Gemini 4: What We Actually Know About Google's Next Frontier Model

September 30, 2026 update. On September 30, Google announced Gemini 4 Argon; see our analysis of Argon’s capabilities and access. Access currently goes to selected Fairwind Program participants and invited testers, and we are not announcing Argon access in GPTunneL. The August 16 article below is preserved as a record of what Google said about Gemini 4 in July, before the announcement. Statements about release dates, benchmarks, prices and model IDs in the historical sections refer to that date. The code advice, FAQ and bottom line have been updated.

As of August 16, Gemini 4 did not exist yet: no API, no benchmarks, no price, no gemini-4 string in any model catalog. What existed was two confirmed statements from Google and a thick layer of analyst inference on top. Below: what Google actually said, what of it was verifiable, and how to structure your code so that switching to a new model costs one config change instead of an integration rewrite.

What Google actually said

On July 21, 2026, Google's blog stated it had started its "most ambitious pre-training run yet" — for Gemini 4. That was the first time the company named the model itself, rather than a leak doing it.

On July 23, 2026, on Alphabet's Q2 earnings call, Sundar Pichai repeated it on the record and added specifics:

• "For the next generation of frontier, you're going to need much larger base models. We are now training Gemini 4" — the next generation requires a base model significantly larger than Gemini 3 Pro.
• "We want to compete at the frontier level of where the frontier will be when Gemini 4 comes out" — they are aiming at where competitors will be at launch, not where they are now.
• "I am very excited by the progress I'm seeing internally on Gemini 4" — the model is already being demoed internally.

Pichai also spelled out the release rhythm: "picking up pace and releasing models almost at a monthly cadence is part of our road map as we are building Gemini 4 as well." Read: the Flash line ships almost monthly and holds the market, while the frontier model trains in the background on its own long cycle.

The stated priorities are explicit: coding and agentic coding. Exactly where Google has been trailing in 2026.

What Google did not say

This is where care matters: nearly everything circulating about Gemini 4 is analyst inference, not a company statement.

The "November–December 2026" estimate is extrapolation from Google's past release cadence, not a commitment. Given that Gemini 3.5 Pro was promised at I/O for June 2026 and had not shipped by mid-August, treat any calendar prediction as a hypothesis.

Why Gemini 4 is needed right now

Context beats rumor. By August 2026, three things had piled up for Google at once.

A stuck Pro. Gemini 3.5 Pro was announced at I/O targeting June. Per Bloomberg reporting, the model is months late: its coding performance fell short of internal expectations against OpenAI and Anthropic models. On July 22 Google shipped three non-frontier 3.x models — and said nothing about Pro.

A leadership shake-up. Demis Hassabis moved from CEO of Google DeepMind to chair of DeepMind and chief scientist of Alphabet. Gemini 4 development is now led by Koray Kavukcuoglu, promoted to SVP and reporting directly to Pichai. On the same day, two of Gemini's co-technical leads left Google — Jeff Dean (27 years at the company) and Oriol Vinyals, both to the startup Discovery Loop.

Flash carrying the load. Gemini 3.7 Flash shipped on August 13, 2026, and it is not a cosmetic bump. DeepSWE v1.1 — 65.3% versus 49.0% for 3.6 Flash. AutomationBench — 30.4% versus 17.0%. WebDev Arena Elo — 1588 versus 1538. FrontierCode 1.1 Main — 43.6% versus 34.4%. On the Artificial Analysis Intelligence Index it scores 56: ahead of Claude Sonnet 5 (55) and its own 3.6 Flash (52), slightly behind GPT-5.6 Terra and Muse Spark 1.2 (57 each).

So the strategy is two-track: cheap Flash absorbs the volume workload and keeps developers, while the heavy base for Gemini 4 trains in the background.

What this means for your code

One practical conclusion: don't hardwire a model name. At an "almost monthly" cadence, a production model lives for months, not years, and Gemini 4 will be one more row in that list.

A base request through GPTunneL is OpenAI-compatible, so switching models is a single field. The model IDs in these examples were checked against the catalog on September 30. Before running the code, confirm them in the current catalog:

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": "Hello!"}]
  }'

In code that means the model name lives in config and nowhere else. Not inside a function body, not in three places, not in a constant next to the prompt.

// models.js — the only place a model name is ever written
export const MODELS = {
  fast: process.env.MODEL_FAST ?? "gemini-3.8-flash",
  smart: process.env.MODEL_SMART ?? "gemini-3.1-pro",
};

// client.js
import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.GPTUNNEL_API_KEY,
  baseURL: "https://gptunnel.ru/v1",
});

export async function ask(tier, messages) {
  const res = await client.chat.completions.create({
    model: MODELS[tier],
    messages,
    temperature: 0.2,
  });
  return res.choices[0].message.content;
}

When the model you need appears in the catalog, migration comes down to a new MODEL_SMART value in your environment plus a test run, with no code deploy. Take the identifier from the catalog at that point: we are not announcing Gemini 4 Argon access in GPTunneL, so no ID for it is given here.

Don't switch blind: a minimal eval

A monthly cadence breaks the habit of "new model shipped, ship it." A new model can be pricier, chattier, or call tools differently — and what was a benchmark win becomes a regression in cost and latency in production.

Keep 20–50 of your own real requests with expected answers, and run every candidate model through them:

import { MODELS } from "./models.js";
import cases from "./eval-cases.json" with { type: "json" };

const CANDIDATES = ["gemini-3.8-flash", "gemini-3.1-pro"];

for (const model of CANDIDATES) {
  let ok = 0, inTok = 0, outTok = 0;
  const t0 = performance.now();

  for (const c of cases) {
    const res = await client.chat.completions.create({
      model,
      messages: [{ role: "user", content: c.prompt }],
      temperature: 0,
    });
    if (res.choices[0].message.content.includes(c.expect)) ok++;
    inTok += res.usage.prompt_tokens;
    outTok += res.usage.completion_tokens;
  }

  console.log(model, {
    accuracy: (ok / cases.length).toFixed(2),
    inTok,
    outTok,
    msPerCase: Math.round((performance.now() - t0) / cases.length),
  });
}

Three columns — accuracy, tokens, latency — answer "should we migrate" better than any press release. The outTok column especially: Gemini 3.6 Flash's headline improvement wasn't intelligence but a 17% cut in output tokens, and that shows up on the invoice more directly than on a benchmark.

What will break when you move to Gemini 4

This list comes from past migrations inside the Gemini line, not from rumors.

• Verbosity. A new model can answer noticeably longer on the same prompt. Watch usage.completion_tokens, not just quality, and keep max_tokens set.
• Tool call shape. The order and number of tool calls in an agent loop shift between versions. If your parser assumes "exactly one call per step," it will break.
• Structured output. Schemas the old model filled reliably may come back with different defaults. Validate against the schema and retry on invalid JSON instead of trusting it.
• Latency and timeouts. A larger base usually means more time to first token. Use streaming and a per-request budget timeout, not a flat "30 seconds for everything."
• Price. GPTunneL pricing for Argon has not been announced. Before switching, estimate the cost of your tasks using our price list once the model appears in the catalog. You can prepare routing between models for different task complexity in advance.

Full request, streaming, and limit specs are in the docs and at docs.gptunnel.ru.

What to use today

Gemini 4 Argon is currently available to selected Fairwind Program participants and invited testers, and we are not announcing Argon access in GPTunneL. For current work, use models from the catalog: in the Gemini family, for example, Gemini 3.8 Flash and Gemini 3.1 Pro. See the current lineup and rates on the Gemini page and the pricing page.

Models run right in your browser, with local payment methods and no subscription: you pay per token. GPT, Claude, Grok and the rest sit behind the same API, so the eval above runs across every candidate on one key.

Gemini 4 FAQ

When will Gemini 4 be released? Google announced Gemini 4 Argon on September 30, 2026, but broad access has not opened yet. Selected Fairwind Program participants and invited testers get the model first; Google names paid API customers and Google AI Ultra subscribers as next, with no date. In July, Google had only confirmed that training was underway; see our Argon analysis for the announcement details.

How does Gemini 4 Argon differ from Gemini 3? According to Google, Argon is built for long multi-step work: complex software engineering, legal and finance documents, and cyber defense. The output limit grew from 64,000 to 1 million tokens. In July, Google only described a significantly larger base model than Gemini 3 Pro. Argon's benchmarks and capabilities are covered in a separate article.

Can I try Gemini 4 Argon through an API already? Access is available to selected Fairwind Program participants and invited testers. For everyone else, Google is opening access in phases: paid API customers and Google AI Ultra subscribers are named next, and no date has been announced. Fairwind rules prohibit sharing or reselling access. We are not announcing Argon access in GPTunneL.

Which Gemini should I use while Argon is unavailable? A model that fits your task from the current catalog: the family includes, for example, Gemini 3.8 Flash and Gemini 3.1 Pro. The lineup and prices change, so check the Gemini page and pricing. Base the decision to switch on your own eval set from the section above.

How do I avoid rewriting code when models change? Keep the model name in an environment variable, use an OpenAI-compatible client, and maintain your own eval set. When the model you need is in the catalog, migration is a config edit and a test run.

Bottom line

In August there were exactly two verifiable facts about Gemini 4: training was underway, and the base would be significantly larger. On September 30, Google announced Gemini 4 Argon, but broad access has not opened and no date has been given. The August advice still holds: move the model name into config, build an eval set from your real requests, and set up routing between a cheap and an expensive model. Then switching models is a line in .env plus a test run, not a migration project.

You can start now: open Gemini models in GPTunneL, run your own scenario on the available options, and read our Argon analysis for what is known about the new model.