Novita AI

# Fix list: Novita AI

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/novita-ai, the October 2026 research run, assessed 9 October 2026. Grade D, 53.5 out of 100.

This 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.

For a coding agent working on Novita AI: 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.

## 1. Reliability, 28 out of 100, up to 14.4 more on the total

Why it scored 28: Hosted reading. No status page was found. None is linked from the home page, the pricing page, the docs introduction, either `llms.txt` or the docs site map (0), so there is no incident record to read (0). The rate limit page documents five account tiers set by monthly top-ups, from under $50 to $10,000 and over, and limits counted in RPM and TPM per model. The figures per model are drawn by script and were not readable, so this line takes part marks (8 of 15). 429 with `RATE_LIMIT_EXCEEDED` or `TOKEN_LIMIT_EXCEEDED`, advice to throttle and use exponential backoff, and a billing page that says 429, 500, 503 and 504 are not charged. No `Retry-After` header was found (10 of 15). No SLA was found. The home page states 99.5% uptime, and the terms say availability is not guaranteed (0). Generally available (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 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.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 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.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 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.

## 2. Payments & pricing, 35 out of 100, up to 8.1 more on the total

Why it scored 35: No machine payment protocol was found in the docs index, the site map, the pricing page or the skill repository (0). Per-token prices for each model on the pricing page without a login, such as `deepseek/deepseek-v4.1-flash` at $0.30 in and $1.20 out per 1M tokens (20). Two language models are listed as free and the FAQ describes a voucher for new users. Whether a card is needed to use either was not stated in the pages read (15 of 20). A person signs up in a browser, keys are created only in the console, and the OAuth route needs the user to sign in and approve (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

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.
- 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.
- 20, a free tier or trial that doesn't need a card.
- 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.

## 3. Schema & documentation, 62 out of 100, up to 6.2 more on the total

Why it scored 62: Model reading. The OpenAPI 3.1.0 document at `/.well-known/openapi.json` is a 964-byte discovery file with four paths and no schemas, and the API reference is written by hand (6 of 25). `llms.txt` on both hosts, a Markdown copy of every docs page and an `llms-full.txt` (10). Guides for function calling, structured output, prompt caching, batch and reasoning say what each is for, and the chat reference describes 136 request and response fields (14 of 20). Fields are typed with required flags and defaults. Enums and ranges are stated in prose only (10 of 15). Curl and Python examples on each guide, error tables by product, a common errors guide and a table of which statuses are billed. The FAQ says `json_schema` and `top_logprobs` are unsupported while the reference and a guide document both (10 of 15). A dated changelog with 84 entries since February 2024. The API has no version scheme beyond `/v1` (12 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 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.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 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.

## 4. Security & auth, 65 out of 100, up to 6.1 more on the total

Why it scored 65: Model reading. Bearer keys with the `sk_` prefix, shown once, up to 10 an account, with an expiry of 24 hours, 30 days, 90 days or none fixed at creation, deletable in the console, and a per-key model access policy and source IP allowlist that can be set by API. OAuth with PKCE and the scopes `api` and `balance:read` is described in `auth.md`, with no token revocation endpoint. No documented way to send a key in a URL was found (26 of 30). The terms say content is not used to train Novita's models or improve the services by default (17 of 20). The terms state zero data retention, with exceptions for law and for what is necessary to run or support the service, and allow automated safety screening. Batch input files are kept for 15 days (11 of 15). Billing can be queried per key, and budget, monitoring and metrics pages are listed in the docs index and were not read (7 of 15). The trust centre at trust.novita.ai is drawn by script and was not read. `/.well-known/security.txt` returns 404, and no disclosure policy or bug bounty was found (4 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 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.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 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.
- 0 to 15, audit logs or per-call visibility for the operator.
- 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.

## 5. Agent ergonomics, 72 out of 100, up to 4.6 more on the total

Why it scored 72: Model reading of the checklist (tool use, structured output, caching, context, batch, SDKs, errors). `tools` with a `strict` flag and a worked example. The list of models that support it is drawn by script, and `tool_choice` is absent from the chat reference (13 of 20). `response_format` takes `json_object` and `json_schema` with `strict` in the reference, while the FAQ says only `json_object` works (9 of 15). Prompt caching is automatic on supported models, cache-read prices are on the pricing page and `cached_tokens` is returned in usage (12 of 15). Context of 1M tokens on DeepSeek V4, GLM 5.3 and Kimi K3 per the pricing page. Output caps were not read (12 of 15). OpenAI-style files and batches with a 24-hour window, one model per file and an introductory 50 per cent discount (10). The OpenAI client works with a changed base URL. Both vendor SDK repositories are archived (5 of 10). Named error codes, a 429 that separates request and token limits, and a `trace_id` for support (11 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 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).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 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.

## 6. Transparency & trust, 56 out of 100, up to 3.9 more on the total

Made of editorial 42, provenance 69.

Why it scored 56: Closed service. The terms set Delaware law and arbitration but name the party only as Novita AI, with no corporate form or address. The skill and the MCP server are MIT (10 of 30). The terms state zero data retention and no training by default, and the privacy policy lists retention periods (account data for the life of the account plus 7 years, technical data 2 years). The privacy policy says it does not cover content processed for customers and points to customer agreements. No DPA was found (16 of 30). Retirement notices in the changelog carry a UTC time and a replacement. The terms reserve the right to discontinue any service without notice (12 of 20). The privacy policy names categories of recipients only and transfers that include the United States. No sub-processor list was found (4 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 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.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 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.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: novita.ai, registered 2023-09-18 (3 years) (7 of 15)
- Terms of service: read, states 6 of the 7 things a reader expects, and has 1 clause that costs points (7.1 of 10)
- Status page: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 62 out of 100, up to 3.3 more on the total

Why it scored 62: Model reading. The newest changelog entry is dated 28 September 2026 (30). That entry retires `deepseek/deepseek-v4-flash` at 16:00 UTC on 9 October 2026 with a named replacement, which is 11 days' notice. No written notice period was found (6 of 12). 11 changelog entries in the 90 days to 9 October 2026, of which one was read (4 of 8). A dated changelog, a Discord server and a support address (12 of 15). `novitalabs/python-sdk` and `novitalabs/javascript-sdk` are archived on GitHub, and issue replies were not examined (2 of 10). The docs rely on the OpenAI client, the SDK page still names the two archived packages, and the skill repository reached version 1.4.0 on 24 September 2026 (5 of 15). The JavaScript SDK repository has a publish workflow only and the Python one has none (3 of 10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 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.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 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.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What 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.

- unchecked: the RPM and TPM figures per model and tier, and the lists of models that support tools, structured output, caching and batch, are drawn by script on the docs pages and were not read
- unchecked: the trust centre at trust.novita.ai is a script-drawn Vanta page, so any SOC 2 or ISO 27001 report, sub-processor list or disclosure policy there was not read. Its robots.txt address returned the same page shell
- unchecked: no request was sent to api.novita.ai or to the MCP endpoint at novita.ai/mcp, so the 429 body, response headers and the live model list rest on the docs
- unchecked: the budgets, model access, network access, monitoring, payment methods, Anthropic compatibility and team guides are listed in the docs index and were not read, to keep to the page limit. About 20 docs pages were read, over the limit of about fifteen
- unchecked: only the newest of the 11 changelog entries in the last 90 days was read. Their dates come from the docs site map
- unchecked: issue replies on the GitHub repositories, the blog and Discord were not read. PyPI download counts were not read because its robots.txt closes `/pypi/`
- No status page, SLA, DPA, sub-processor list, security.txt or legal entity name was found. Whether they exist on request is not established
- Whether a card is needed to use the free models or the new-user voucher was not stated in the pages read, and the voucher amount was not found
- The LLM FAQ says `json_schema` and `top_logprobs` are not supported, while the chat reference and the structured outputs guide document both. Not checked by a call
- The home page shows DeepSeek V4 Pro at $1.74 in and $3.48 out per 1M tokens, and the pricing page shows $1.60 and $3.20 on the same day
- The home page and the docs introduction carry text for AI agents that asks them to read a skill file and follow it. It was not acted on, and it is recorded as a fact with no deduction
- The lead was right about the interface and the skill repository. It did not mention the OAuth route, the MCP server or that both SDK repositories are archived
- `lastRelease` is the date of the newest changelog entry, since the service has no version
- Agent Sandbox, GPU instances, serverless GPUs and dedicated endpoints share the key and terms and were not graded here

## Weaknesses

- No status page or SLA was found on the home page, the docs index or the docs site map. The home page states 99.5% uptime with no supporting document
- The rate limit page explains five tiers by monthly top-up, and the RPM and TPM figures per model are drawn by script and were not readable
- `deepseek/deepseek-v4-flash` was retired on 9 October 2026 on a notice dated 28 September 2026, and the named replacement is priced higher
- The Python and JavaScript SDK repositories are archived, last changed in November 2024 and March 2025, and the SDK docs page still tells users to install them
- The terms name the contracting party only as Novita AI, with no corporate form or address, and no DPA or sub-processor list was found
- The Acceptable Use Policy forbids scripts, bots or crawlers that extract data from the Site. Recorded as a fact, and it matters before any probe is run

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Set the OpenAI client base URL to `https://api.novita.ai/openai` and send the key as `Authorization: Bearer`. Keys start with `sk_`
- Do not close a connection to stop generation. A request that reached the model is billed in full under status 499, so cap output with `max_tokens`
- On 429, check whether the code is `RATE_LIMIT_EXCEEDED` or `TOKEN_LIMIT_EXCEEDED` and back off exponentially. No `Retry-After` header is documented
- Read the changelog for retirement notices before pinning a model id. One retirement in October 2026 came with 11 days' notice
- Ask the account owner for a key with an expiry, a model access policy and an IP allowlist. Keys are created and deleted only in the console

## When it's done

Send 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.