As of June 2026, Gemma 4 31B costs approximately $0.12–$0.15 per 1M input tokens and $0.35 per 1M output tokens. The Gemma 4 26B MoE variant is cheaper at $0.06 per 1M input and $0.30–$0.33 per 1M output. Prices differ across providers such as OpenRouter, DeepInfra, and Cloudflare, so comparing before committing can save meaningful money at scale.

Gemma 4 31B is a dense transformer with 30.7B parameters suited to tasks requiring deep reasoning. Gemma 4 26B MoE is a Mixture-of-Experts model that activates only 3.8B parameters per token, delivering near-31B benchmark quality at roughly half the inference cost and with lower latency. For most production workloads the MoE variant offers a better cost-to-performance ratio. On SWE-Bench Pro, Gemma 4 31B scores 73+, comparable to leading frontier models at a fraction of the price.

Input tokens are the text sent to the model — your prompt, system instructions, conversation history, and any documents in context. Output tokens are the generated reply the model returns. Output tokens are typically priced 2–3× higher than input tokens. For Gemma 4, output tokens cost roughly 2.5× more per million. Budgeting tip: a rough rule is 1 token ≈ 0.75 words, and 1,000 tokens ≈ a few paragraphs of prose.

Yes. Through providers such as Requesty and DeepInfra, prompt caching can reduce effective input costs by 30–80% when the same system prompt or document context is reused across many requests. This is especially valuable for RAG pipelines, coding assistants, and chatbots that include a large fixed system prompt in every API call.

The Gemma 4 model weights are free to download under an Apache 2.0 licence with no commercial restrictions. However, running the model yourself requires GPU hardware with real infrastructure costs. For self-hosted inference, effective cost typically lands between $0.001 and $0.005 per 1M tokens on consumer GPU hardware. The E2B variant is also available free via Google AI Studio for low-volume experimentation. Managed API access through providers is billed per token.

Gemma 4 is dramatically cheaper than closed-source frontier models. As of June 2026, GPT-5 input tokens cost around $2 per 1M and Claude Opus 4.6 input costs around $15 per 1M. Gemma 4 31B at $0.12–$0.15 per 1M input is 10–100× cheaper, making it highly attractive for cost-sensitive production workloads where its benchmark performance is sufficient for the task.

All Gemma 4 variants support a context window of up to 262K tokens (approximately 200,000 words). This makes it well suited for long-document summarisation, large codebase analysis, and retrieval-augmented generation. Be aware that using large context windows increases token costs proportionally — every token in the context counts toward your input bill on each request, so a 64K context window used for every call can double your bill versus a 4K sliding window.

As of mid-2026, DeepInfra and Cloudflare Workers AI typically offer the lowest blended per-token rates for Gemma 4, especially on high-throughput workloads. OpenRouter provides access to multiple underlying providers and routes automatically by cost or speed. Google AI Studio offers direct managed access with potentially different SLAs. Always compare current pricing on each provider’s page before committing, as rates fluctuate.