Gemini 3.8 Flash: what changed and how to call it via API

Google shipped Gemini 3.8 Flash on September 2, 2026 — the third Flash release in six weeks. The model is already in the GPTunneL catalog under the id gemini-3.8-flash. In short: not a single API parameter changed, the limits are the same, and the gains are in long agentic tasks and professional domains. If your code targets 3.7 Flash, migration is a one-line change of the model name.

What shipped

From the model card:

The table matches 3.7 Flash line for line. Google says outright that speed and price in its own API are unchanged; what changed is behavior: the model takes more reasoning steps and calls tools repeatedly before returning an answer.

Minimal request

The GPTunneL gateway is OpenAI-compatible; the key goes in Authorization without the Bearer prefix:

curl https://gptunnel.ru/v1/chat/completions \
  -H "Authorization: <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.8-flash",
    "messages": [{"role": "user", "content": "One word: the capital of France?"}]
  }'

The response I got before publishing:

{
  "model": "gemini-3.8-flash",
  "choices": [{"message": {"role": "assistant", "content": "Paris"}, "finish_reason": "stop"}],
  "usage": {
    "prompt_tokens": 11,
    "completion_tokens": 108,
    "total_tokens": 119,
    "prompt_cost": 0.0033,
    "completion_cost": 0.162,
    "total_cost": 0.1653,
    "prompt_tokens_details": {"cached_tokens": 0}
  }
}

One word — 108 output tokens. Thinking is billed as output, and by Google's design 3.8 Flash does more of it. Track spend by total_cost in usage, not by the length of the visible answer.

Through the gateway, reasoning depth is set with reasoning_effort taking low, medium, high; Google's thinking_level is silently ignored. Per-level measurements on 3.7 Flash are in the previous article; the mechanics on 3.8 are the same.

What moved on benchmarks

Google compares 3.8 Flash not with its predecessor but with competitors' flagships:

On DeepSWE v1.1 — long software engineering tasks from the ticket to a finished patch — Google shows a result above 70% at a cost well below the large models; the announcement gives no exact figure. Third-party reviews quote Terminal-Bench 2.1 at 90.8% versus 81.6% for 3.7 Flash.

Read this soberly. The gain over 3.7 Flash is 1–2 percentage points on professional agentic tests, not a jump like DeepSWE 49% → 65% a month ago. Beating Opus 5 and Sol happens on narrow domain benchmarks in finance and law. On general agentic tasks like OSWorld the flagships, according to third-party summaries, stay ahead. If you run a multi-step agent with tools in a specialized domain, it is worth a re-run. If it is a classifier or a summarizer, you will not see a difference.

3.8 Flash Cyber

Alongside the main model Google released a cybersecurity variant: vulnerability discovery across twenty languages with a success rate above 70% and automated patching at 47.2% pass@1 on CWE-Bench. It did not make it into the public API: access goes only through the Fairwind program for government bodies, infrastructure operators and maintainers. It is not available through GPTunneL and will not be until Google opens the model publicly.

What to check before switching

• max_tokens. The cap eats thinking before the answer: on 3.7 Flash, max_tokens: 16 returned empty content with finish_reason: "length". 3.8 reasons more — leave an even larger margin.
• Timeouts. Extra steps and repeated tool calls take time. A 30-second client timeout will cut off normal answers on high; use 180 seconds and retries with exponential backoff.
• The bill. More reasoning at the same per-token price means more tokens per task. Compare total_cost on your own requests before and after, not the catalog rates.
• Structured outputs and streaming work as before: response_format with a schema and stream: true. The gateway still sends the keepalive comments : HELLO and : PROCESSING before the first chunk — the parser has to skip lines without data: .

What it costs

In GPTunneL, Gemini 3.8 Flash is billed per token: $3.00 per 1M input tokens and $15.00 per 1M output tokens, with a 1M-token context. That is twice the price of 3.7 Flash, which stays in the catalog at $1.50 and $7.50 — so for tasks where 3.8 shows no gain, keeping the older model makes sense. Google itself runs an introductory rate of $0.75 and $3.75 per 1M tokens through December 31, 2026, then $1.50 and $7.50 from January. Current numbers for every model live on the pricing page. No subscription, pay-as-you-go by the token, local payment methods.

Try it yourself

Swap gemini-3.8-flash for gemini-3.7-flash in your code with base_url=https://gptunnel.ru/v1 and run the same agentic loop. Compare two numbers: total_cost and the share of tasks carried to completion. Or open Gemini 3.8 Flash in GPTunneL and test it on your own task in the chat.