Gemini 3.8 Flash Doubles in Price in 2027: Cut Tokens Now

Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash double in price on January 1, 2027: input goes from $0.75 to $1.50 per million tokens and output from $3.75 to $7.50. You cannot negotiate the rate, but you can send fewer tokens. Encoding structured input as TOON offsets part of the increase.

What exactly changes in Gemini Flash pricing?

Google's Gemini API pricing page (checked October 1, 2026) lists a dated schedule for gemini-3.8-flash: input is "$0.75 through December 31, 2026. $1.50 starting January 1, 2027." The same schedule applies to gemini-3.7-flash and gemini-3.6-flash. All prices are per million tokens, standard tier.

Every line doubles, so no billing mode escapes it. For comparison, the same page lists gemini-3.5-flash at $1.50 input and $9.00 output with no dated change, and gemini-3.1-flash-lite at $0.25 input and $1.50 output for text. After January 1, the newer Flash models cost the same as 3.5 Flash for input and less for output.

How much will the Gemini Flash increase cost you?

Since input and output both double, your Flash bill doubles too, unless usage changes. Take an example workload of 500 million input tokens and 50 million output tokens per month on gemini-3.8-flash:

The TOON row assumes 60% of input tokens are structured data (tool results, retrieved rows, catalogs) and that TOON shrinks that part by 40%. That removes 120 million input tokens, saving $180 a month at the 2027 rate, or about a third of the $562.50 increase. Plug in your own split. If your prompts are mostly prose, the offset is small. If they are mostly tables, it is larger. Our LLM cost calculator guide walks through the same math.

How much does TOON save on Gemini input?

The official TOON benchmark (spec v4.1, 5,856 LLM calls) reports 42.6% fewer tokens than JSON, with 72.2% accuracy vs 71.4% for JSON (retrieval-accuracy results). Savings depend on data shape: 58.7% on flat tables and 32.7% on mixed structures (token-efficiency results). Those counts use OpenAI's o200k_base tokenizer. Gemini tokenizes differently, and the benchmark notes that relative differences "hold directionally." Measure on Gemini's own token counter before you budget.

Here is a small product list, as pretty JSON and as TOON (4.1.1 encoder output):

{
  "products": [
    {
      "sku": "SKU-1042",
      "name": "Trail Running Shoe",
      "price": 129.99,
      "stock": 42,
      "rating": 4.6
    },
    {
      "sku": "SKU-1043",
      "name": "Merino Base Layer",
      "price": 79.5,
      "stock": 118,
      "rating": 4.8
    },
    {
      "sku": "SKU-1044",
      "name": "Packable Rain Shell",
      "price": 189,
      "stock": 7,
      "rating": 4.4
    }
  ]
}

products[3]{sku,name,price,stock,rating}:
  SKU-1042,Trail Running Shoe,129.99,42,4.6
  SKU-1043,Merino Base Layer,79.5,118,4.8
  SKU-1044,Packable Rain Shell,189,7,4.4

With o200k_base we counted 153 tokens for the pretty JSON, 92 for minified JSON and 73 for TOON. Even against minified JSON, the TOON version is about a fifth smaller, and the gap grows with more rows.

Does Gemini read TOON accurately?

gemini-3.6-flash is one of the four models in the current benchmark. It scored 69.3% with TOON and 68.4% with JSON across 244 questions (source). The intervals (about ±5.8 points) overlap, so the fair reading is "no accuracy loss," not "TOON is more accurate." Minified JSON scored lower for this model, at 63.5%. Our TOON with Google Gemini guide covers SDK setup and prompt wording.

What else cuts the Gemini bill before January?

• Batch API. Batch is 50% of standard on every listed model. Moving async jobs to batch roughly cancels the increase for those jobs.
• Context caching. Cached input will cost $0.15 per million, a tenth of the new input rate. Caching and TOON stack: a smaller cached block is cheaper to write and to re-read. See TOON for prompt caching.
• Route simple tasks to Flash-Lite. gemini-3.1-flash-lite is listed at $0.25 input and $1.50 output.
• Trim fields before encoding. A column the model never uses costs tokens in any format.

Keep TOON's limits in mind. It shrinks input, not output, so output-heavy workloads such as long-form generation gain little. On deeply nested or semi-uniform data, minified JSON can be smaller. If your Flash calls are mostly about reading tables, rows and tool results, now is a good time to switch, while the cheaper rate still applies and you can measure the difference.

Frequently Asked Questions