Gemma 4 & EmbeddingGemma 2 Pricing
Both models are free. What you pay for is the computing that runs them: your own hardware, a rented graphics card, or a hosted provider's per-token rate. Here is every price we could verify, and a calculator for your own numbers.
What is free, and what is not
Free to download, use commercially, fine-tune and redistribute under Apache 2.0.
Also Apache 2.0. No seat fees, revenue thresholds or usage reporting.
No per-token fee. You pay for the hardware and the electricity.
Per million tokens for the 26B A4B; $0.22 per million out.
Running a model yourself is the free route if you already own suitable hardware: the 12B needs about 6.7 GB of memory at 4-bit. See the VRAM guide and the install guide.
Estimate your monthly bill
Enter how many tokens you expect to send and receive each month. A token is roughly three-quarters of an English word.
Uses the list prices in the tables below. It ignores cache discounts, volume deals and taxes, and the rental comparison assumes one card running all month (730 hours) with no staff or storage costs. Treat it as a rough guide.
Every hosted price we could verify
Per million tokens. The same model costs very different amounts depending on who serves it.
Gemma 4 31B
| Provider | Input | Output | Note |
|---|---|---|---|
| DeepInfra Turbo | $0.09 | $0.34 | Lowest price |
| CoreWeave | $0.10 | $0.34 | |
| Venice | $0.12 | $0.36 | |
| Chutes | $0.12 | $0.37 | |
| Crusoe | $0.14 | $0.40 | |
| Friendli | $0.14 | $0.40 | |
| NovitaAI | $0.14 | $0.40 | |
| Parasail | $0.15 | $0.40 | |
| DeepInfra | $0.20 | $0.40 | Standard tier |
| io.net | $0.36 | $1.09 | |
| SambaNova | $0.38 | $1.15 | Highest throughput listed |
| Together AI | $0.39 | $0.97 | Direct, serverless |
| ModelRun | $0.75 | $1.00 | Lowest latency listed |
| SiliconFlow | $0.75 | $1.00 |
Gemma 4 26B A4B
| Provider | Input | Output | Note |
|---|---|---|---|
| Darkbloom | $0.042 | $0.22 | Lowest price |
| io.net | $0.052 | $0.23 | Discounted rate |
| DekaLLM | $0.06 | $0.33 | |
| DeepInfra | $0.07 | $0.34 | |
| Makora | $0.08 | $0.32 | |
| NextBit | $0.09 | $0.30 | |
| CoreWeave | $0.10 | $0.30 | |
| Cloudflare | $0.10 | $0.30 | |
| Venice | $0.13 | $0.40 | |
| Parasail | $0.13 | $0.40 | |
| NovitaAI | $0.13 | $0.40 | |
| SiliconFlow | $0.14 | $0.40 | |
| Google Vertex AI | $0.15 | $0.60 | Google's own cloud |
An eight-fold spread
For the 31B, input runs from $0.09 to $0.75 per million tokens for the same weights. It pays to compare.
Price is not everything
Providers differ in speed, uptime, context limit and whether they serve a compressed build. The cheapest may not suit a production app.
The smaller sizes
Hosted 12B, E4B and E2B are less widely offered and their prices could not be confirmed for this page. They are the sizes easiest to run yourself.
From OpenRouter's 31B and 26B A4B listings and Together AI's pricing page, October 7, 2026. Output usually costs three to five times as much as input. How to call these providers is on the API page.
What Google charges
| Route | Models | Price |
|---|---|---|
| Google AI Studio, in the browser | 31B, 26B A4B | Free to try, within usage limits |
| Gemini API | gemma-4-31b-it, gemma-4-26b-a4b-it | No price listed on Google's pricing page; access is rate-limited |
| Google Vertex AI | 26B A4B | $0.15 in, $0.60 out per million tokens, as listed on OpenRouter |
| EmbeddingGemma 2 | - | No hosted price yet; Google says a Model Garden listing is coming soon |
Renting a graphics card instead
If you do not own suitable hardware, you can rent it by the hour. You pay for the time, whether or not the model is busy.
| Graphics card | Memory | Per hour | Per month, always on | Runs comfortably |
|---|---|---|---|---|
| NVIDIA L4 | 24 GB | $0.49 | $358 | E2B, E4B, 12B at 4-bit |
| RTX A6000 | 48 GB | $0.53 | $387 | 26B A4B, 31B at 4-bit |
| RTX 4090 | 24 GB | $0.74 | $540 | 12B; 26B A4B and 31B at 4-bit with a short context |
| RTX 5090 | 32 GB | $0.99 | $723 | 26B A4B, 31B at 4-bit |
| L40S | 48 GB | $1.09 | $796 | 26B A4B, 31B at 4-bit, with room for context |
| A100 | 80 GB | $1.59 | $1,161 | 31B at 8-bit |
| H100 PCIe | 80 GB | $2.89 | $2,110 | 31B at full precision |
| H100 SXM | 80 GB | $3.49 | $2,548 | 31B at full precision, production |
Hosted is cheaper when
Your volume is low or comes in bursts. At a few cents per million tokens, light use costs far less than a card left switched on.
Self-hosting wins when
You have steady, high volume, already own the hardware, need the smaller models, or cannot send data to a third party.
Count the extras
Storage for the weights, monitoring, and the time of whoever keeps it running. A rented card is rarely the whole bill.
On-demand prices from RunPod's pricing page, October 7, 2026; monthly figures are the hourly price times 730. "Runs comfortably" is this site's guidance, based on Google's memory figures. Setting one up is covered under self-hosting; the 26B A4B is the efficient pick for serving - see the model comparison.
EmbeddingGemma 2 pricing
EmbeddingGemma 2 is free and small, so the model itself costs almost nothing to run. With embeddings, the cost that grows is storage: every document you index becomes a vector you have to keep.
Apache 2.0, including commercial use.
When you run it yourself. No hosted price exists yet.
Or 378 MB for text only, on Ollama.
Google's figure for the full model on a Pixel 11 Pro, with quantisation.
Vector storage calculator
Why this matters for cost
Paid embedding services charge for every token you embed, and again whenever you re-index. With an open model you can re-embed as often as you like for the price of the computing.
It runs on what you have
A laptop, a phone or a small server is enough. You do not need a graphics card to get started.
Budget for re-embedding
Vectors from version 1 and version 2 are not interchangeable. If you upgrade, plan the computing time to index everything again.
Storage figures from Google's EmbeddingGemma 2 developer guide: about 1.5 GB for a million 768-number vectors and 250 MB at 128. The calculator scales those; a vector database adds its own overhead. Quality by vector length is on the limitations page; setup is in how to use and on the Ollama page.
Is Gemma 4 free?
The model is: the weights are Apache 2.0 and free to download and use commercially. Running it costs something - your own hardware and electricity, or a hosted provider's per-token rate from $0.042 per million input tokens.
Is EmbeddingGemma 2 free?
Yes. It is Apache 2.0 and there is no hosted, paid version from Google yet. You run it yourself, which needs very little hardware.
What is the cheapest way to use Gemma 4?
Your own machine, if a model fits: see the download hub. Otherwise the 26B A4B at the lowest hosted rate, $0.042 in and $0.22 out per million tokens.
Is there a free tier?
Google AI Studio is free to try in the browser, within limits. Google's API pricing page lists no price for Gemma models. This page found no free hosted tier on OpenRouter for the 31B or 26B A4B.
Why is the 26B cheaper than the 31B?
The 26B A4B activates only about 3.8B of its parameters for each token, so it takes less computing to serve. See features.
Should I self-host to save money?
Usually not for cost alone. A rented card costs the same whether busy or idle, and light use is far cheaper hosted. Self-host for privacy, offline use, the smaller models or steady high volume.
Are thinking tokens charged?
Yes, as output. Turn thinking down for simple tasks to save money.
How current are these prices?
They were recorded on October 7, 2026 from the providers' own listings. Prices move often; follow the source links before you commit. Bigger hosted models cost more - see Gemma 4 vs Qwen3.8-Max and Gemma 4 vs Claude Code.
The cheapest token is the one you run yourself
Both models are free to download and run on hardware you already own.