Novita AI Pricing in 2026: Every Model Priced, Verified

Novita AI prices more than sixty text models per token as of 2026-11-05, from $0.02 per million input tokens on Meta's Llama 3.1 8B up to $4.00 per million on Moonshot AI's Kimi K3.1 Pro Highspeed, and on the shared gpt-oss-120b checkpoint its $0.07 blended rate is the third cheapest of the 16 providers third-party benchmarking firm Artificial Analysis measures. Two things complicate the picture: a documented November 26, 2026 deprecation that removes 13 models — including gpt-oss-120b itself — from the catalog with no replacements, and a privacy policy that explicitly routes customer content to a separate agreement rather than a public attestation.

Key facts

• Frontier rows: DeepSeek V4.1 Flash (1017) lists at $0.30 input / $1.20 output per million tokens with cache read at $0.006 — a 50x cache discount (Novita pricing, accessed 2026-11-05).
• Moonshot AI's Kimi K3 lists at $3.00 / $15.00 per million with three-tier cache billing: cache read $0.30, 5-minute cache write $3.00, 1-hour cache write $6.00 (Novita pricing).
• Cheapest cache: Xiaomi's MiMo V2.6 Flash bills cached input at $0.0028 per million against $0.14 uncached on a 1M-token context — among the steepest cache discounts DeAI has verified on any rate card (Novita pricing).
• Third-party check: Artificial Analysis ranks Novita at $0.07 blended (7:2:1 cache/input/output) on gpt-oss-120b — third of 16 providers behind CoreWeave ($0.04) and DeepInfra ($0.05) — with 91% Endpoint Accuracy, a 12.59s first answer token, and a 0.0% error rate (Artificial Analysis, accessed 2026-11-05).
• Deprecation: 13 models retire at 16:00 UTC on November 26, 2026, with no replacements, including gpt-oss-120b, Llama 3.3 70B, and both Gemma 4 rows (Novita deprecation notice, published 2026-11-05).
• Batch discount: the OpenAI-compatible Batch API runs at an introductory 50% off input and output tokens for supported models, with a 24-hour completion window (Novita Batch Inference docs; Novita pricing).

What Novita is

Novita AI is an AI-native cloud that bundles three products behind one account: serverless model APIs (the LLM rate card this page prices), on-demand GPU instances, and an "Agent Sandbox" runtime for executing agent code. The LLM surface is OpenAI-compatible at base URL https://api.novita.ai/openai — the standard chat-completions swap documented in our OpenAI-compatible API explainer — and the company's marketing claims "200+ models" across text, image, audio, and video. DeAI counted roughly 60 text-generation rows on the serverless card on 2026-11-05, plus embeddings, image, video, speech, and an AI-search API reselling Exa and Tavily endpoints.

Unlike DeepInfra, which builds its own GPU fleet and publicly discusses its hardware strategy, Novita publishes no company-footing page DeAI could find: no funding announcements, no datacenter count, no disclosed infrastructure partners. Its site lists a Trust Center link in the footer navigation, but the URL did not resolve on 2026-11-05. That opacity is itself a data point for teams evaluating vendor risk on a budget host.

The full rate card, priced (2026-11-05)

Everything below is Novita's first-party serverless list rate, accessed 2026-11-05. "Cache read" is the per-token rate for cached prompt-prefix hits. This is the selection that matters for builders, not the whole catalog — image, video, audio, and search APIs are priced separately.

Open-weight frontier and near-frontier

Open-weight workhorses

Three structural notes. First, the Kimi K3.1 Pro rows carry a three-way cache structure — cache read, 5-minute cache write, and 1-hour cache write priced separately — that matches Anthropic's prompt-caching model rather than the implicit free caching most open-model hosts run; paying more to hold a prefix for an hour ($6.00/M on K3) is unusual in this market. Second, Novita still lists DeepSeek R1 0528 — a January-2025-era reasoning model — alongside the V4 generation, which tells you the card is additive rather than curated. Third, "Apodex 1.1 Mini" is listed as Free, the only zero-cost row.

The cache economics are the real story

Novita's headline rates sit mid-pack for the hosts DeAI tracks — our DeepInfra price card lists DeepSeek V4 Flash 0731 at $0.06/$0.18 against Novita's $0.44/$1.32 — but the cache-read column changes the arithmetic for agent-shaped workloads:

• DeepSeek V4.1 Flash: $0.006/M cached vs $0.30 uncached — a 50x discount, the deepest published cache-read rate DeAI has verified on a current frontier checkpoint.
• MiMo V2.6 Flash: $0.0028/M cached vs $0.14 — a 50x discount on a 1M context.
• GLM 5.3 Flash: $0.03/M cached vs $0.15 — 5x.
• Kimi K3: $0.30/M cache read vs $3.00 — 10x, but only for the 5-minute window; the 1-hour write costs full input price.

At Artificial Analysis's default 7:2:1 cache/input/output blend, Novita lands at $0.07 per million tokens on gpt-oss-120b — behind CoreWeave ($0.04) and DeepInfra ($0.05) but ahead of Groq ($0.15) and Cerebras ($0.39). For a long-system-prompt agent loop where most input tokens are cache hits, Novita's effective rate on DeepSeek V4.1 Flash approaches the cache-read floor; for cache-light workloads, DeepInfra and the budget rows on our cheapest-LLM-API roundup stay cheaper.

Speed and reliability: the other side of the trade

Artificial Analysis's live measurements on gpt-oss-120b (high) put Novita's first answer token at 12.59 seconds — roughly 6x slower than Cerebras (1.88s), 2.6x slower than Groq (4.73s), and slower than DeepInfra's standard endpoint (37.74s TTFT is DeepInfra Turbo; standard DeepInfra measured 44.82s, so Novita beats the standard DeepInfra endpoint while trailing the fast-silicon hosts badly). Its Endpoint Accuracy Index of 91% beats DeepInfra's standard endpoint (84%) and Cerebras (87%), and its 0.0% measured error rate is the cleanest in Novita's cohort. Read together: Novita behaves like a throughput-tolerant commodity fleet — accurate and stable, not fast. For user-facing streaming, benchmark the custom-silicon hosts in our Cerebras vs Groq comparison; for batch and background agents, the cache discounts dominate.

Novita offers 61 models on Artificial Analysis's tracker, and its measured intelligence leader is DeepSeek V4.1 Flash 1017 (Intelligence Index 33), followed by GLM-5.3 (32) and MiniMax-M3.1-Flash (29) — a catalog weighted toward the current Chinese open-weight frontier rather than the Llama/Gemma long tail.

Rate limits: tier-shaped, and the console holds the numbers

Novita's LLM rate limits are documented as a five-tier system keyed to top-ups over the trailing three calendar months: T1 for accounts that never exceeded $50/month, T2 at $50–$500, T3 at $500–$3,000, T4 at $3,000–$10,000, and T5 for at least $10,000 in one of the last three months. Limits are expressed as RPM (requests per model per minute) and TPM (tokens per model per minute), and they vary per tier — but the numeric table that should list the actual RPM/TPM values renders as a blank block in Novita's documentation as of 2026-11-05, so the real ceilings are effectively console-only. Increases route through Discord or a sales call, not a self-serve form.

Two billing rules are worth internalizing before building on Novita. Rejected requests (HTTP 400/401/403/429) are not billed, and server-side failures (500/503/504) are not billed — but HTTP 499, a client disconnect after inference has started, is billed in full, which makes aggressive client-side timeouts a direct cost. The Batch API mirrors OpenAI's: JSONL input, a 24-hour completion window, same endpoints, and the introductory 50% discount on both input and output tokens.

Privacy: a policy with a deliberate carve-out

Novita's privacy policy (effective May 13, 2026) states plainly that personal information "will not be used for model training." Read the scope clause before relying on it: the policy "does not apply to content that we process on behalf of customers while providing services," which is governed instead by customer agreements. In other words, the public no-training commitment covers account-holder personal information, not prompts — the data builders actually care about has no public, contractually specific commitment at all. DeAI found no independent attestation (no SOC 2 summary, no TEE claim, no third-party audit reference) behind any data-handling statement.

That places Novita in the same "policy, not attestation" category as DeepInfra, Groq, and most of the hosts we track — distinguishable from the TEE-attested tier where confidentiality holds by construction. Our zero-data-retention explainer covers how to tell the two categories apart and why policy statements alone never earn the zero-retention label.

The November 26 cliff

The single most date-sensitive fact on this page: Novita's own deprecation notice (published 2026-11-05) retires 13 models at 16:00 UTC on November 26, 2026, with no replacements offered. The affected list includes OpenAI's gpt-oss-120b — the benchmark checkpoint every provider is compared on — plus Meta's Llama 3.1 8B and Llama 3.3 70B, Google's Gemma 3 27B and both Gemma 4 rows, NVIDIA's Nemotron 3 Nano 30B, DeepSeek OCR 2, the Sao10K Stheno fine-tune, AutoGLM Phone 9B, and the BGE reranker. Affected model IDs simply stop working; applications must migrate before the cutoff.

Two readings. Charitably, this is catalog hygiene — the retired rows are the oldest models on the card, and a serverless operator managing real GPU capacity has to draw a line somewhere. Less charitably, it is the sharpest catalog shrink among the hosts DeAI tracks — Groq's August repricing at least kept its GPT-OSS rows, and DeepInfra's card has only added models — and it lands on the exact checkpoint that anchors third-party comparisons. Either way, if gpt-oss-120b or Llama 3.3 70B is load-bearing in your stack, this page's gpt-oss rows have a 21-day shelf life from publication; our Groq price card and DeepInfra price card both still serve those checkpoints after the cutoff.

GPU instances: the escape hatch, priced

Novita's GPU cloud rents by the hour alongside the serverless API: RTX 4090 24GB at $0.33/hr on-demand ($0.17 spot), RTX 5090 32GB at $0.56/hr, RTX 6000 Ada 48GB at $0.77/hr ($0.39 spot), and H100 SXM 80GB at $3.39/hr on-demand ($1.70 spot). For comparison, our Fireworks and Cerebras price cards list reserved H100/H200 capacity at $8/hr on the enterprise hosts — Novita's H100 rate is among the cheapest single-GPU H100 rates DeAI has verified, though the comparison is not apples-to-apples: Novita rents bare instances while Fireworks' rate is a managed dedicated deployment. For steady workloads on models Novita retires, a 4090 running an 8B–32B open model via vLLM is the classic self-host fallback.

How to choose

• Cache-heavy agent loops on DeepSeek or MiMo checkpoints: Novita's $0.006/M cache-read on DeepSeek V4.1 Flash and $0.0028/M on MiMo V2.6 Flash are the deepest verified discounts in the field — rerun your own blend before concluding.
• Cheapest per-token floor on a current frontier model: DeepInfra still wins (V4 Flash 0731 at $0.06/$0.18 vs Novita's $0.44/$1.32 on the same checkpoint).
• Stability over latency: Novita's 91% Endpoint Accuracy and 0.0% error rate on gpt-oss-120b are strong for a commodity fleet; pay fast-silicon prices only if someone watches the stream.
• Batch, offline, eval sweeps: the 50%-off Batch API with a 24-hour window is straightforward OpenAI-compatible volume discounting.
• Confidentiality by construction: none here — Novita's posture is a policy document with an explicit carve-out for customer content, not an attestation.

Switching to Novita is a base-URL swap from any OpenAI-shaped client:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.novita.ai/openai",
    api_key=os.environ["NOVITA_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek/deepseek-v4.1-flash-1017",  # exact slug varies by provider
    messages=[{"role": "user", "content": "ping"}],
)

Model slugs are namespaced (deepseek/deepseek-v4.1-flash-1017, moonshotai/kimi-k3) and are not portable across providers; quantization and sampling defaults differ per endpoint and show up in eval scores, not the price table. A cheaper model that fails your evals is not cheaper after retries.

Current state (November 2026)

Novita is the mid-pack price-setter with the deepest cache discounts DeAI has verified, a strong accuracy/error-rate profile on third-party measurement, and the weakest public-company footing of any host in this series — no funding disclosure, no infrastructure transparency, a non-resolving Trust Center link. The November 26 deprecation is the near-term test: shrinking away from the comparison checkpoint (gpt-oss-120b) while the Chinese frontier rows (DeepSeek V4.1, Kimi K3.1, GLM 5.3) expand is a clear strategic bet on where its demand actually sits. DeAI's Price Index re-shoots this spread weekly precisely because rate cards this churn-prone go stale fast. All prices here are dated 2026-11-05 and will move.

FAQ

How much does Novita AI cost per million tokens?

As of 2026-11-05, Novita's serverless rate card spans $0.02/$0.05 per million input/output tokens on Llama 3.1 8B up to $4.00/$20.00 on Kimi K3.1 Pro Highspeed. Current-frontier rows: DeepSeek V4.1 Flash $0.30/$1.20 (cache read $0.006), DeepSeek V4 Pro 0813 $1.32/$3.96, Kimi K3 $3.00/$15.00, GLM 5.3 Flash $0.15/$0.50.

Is Novita AI cheaper than DeepInfra or Groq?

On gpt-oss-120b, Artificial Analysis's blended (7:2:1) ranking puts Novita at $0.07 per million tokens — third cheapest of 16 providers, behind CoreWeave ($0.04) and DeepInfra ($0.05). Novita's input rate on that checkpoint ($0.05/M) matches DeepInfra's, but its output rate ($0.25/M) is higher. Rates move weekly; compare on the day you buy.

What are Novita AI's rate limits?

Novita tiers accounts T1 through T5 by top-ups over the trailing three calendar months (T1 under $50/month, T5 at least $10,000 in one month), with per-model RPM and TPM limits per tier. The numeric tier table is available only in the console. Rate-limit rejections (HTTP 429) are not billed; requests disconnected after inference starts (HTTP 499) are billed.

Does Novita AI train on your data?

Novita's privacy policy (effective May 13, 2026) states that personal information will not be used for model training, but the policy explicitly excludes content processed on behalf of customers while providing services — that data is governed by customer agreements. That is a self-published policy posture, not an independent verification.

Which Novita models are being retired in November 2026?

Thirteen models retire at 16:00 UTC on November 26, 2026, with no replacements: gpt-oss-120b, Meta's Llama 3.1 8B and Llama 3.3 70B, Google's Gemma 3 27B and both Gemma 4 open rows, NVIDIA's Nemotron 3 Nano 30B, DeepSeek OCR 2, AutoGLM Phone 9B, the BGE-M3 reranker, and the Sao10K Stheno fine-tune, among others.