Official API vs Gateway Pricing: OpenRouter, Novita and Closed Models

Buying tokens from OpenAI, Anthropic or Google is one price path. Buying the same open-weight weights from a hosting gateway (Novita, OpenRouter, Together, Groq) is another. The sticker rates look similar in screenshots and are not the same product: official APIs sell the maker’s closed model; gateways usually sell hosting for open weights, or they pass through a maker’s list price and take a platform fee somewhere else. This page separates those paths with numbers we can source, and labels everything else as Not published rather than guessing.

Disclosure: links to Novita AI are affiliate links — we may earn a commission at no extra cost to you. It never affects order: every table here is sorted by price only. Read our disclosure.

What “official” and “gateway” mean here

We do not treat a gateway row as a discount on a closed flagship. If Novita hosts gpt-oss-120b, that is an open-weight model price — not “OpenAI GPT-6 for $0.05”. Closed GPT-6 / Claude / Gemini rates below are official only.

Gateway prices we track (Novita)

Four of these rows are stored in our pricing data as price_type: gateway (GPT OSS 120B, Llama 4 Maverick, DeepSeek R1 0528, Qwen3.8 Max), verified against Novita’s pricing page on 2026-10-26. DeepSeek V4 Pro is stored as DeepSeek’s official peak rate; Novita lists the same SKU (V4 Pro 0813) at that peak list price, so we include it here for the gateway path and unpack the official off-peak gap in the next section. Per-model batch rates are Not published on Novita for these SKUs, so the calculator shows that instead of inventing a 50% cut.

Workload for the monthly column: 10M tokens a month, 8M input / 2M output, no caching, standard (non-batch) rates. Sorted by monthly cost.

Affiliate link for the host (does not change the sort order): Novita AI. Provider overview: /providers/novita/.

Same workload on official closed APIs

Same mix (8M in / 2M out per 10M tokens). Official list prices only — sorted by monthly cost.

At this mix, the cheapest gateway row in our dataset ($0.90 on GPT OSS 120B via Novita) undercuts the cheapest official closed tier ($1.80 on Haiku 5.5 / GPT-6 Luna). That is a quality and capability comparison, not a like-for-like discount on Sonnet or Sol.

The clean price-gap case: DeepSeek V4 Pro

DeepSeek publishes an official API price with peak / off-peak tiers. Novita hosts the same model version (DeepSeek V4 Pro 0813) at the peak official rate. Details and sources are on the DeepSeek V4 Pro model page and in our DeepSeek vs closed models article.

Peak hours on DeepSeek’s docs: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays; all other hours are off-peak at half price. Off-peak direct is the only row that beats Novita on this SKU; at peak, the rates match.

OpenRouter vs OpenAI API price (what we can and cannot say)

People search “OpenRouter vs OpenAI API price” expecting a side-by-side of GPT-6 on both. On CostMyTokens we do not store OpenRouter per-model rows in pricing.json, so we will not invent an OpenRouter column for GPT-6 Astra / Sol / Luna.

What OpenRouter’s own FAQ states (fetched 2026-10-26 from openrouter.ai/docs/faq):

• Inference is billed at the underlying provider’s list price — “no markup on inference pricing”.
• OpenRouter charges a 5.5% fee (minimum $0.80) when you purchase credits, and 5% for crypto. BYOK has a separate allowance and fee above that allowance.

So for a closed OpenAI model, the token rate on OpenRouter is designed to match OpenAI’s list price; the extra cost shows up as a credit-purchase fee (and routing / uptime trade-offs), not as a cheaper $/1M on the model card. That is the opposite of Novita’s open-weight hosting rows above, where the gateway is the price you pay per token for that host.

We also do not have Together AI or Groq prices in our dataset. Their per-model rates are Not published here. OpenRouter, Together and Groq have no public affiliate / referral program we could verify on official pages (see our affiliate notes); Novita does.

When to use official vs gateway

Prefer official APIs when:

• You need a specific closed model (GPT-6.x, Claude 5.5, Gemini 4) with maker SLAs, enterprise terms, or features only on the first-party API.
• You already get volume discounts, committed use, or org billing with the maker.
• You care about first-party batch / cache / long-context rules exactly as documented (e.g. OpenAI’s ~272K long-context tier, Anthropic Haiku’s 100K cliff).

Prefer a gateway when:

• You want open-weight models (Llama, GPT-OSS, Qwen, DeepSeek hosted copies) without running GPUs yourself.
• One API key across many open models matters more than talking to each host separately.
• For DeepSeek V4 Pro specifically, you are fine paying peak rates for simpler billing — or you will call DeepSeek direct off-peak to cut the bill in half.

Prefer OpenRouter-style routers when:

• You want failover across providers and a single OpenAI-compatible endpoint, and you accept credit fees on top of list prices.
• Compare the fee on credits plus list price against going direct — do not assume the model card $/1M is a discount vs OpenAI.

Fine print that moves the numbers

• GPT-6.1 Sol: prompts above ~272K input tokens enter a long-context tier (~2× input, ~1.5× output). Our calculator uses standard-tier rates.
• Claude Haiku 5.5: prompts over 100K tokens are billed entirely at the higher tier ($0.50 / $2.50). Calculator uses the base tier.
• Gemini 4 Argon: introductory $2 / $10 through 2026-11-30; from 2026-12-01 rates double to $4 / $20.
• Novita batch: introductory 50% batch discount exists on Novita for supported models, but per-model batch figures for the Novita rows we track are Not published in our data — we do not fill them in.
• Quality ≠ price rank: a $0.90/month open-weight host only wins if the model finishes your task. Benchmark on your own prompts.

Run your own mix

Fixed 8M/2M tables are a starting point. Plug real input, output and cache share into the calculator:

• Open calculator
• Deep links: GPT OSS 120B · Llama 4 Maverick · DeepSeek V4 Pro · GPT-6.1 Sol · Claude Sonnet 5.5

Related

• Cheapest LLM API in 2026 (ranked)
• DeepSeek V4 Pro vs closed models
• DeepSeek official API prices (all models)
• Novita AI gateway prices
• OpenAI vs Anthropic vs Google pricing
• Affiliate / sponsored disclosure

Disclosure: links to Novita AI are affiliate links — we may earn a commission at no extra cost to you. It never affects order: every table here is sorted by price only. Read our disclosure.