Published Updated
In one or two sentences, before anything else.
Novita AI is an inference host for open-weight models such as DeepSeek, GLM, Kimi and Qwen, with image, video, GPU and sandbox services; it does not resell GPT, Claude or Gemini. Its per-token prices match or undercut the makers' own, its terms promise zero data retention by default, and it does not name the company behind it.
What it can do, one sentence each, and whether we actually tested it.
96 chat models in its public model list, including DeepSeek V4.1 Flash and V4 Pro 0813, GLM 5.3, Kimi K3 and K2.6, Qwen3.8, MiniMax M3, Gemma 4 31B and gpt-oss-120b.
Count from Novita's public model list on October 21, 2026. The home page claims 200+ models including media. No GPT, Claude or Gemini model is offered.
Novita's terms say request content is not logged for human review or kept beyond producing the output, and is not used to train Novita's models by default.
Section 10.2 of Novita's terms; Standard Compute's provider listing also reads it as the default. Automated safety screening may still apply.
Chat, completions, embeddings, rerank and batch at https://api.novita.ai/openai, and an Anthropic-compatible endpoint at https://api.novita.ai/anthropic, with function calling, structured outputs, reasoning and prompt caching.
Per Novita's docs. The Anthropic page does not list which models it supports.
A command-line tool, novi, that starts Claude Code, Codex, OpenCode, Pi or DeepSeek Harness on Novita models for one session without changing your own settings.
Beta, macOS and Linux. Novita says tools, MCP servers, subagents and web search keep working.
The novitalabs/novita-skills repository holds SKILL.md skills for Novita's models, GPUs, sandbox and integrations, installed with npx skills add.
In Novita's GitHub organization; the README does not name the agents it targets.
@novitalabs/novita-mcp-server lets an MCP client list, create, start, stop and delete GPU instances, templates and network storage.
Beta. Its README says it only supports GPU instance management; it does not call models.
GPU instances, serverless GPUs and agent sandboxes are billed from the same balance; an H100 SXM 80GB costs 3.39 USD an hour on demand or 1.70 USD spot.
H100 prices from a UsagePricing capture of Novita's GPU page on September 23, 2026.
Novita serves 48 models as a provider on OpenRouter, runs 36 bots on Poe and is listed in Hugging Face Inference Providers.
Prices through OpenRouter can differ from Novita's own list.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
About 4,000 tool-calling requests replayed against Novita's Kimi K2-0905 and K2 Thinking endpoints and compared with Moonshot's own API.
What happened
Novita passed on K2-0905 with 100% schema accuracy (1,299 of 1,299). On K2 Thinking it scored 72.22% trigger similarity, just under Moonshot's 73% bar, with 96.46% schema accuracy. The test is a year old and covers one model family.
Case 2
Input
Repeated speed and latency measurements of Kimi K2.6 and DeepSeek V4 Pro 0813 across every host that serves them.
What happened
Novita was 7th of 8 hosts on Kimi K2.6 at 74 tokens per second and the slowest to first chunk (3.88 seconds); on DeepSeek V4 Pro 0813 it was 6th of 10 at 70 tokens per second. On smaller models it did better, such as 194 tokens per second on DeepSeek V4.1 Flash.
Case 3
Input
export ANTHROPIC_BASE_URL=https://api.novita.ai/anthropic, ANTHROPIC_AUTH_TOKEN set to a Novita key, ANTHROPIC_MODEL=moonshotai/kimi-k2-instruct, then claude in the project folder.
What happened
Claude Code talks to Novita's Anthropic-compatible endpoint with a Kimi model in place of Claude. The guide does not list the supported models or say whether prompt caching and web search work; Novita Connect's novi claude does the same setup per session.
Case 4
Input
Daily coding sessions with about 150,000-token DeepSeek V4 Flash contexts, run on Novita and on DeepSeek's own API.
What happened
The user reported about 99.5% cache hits on Novita against 99.9% on DeepSeek, and occasional whitespace-counting errors only on Novita. One user, no logs.
What it does badly, with the evidence, and what we could not settle.
Weak
No closed models.
Evidence: No GPT, Claude or Gemini model appears on Novita's pricing page, home page or public model list on October 21, 2026; the only Google and OpenAI models are the open-weight Gemma and gpt-oss.
Slow first token on large models.
Evidence: Artificial Analysis measured Novita slowest of 8 hosts to first chunk on Kimi K2.6 (3.88 seconds) and 9th of 10 on DeepSeek V4 Pro 0813 (2.18 seconds).
Just under Moonshot's tool-call bar on Kimi K2 Thinking.
Evidence: Moonshot's K2 Vendor Verifier (November 2025) scored Novita at 72.22% trigger similarity against an acceptance bar of 73%.
No off-peak discount on DeepSeek.
Evidence: Novita lists DeepSeek V4 Pro 0813 at 1.32 / 3.96 USD per million tokens at all hours; DeepSeek's own API charges 0.66 / 1.98 USD off-peak.
Fees are non-refundable and the Coding Plan draws complaints.
Evidence: Novita's terms say "All fees are non-refundable" except verified outages, billing errors or legal requirements claimed within 30 days. Trustpilot (2.5 out of 5 from 9 reviews) has 2026 reports of hard-to-find cancellation and a 40 USD charge where 20 USD was shown.
Not for buyers who need closed models or a named counterparty.
Evidence: Anyone who needs GPT, Claude or Gemini, a legal entity to contract with, refunds or the lowest latency on large models should look elsewhere: Novita offers none of these, and its terms and privacy policy name no company or address.
Uncertain
Where inference runs, and whether you can keep it in one region.
Why unsure: Novita's privacy policy allows transfers to the US and elsewhere and names no inference regions; Standard Compute's listing records the location as not disclosed and no region pinning.
How each model is quantized on Novita's own API.
Why unsure: OpenRouter lists fp8 for most of Novita's endpoints and unknown for Kimi K2.6; Novita's pricing page does not say.
The new-user voucher amount and the rate limits.
Why unsure: Novita's FAQ confirms new-user vouchers without an amount (third parties report 0.50 USD or about 10 USD), and its docs describe rate limits both by top-up tier and by verification level.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Operator | Novita AI; no legal entity or address named in its terms or privacy policy; terms under Delaware law | Job listings point to San Francisco; other listings cite Hong Kong or Singapore. No funding round found. Source: novita.aichecked |
| DeepSeek V4 Pro 0813 on Novita, per million tokens | 1.32 USD input, 3.96 USD output | DeepSeek's own price: the same at peak, 0.66 / 1.98 USD off-peak. Source: novita.aichecked |
| Platform fee | None stated ("No hidden fees"); cached input about a tenth of normal input; batch at an introductory 50% off | Source: novita.aichecked |
| Minimum top up | Manual top-ups must be more than 10 USD; auto top-up available | Source: docs.novita.aichecked |
| Payment methods | Stripe for all payments; no crypto, PayPal, Alipay or invoicing found in the docs | Source: docs.novita.aichecked |
| Free tier | New-user and referral vouchers of unstated size, a few 0 USD models, and 200 USD over 3 months for eligible startups | Novita's FAQ confirms the vouchers without an amount; the startup program is described on Novita's blog. Source: docs.novita.aichecked |
| Models | 96 chat models in the public list, all open-weight; image, video and audio APIs besides | Source: api.novita.aichecked |
| API formats | OpenAI-compatible chat, completions, embeddings, rerank and batch; Anthropic-compatible Messages; function calling, structured outputs, prompt caching | No Responses format or web search found in the docs index. Source: docs.novita.aichecked |
| MCP | @novitalabs/novita-mcp-server (beta) manages GPU instances only; no MCP pass-through in API requests documented | Source: github.comchecked |
| Agent Skills | novitalabs/novita-skills: novita-ai, novita-sandbox and a brand design skill (SKILL.md) | Source: github.comchecked |
| Coding agents | Claude Code and Cursor guides; Novita Connect (beta) launches Claude Code, Codex, OpenCode, Pi and DeepSeek Harness | Source: docs.novita.aichecked |
| Privacy and data retention | Zero data retention by default; no training on content by default; SOC 2 badge in the site footer | Source: novita.aichecked |
| Status page and uptime | No public status page; OpenRouter one-day uptime 99.93% on DeepSeek V4 Pro and 100% on Kimi K2.6 | Novita's home page claims 99.5% uptime; its FAQ says to ask on Discord about platform status. Source: openrouter.aichecked |
| Refunds | All fees non-refundable, except verified Novita outages, billing errors or legal requirements claimed within 30 days | Source: novita.aichecked |
Answer first, then why, an example, the limits and what to do next.
Novita AI matches DeepSeek's peak list price and undercuts Moonshot's own Kimi K2.6 price. DeepSeek V4.1 Flash costs 0.30 / 1.20 USD per million tokens on both, and Kimi K2.6 costs 0.80 / 3.40 USD on Novita against 0.95 / 4.00 USD from Moonshot.
Limits
DeepSeek halves its price off-peak and Novita does not, and Novita's endpoints on OpenRouter were cheaper than its own list for DeepSeek V4.1 Flash and GLM 5.3.
Novita AI does not offer GPT, Claude or Gemini models. It hosts open-weight models only, including Google's Gemma and OpenAI's gpt-oss, so for closed models you need the makers' APIs or a relay such as OpenRouter or Atlas Cloud.
Novita AI is a working host with a zero-data-retention policy, but it does not say which company you are dealing with. Its terms name no legal entity or address, fees are non-refundable, and it does not say where inference runs.
Next step
Test with a small top-up, and pin Novita through OpenRouter if you want a fallback.
Novita AI does not name its owner, legal entity or funding in its terms or privacy policy. The terms follow Delaware law; job listings point to San Francisco, while other listings cite Hong Kong or Singapore, and no funding round was found in news or databases.
Novita AI publishes an MCP server, but it only manages GPU instances. The beta @novitalabs/novita-mcp-server cannot call models; for agents, Novita offers SKILL.md skills in novitalabs/novita-skills and the Novita Connect CLI, which it says keeps your MCP servers working.
Chutes, Atlas Cloud and DeepInfra are the closest alternatives for open models. Chutes sells Kimi K2.6 at 0.45 / 1.95 USD in trusted execution environments, Atlas Cloud adds closed and media models, and Perkstack found DeepInfra 15% to 40% cheaper than Novita on large models.
Why
OpenRouter lets you compare all of them per model and fall back between hosts.
This page is built from Novita AI's site, docs, terms and public model list, DeepSeek's and OpenRouter's price data, Artificial Analysis measurements, Moonshot AI's K2 Vendor Verifier, Trustpilot reviews and Hacker News comments, all checked on October 21, 2026. Novita's claims are labelled as its own.
Novita AI hosts open-weight models and sells access by the token. Its public OpenAI-compatible model list had 96 chat models on October 21, 2026: 15 from DeepSeek, 14 GLM, 14 Qwen, 8 Kimi, 5 Llama, 5 MiniMax, 4 Xiaomi MiMo and 4 Gemma, plus gpt-oss-120b and models from Tencent, Baidu and StepFun. The home page claims 200+ models, counting image, video and audio APIs such as Kling v3.0, Wan 2.7, Seedance 1.5 Pro and Fish Audio text-to-speech. Novita also rents GPU instances, serverless GPUs and agent sandboxes.
It does not resell closed models. No GPT, Claude or Gemini model appears on its pricing page, home page or model list; the only Google and OpenAI models are the open-weight Gemma and gpt-oss. Many developers use Novita without visiting it: it serves 48 models as a provider on OpenRouter and runs 36 bots on Poe.
Novita's pricing page says there are no hidden fees. Cached input costs about a tenth of normal input, and its batch API has an introductory 50% discount.
| Model, per million input / output tokens | Novita list | Novita on OpenRouter | Maker's own price |
|---|---|---|---|
| DeepSeek V4 Pro 0813 | 1.32 / 3.96 USD | not listed separately | 1.32 / 3.96 USD peak, 0.66 / 1.98 USD off-peak (DeepSeek) |
| DeepSeek V4.1 Flash | 0.30 / 1.20 USD | 0.195 / 0.78 USD | 0.30 / 1.20 USD peak, 0.15 / 0.60 USD off-peak (DeepSeek) |
| Kimi K2.6 | 0.80 / 3.40 USD | 0.80 / 3.40 USD | 0.95 / 4.00 USD (Moonshot AI's endpoint on OpenRouter) |
| GLM 5.3 | 1.40 / 4.40 USD | 0.63 / 1.98 USD | 1.40 / 4.40 USD (Z.AI's endpoint on OpenRouter) |
Novita's prices come from its pricing page. Two things stand out. Novita charges DeepSeek's peak price at all hours and does not pass on DeepSeek's off-peak half price. And its endpoints on OpenRouter were cheaper than its own list for DeepSeek V4.1 Flash and GLM 5.3, so the same model can cost less through OpenRouter. Perkstack's weekly price checks (updated July 23, 2026) found Novita tied for cheapest on small models and typically 15% to 40% above DeepInfra on large ones.
Novita's own home page claims 99.5% uptime.
On Hacker News, OpenRouter users mostly speak well of Novita. Comments in 2026 recommend it for production, praise its zero-retention policy and cache hit rates, and one user named it among the few OpenRouter providers that did not degrade results for them (September 29). One user running DeepSeek V4 Flash sessions of about 150,000 tokens saw about 99.5% cache hits on Novita against 99.9% on DeepSeek's own API, and occasional whitespace-counting errors only on Novita (May 25). Complaints include three of four requests failing on an old Kimi K2 model (June 27) and a quoted view that Novita hosts many models cheaply without testing them well.
Trustpilot shows 2.5 out of 5 from 9 reviews on an unclaimed profile. The 2026 one-star reviews are about the Coding Plan subscription: cancellation hard to find, credits gone in about two hours, a 40 USD charge where 20 USD was shown, and slow email support. One reviewer says a 100 USD balance is needed to run more than three GPU instances. A five-star review calls Novita "by far the best OpenRouter provider for some models".
@novitalabs/novita-mcp-server in its GitHub organization. It is in beta and only manages GPU instances, templates, registry credentials and network storage; it does not call models. No MCP pass-through in API requests is documented. Novita Connect's page says MCP servers keep working when agents run on Novita models.SKILL.md skills: novita-ai (models, GPUs, sandbox, integrations, troubleshooting), novita-sandbox and a brand design skill, installed with npx skills add novitalabs/novita-skills. It does not name the agents it targets.https://api.novita.ai/anthropic; the guide's example uses Kimi K2 and does not list supported models or features. Novita Connect, a beta CLI called novi for macOS and Linux, starts Claude Code, Codex, OpenCode, Pi and DeepSeek Harness on Novita models without touching your own settings. There is also a Cursor guide.https://api.novita.ai/openai, with guides for function calling, structured outputs, reasoning and prompt caching. The Responses format and web search are not in its docs index.llms.txt for its docs.Section 10.2 of Novita's terms adopts a "Zero Data Retention" policy: content is not logged for human review or kept beyond producing the output unless law or support needs it, and it is not used to train Novita's models by default. Standard Compute's provider listing reads zero retention as the default, and records the inference location as not disclosed with no region pinning. The privacy policy (May 13, 2026) allows transfers to the US and elsewhere and does not say where inference runs. The site footer shows a SOC 2 badge.
The terms and privacy policy name no legal entity or address; the terms follow Delaware law. Job listings point to San Francisco, while other listings cite Hong Kong or Singapore, and no funding round was found. Fees are non-refundable except for verified Novita outages, billing errors or legal requirements claimed within 30 days. Manual top-ups must be more than 10 USD, paid through Stripe. There is no public status page; the FAQ says to ask on Discord.
Skip Novita AI if you need GPT, Claude or Gemini, a named company to contract with, refunds, or the lowest first-token latency on large models. It fits developers who want zero-retention access to new open models, directly or as a pinned OpenRouter provider.
Sources
Changelog
APIMODELS is a pay-as-you-go AI API relay that sells about 150 models through one key at its own prices, mostly below the makers' list prices, and spends the margin on backup channels: 98.62% of requests succeeded over 30 days, moderation rejections excluded. APIMODELS is our own product: it is made by the team behind GenGrades.
ZenMux is an LLM gateway, operated under its terms by NexaMind Singapore Pte. Ltd., that sells Claude, GPT, Gemini and many Chinese models through OpenAI, Anthropic and Vertex-style endpoints at the makers' list token prices. It adds automatic credits for bad or slow answers, which it calls insurance, and 20 to 200 USD monthly coding plans.
Vercel AI Gateway is Vercel's hosted API for about 480 text, image, video and speech models behind one key, billed from prepaid credits at the providers' list prices with no markup, including when you bring your own keys. It became generally available on August 21, 2025, and gives each Vercel team 5 USD of free credit a month.
Requesty is a London-based LLM gateway, run by Requesty Ltd and seed-funded by 20VC in 2025, that sells OpenAI- and Anthropic-compatible access to more than 600 models at list price plus a 5% top-up fee. It stands out for routing policies, spend controls and an EU endpoint in Frankfurt.
Poe API is Quora's developer API for the bots in its Poe chat app, opened on July 31, 2025, with OpenAI-compatible endpoints and an Anthropic-compatible one for official Claude bots. Calls spend the same points as a Poe subscription, so there is no separate API bill; current Claude bots cost about 1% more than Anthropic's list price.
OpenRouter is a New York company's API relay that sells access to more than 680 models from over 100 hosting providers through one key, at each provider's list price plus a 5.5% fee on credit purchases. Stripe agreed to buy it on August 19, 2026; no closing had been announced by October 21, 2026.