic device' that accesses the Services to monitor or collect data, and forbid using the Services to develop competing products. No benchmarking clause was found. Recorded as a fact with no deduction. It matters before any probe is run",
      "docs.cartesia.ai's robots.txt carries `Content-Signal: ai-train=yes, search=yes, ai-input=yes`. www.cartesia.ai allows every path. status.cartesia.ai and trust.cartesia.ai serve no robots.txt",
      "The lead described a REST batch API for Ink 2. The docs say batch `/stt` accepts `ink-whisper` only and mark `ink-2` batch as 'Not available yet'. The listing is named Cartesia Ink and covers both models",
      "The model page dates `ink-2` 17 September 2026 while the changelog announced Ink 2 in May 2026. The September date is read as the multilingual release, and `lastRelease` uses it",
      "Cartesia's pricing FAQ quotes $0.39 an hour for Ink on the Scale plan. Plan arithmetic (10,800 credits an hour at $299 for 8 million) gives $0.40",
      "The privacy policy says the services are designed for users in the United States only, while the docs sell regional deployments in the EU, UK, India and Australia",
      "The access token page lists an OpenAI-compatible `/audio/transcriptions` endpoint under the STT grant. No reference page for it was found in the docs index",
      "The `cartesia-mcp` repository has no licence file in its root. The SDK repositories are Apache-2.0",
      "Both SDK repositories returned 134 stars and 4 open issues from the GitHub API on 9 October. The match was not investigated",
      "Whether this listing should stand apart from the Cartesia text-to-speech listing is the editor's call. It shares the API host, keys, plans, terms and status page, and has its own models, endpoints, prices and status components"
    ],
    "weaknesses": [
      "The terms let Cartesia train models on inputs and outputs unless otherwise agreed. Opting out is a form in the Playground's data controls",
      "Zero data retention is an Enterprise plan setting, and no retention period for other plans was found in the terms, privacy policy or docs",
      "`ink-2` supports English, French, Hindi, Japanese and Spanish only and has no batch endpoint. `POST /stt` accepts `ink-whisper` only",
      "No diarisation or speaker labels in the reviewed documentation, and realtime input is raw mono PCM with `encoding` and `sample_rate` required",
      "The terms of 14 June 2024 forbid automation software (bots) and any robot or scraper that accesses the Services to collect data, which matters before any probe is run",
      "No SLA was found, and a 429 for exceeding the concurrency limit is documented without a `Retry-After` header or backoff guidance"
    ],
    "agentNotes": [
      "Use `wss://api.cartesia.ai/stt/turns/websocket` with `model=ink-2`, `encoding`, `sample_rate` and `cartesia_version=2026-08-14`. All four are required.",
      "Send raw mono audio in chunks of about 100 ms at the speed it was spoken. Pushing a whole file into the socket can return an internal server error.",
      "Read the final text from `turn.end` only. `transcript` is cumulative within a turn, so joining `turn.update` events duplicates text.",
      "Send `{\"type\": \"close\"}` after the last audio and keep reading until the server closes the socket, or the buffered tail is lost.",
      "Check `encoding` and `sample_rate` against the source before sending. The docs say the server might not return an error when they are wrong."
    ],
    "recheck": "https://www.anchorterminal.com/builders/#disputes"
  },
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-10",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.4",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run"
  }
}