title: "Gemini 4 Is Confirmed: What Google Actually Said, What's Still Unknown, and What to Do While You Wait" slug: "gemini-4-confirmed-what-we-know-2026" excerpt: "Google confirmed Gemini 4 is in pre-training on July 21, 2026. Here's every confirmed fact, what's still unknown, which current Gemini model to use now, and how to prepare." category: "artificial-intelligence" tags: ["gemini 4", "google ai", "gemini models", "ai agents", "coding agents", "frontier models"]
Gemini 4 Is Confirmed: What Google Actually Said, What's Still Unknown, and What to Do While You Wait
Verdict: Google has confirmed that Gemini 4 is in pre-training as of July 2026, describing it as the company's "most ambitious pre-training run yet." CEO Sundar Pichai stated Gemini 4 will use a larger base model than Gemini 3 Pro and that coding and autonomous agents are the stated priorities for the generation. No release date, parameter count, benchmark score, context window, or price has been disclosed. If you are building with AI today, the smartest move is to master Gemini 3.6 Flash now and design your workflows to be model-agnostic so swapping in Gemini 4 later is a one-line change, not a rewrite.
Last verified: 2026-07-30
- Gemini 4 pre-training confirmed by Google on July 21, 2026, and by Pichai on the July 22 Q2 earnings call. (Google blog)
- Larger base model than Gemini 3 Pro — confirmed by Pichai on the earnings call. (TheStreet)
- Coding and agents are stated priorities — confirmed. (AIToolsReview)
- Alphabet raised 2026 capex to $195–$205 billion, up from $180–$190 billion. (Metirai)
- No release date, parameter count, context window, price, or benchmark — all unconfirmed.
What did Google actually confirm about Gemini 4?
Google confirmed exactly four things about Gemini 4, all traceable to its own words spoken between July 21 and July 23, 2026.
1. Pre-training has started. On July 21, 2026, Google published a blog post introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The final paragraph read: "We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress." (Google blog) That single line — buried in a smaller model's announcement — was the first time Google confirmed the name "Gemini 4" in its own words.
2. It will use a larger base model than Gemini 3 Pro. Two days later, on the July 22 Alphabet Q2 2026 earnings call, Pichai told analysts that Gemini 4 will require "a massive leap in base model scale" to "compete at the frontier level of where the frontier will be when Gemini 4 comes out." (TheStreet) Translation: Google is training toward a target that keeps moving while the model is still in pre-training.
3. Coding and autonomous agents are the stated priorities. Pichai specifically admitted Google "runs a bit behind" on agentic coding and is dedicating significant resources to close the gap against OpenAI's Codex and Anthropic's Claude for coding. (AIO Apex) The Gemini 3.6 Flash refresh already made its biggest gains in agentic evals — DeepSWE, MLE-Bench, and OSWorld-Verified — pointing toward Gemini 4 being positioned first as a stronger agentic and coding model.
4. Alphabet is spending record amounts on compute. The company raised its full-year 2026 capital expenditure forecast to $195–$205 billion, up from $180–$190 billion the prior quarter. (Metirai) Google Cloud revenue grew 82% year-over-year in Q2 to $14.4 billion, driven by enterprise AI adoption. The capex increase funds the infrastructure that will train and serve Gemini 4.
Everything else circulating — specific release windows, rumored parameter counts, benchmark comparisons to unreleased rivals — is analysis or extrapolation, not Google disclosure.
What has Google NOT confirmed about Gemini 4?
The gap between "pre-training started" and "model released" is where every rumor lives. Here is what is still unknown.
| Claim | Source type | Status |
|---|---|---|
| Pre-training for Gemini 4 has started | Official Google blog, July 21, 2026 | Confirmed |
| Larger base model than Gemini 3 Pro | Pichai on Q2 earnings call | Confirmed |
| Coding + agents are development priorities | Pichai on Q2 earnings call | Confirmed |
| Near-monthly Flash release cadence in interim | Pichai on Q2 earnings call | Confirmed |
| "Release in late 2026" | Analysts citing historical cadence | Speculative |
| "100 trillion parameters" | Social media, no source | False |
New gemini-4-* model ID in API |
Code miners checked model catalogs | False (no such ID exists as of July 22, 2026) |
| Specific benchmark scores | Not disclosed by Google | Unknown |
| Context window size | Not disclosed by Google | Unknown |
| Pricing | Not disclosed by Google | Unknown |
Based on Google's historical cadence — roughly seven months between Gemini 2.5 and Gemini 3, and a similar gap through the 3.x refreshes — independent outlets estimate a possible window of November or December 2026. (AIToolsReview) But that is an inference from past patterns, not a Google commitment. Google's own track record on dates is uneven: Gemini 3.5 Pro was promised for June 2026 and has still not shipped as of this writing. (9to5Google)
Why is Gemini 3.5 Pro delayed — and what does that mean for Gemini 4?
Gemini 3.5 Pro, originally announced at I/O in May 2026 for a June release, is months late. According to Bloomberg, Google delayed it because it "fell short of internal goals" — specifically in coding performance. The company updated training data in late June but the results were "disappointing." (9to5Google) (Technology.org) For a deeper guide on what to use in the meantime, see our Gemini 3.5 Pro delay: what to use instead.
This matters for two reasons.
First, it tells you where Gemini 4's bar is set. Pichai's framing — training to compete at "where the frontier will be when Gemini 4 comes out" — implicitly concedes that Google is training against a moving target. Claude Opus 5 shipped days before the earnings call. OpenAI's GPT-5.6 launched the same week. Chinese labs like Zhipu released GLM-5.2, which matches Anthropic's Opus 4.8 on coding benchmarks at a fifth of the cost. (Technology.org) Google cannot ship a model that merely matches today's frontier — it needs to beat the frontier that exists months from now.
Second, two model generations are now moving at once inside the same company. Gemini 3.5 Pro is still in partner testing. Gemini 4 is already training in the background. Google has said it plans near-monthly Flash releases in the interim, meaning the 3.x line continues while the next full generation builds. (AIToolsReview) For anyone building on Gemini, that means the model you deploy today will likely be superseded twice before Gemini 4 lands.
Which Gemini model should you use right now?
You do not need Gemini 4 to do real work today. Google shipped three highly capable models on July 21, 2026, and the 3.1 Pro is still available for frontier-grade reasoning. Here is what each is best for.
| Model | Best for | Price (per 1M tokens) | Key score |
|---|---|---|---|
| Gemini 3.6 Flash | Agentic workflows, coding, computer use, multimodal knowledge work | $1.50 in / $7.50 out | DeepSWE 49%, OSWorld-Verified 83% |
| Gemini 3.5 Flash-Lite | High-volume routing, classification, document extraction, agentic search | $0.30 in / $2.50 out | Fastest in series at 350 tokens/s |
| Gemini 3.1 Pro (preview) | Complex reasoning, long-context analysis where Flash falls short | ~$15 in / ~$60 out (estimate) | Frontier multimodal |
| Gemini 3.5 Pro | Not yet available — still in partner testing | TBD | TBD |
(Google blog) (AIToolsReview) (Bind AI)
The practical rule: for most tasks, Gemini 3.6 Flash is the right default — it is cheaper and more token-efficient than its predecessor while beating it on every agentic benchmark Google published. Use 3.5 Flash-Lite when you need raw throughput on simple tasks. Reserve 3.1 Pro for deep reasoning where cost is secondary to quality. And keep an eye on our Gemini 3.6 Flash vs 3.5 Flash-Lite routing for the decision rule.
For a deeper look at the newest workhorse model, see our hands-on guide to Gemini 3.6 Flash for AI agents.
What will Gemini 4 actually be good at?
Based on Google's stated priorities — not leaks — here is what Gemini 4 is being built to do, and what is still inference.
Coding and agentic coding (confirmed priority)
Pichai specifically flagged coding, particularly agentic coding (agents that write, test, and deploy code autonomously in long-running loops), as an area where Google trails. Google has formed a dedicated DeepMind coding team and is unifying internal AI coding tools. (Technology.org) The gains in Gemini 3.6 Flash were already concentrated in agentic evals — DeepSWE jumped from 37% to 49%, MLE-Bench from 49.7% to 63.9% — so Gemini 4 is being aimed squarely at the agentic coding space where Claude Code and OpenAI Codex currently dominate.
Autonomous agents (confirmed priority)
An agent does not just answer a question — it plans steps, uses tools, and finishes a job with far less human input than a single chat prompt. Google's Project Mariner agent prototype is already in official research publications. (FelloAI) The Gemini Enterprise Agent Platform, launched in 2026, gives enterprises a registry, gateway, and governance layer for agent deployment. (Google Cloud) The infrastructure is being built now; Gemini 4 is the model intended to power it at the frontier level.
Improved reasoning (inferred, not confirmed)
Google has not named a specific reasoning benchmark target for Gemini 4. Any claimed scores you see online are fabricated. What is confirmed is that Gemini 3.1 Pro already introduced Deep Think reasoning mode (Google's equivalent of extended thinking), and Gemini 3.5 Pro is expected to carry a 2-million-token context window — doubling Claude Opus 4.8. (Bind AI) Whether Gemini 4 inherits or extends these is unstated.
Video and multimodal (inferred from cadence)
Gemini Omni shipped as Google's video-creation series in May 2026. The Gemini 3 series already handles native video understanding. A generational leap to real-time video streams is plausible — but that is direction, not disclosure.
How to prepare for Gemini 4 without waiting
You can replicate most anticipated Gemini 4 workflows right now using models that already exist. Here is the practical playbook.
1. Switch to Gemini 3.6 Flash as your default agent model. It is the closest proxy to what Gemini 4 will do at the Flash tier — better agentic coding, fewer wasted tokens, cheaper per-task cost. If your workflow works on 3.6 Flash today, it will work on Gemini 4 tomorrow with minimal changes.
2. Design agentic workflows, not just prompts. The direction across every major lab in 2026 is the same: agents that plan, execute, and verify are the real productivity step-up, not marginally better one-shot answers. Stop crafting the perfect single prompt; start building multi-step agent loops that break a goal into steps, call tools, and self-check results. This is the skill set Gemini 4 is being built to reward.
3. Make your stack model-agnostic. Whether you ultimately use Gemini 4, Claude Opus, or GPT-5.6 is less important than whether your pipelines can swap models cheaply. Route by task type — a planner-executor pattern using a cheap model to decompose a task and a capable model to execute — keeps you flexible. When Gemini 4 lands, you point your router at it; you do not rewrite six months of integration work.
4. Start using Google's agent infrastructure now. The Gemini Enterprise Agent Platform is live — Agent Registry, Agent Gateway, IAM controls, and Model Armor are shipping today. Teams that already have agents registered and governed when Gemini 4 launches will be first to route the new model through production. Teams starting from scratch will spend weeks catching up.
5. Watch the official release notes, not the rumor mill. When Gemini 4 ships, the confirmation will appear in Google's Gemini release notes and the API model catalog — not a social media post with a screenshot. Check those two URLs directly. Ignore everything else until one of them confirms it.
What this means for you
If you run a small team or business and use AI for research, content, or automation, the Gemini 4 confirmation changes your planning horizon, not your to-do list today.
- If you have not adopted agentic workflows yet, start now with Gemini 3.6 Flash. The skill of designing agent loops — not the availability of Gemini 4 — is your actual bottleneck. Teams that learn the workflow shape today will swap in Gemini 4 as a drop-in upgrade. Teams that wait will start from zero when it lands.
- If you are choosing between AI platforms, the gap between Google, OpenAI, and Anthropic is narrower than the headlines suggest. Claude Opus 5 is here. GPT-5.6 is here. Gemini 3.6 Flash is a serious workhorse. The smart move is to test them side by side on your own workload and pick per task, not per brand.
- If you depend on Google's ecosystem (Workspace, Vertex AI, Android Studio), the infrastructure for Gemini 4 is being built into the products you already use. Google reports nearly 90% of the Fortune 100 uses Gemini Enterprise. (TheStreet) Your upgrade path will be largely automatic when the model ships.
- If you are budgeting for 2026–2027 AI spend, expect the per-task cost to keep falling. Flash-tier output pricing dropped from $9 to $7.50 per million tokens in the July refresh. (Google blog) Competition drives prices down whether Gemini 4 is revolutionary or incremental.
FAQ
Q: Is Gemini 4 released yet?
A: No. As of July 30, 2026, there is no public Gemini 4 model, no API access, no documentation, and no gemini-4-* model ID in Google's API catalogs. Google has confirmed only that pre-training has started.
Q: When will Gemini 4 be released? A: Google has not given a release date. Independent analysts estimate November or December 2026 based on the historical ~7-month gap between major Gemini generations, but this is an inference, not a Google commitment. Google delayed Gemini 3.5 Pro from June to an unspecified date, so its own date estimates are not reliable.
Q: How big will Gemini 4 be? A: Pichai told analysts Gemini 4 will need "a larger base model to compete at the frontier," confirming it will be meaningfully bigger than Gemini 3 Pro. No parameter count, architecture detail, or training-compute figure has been disclosed.
Q: What will Gemini 4 be best at? A: Google has stated coding and autonomous agents are the development priorities for the generation. This is consistent with the gains already seen in the Gemini 3.6 Flash refresh (DeepSWE, MLE-Bench, OSWorld-Verified). No specific new capability has been confirmed beyond that directional priority.
Q: Will Gemini 4 replace Gemini 3.5 Pro? A: Not immediately. Gemini 3.5 Pro has still not shipped to general availability and remains in partner testing. Google is running two release tracks in parallel: finishing the Gemini 3.x line (including 3.5 Pro) while pre-training Gemini 4 as the next full generation.
Q: Should I wait for Gemini 4 or use a current model? A: Use a current model. Gemini 3.6 Flash ($1.50/$7.50 per million tokens) is available now and handles agentic workflows, coding, and computer use at a fraction of frontier-tier pricing. Building production workflows on 3.6 Flash today means you are ready to swap in Gemini 4 the day it launches — not six months later.
Q: How much is Google spending on Gemini 4? A: Alphabet raised its 2026 full-year capital expenditure forecast to $195–$205 billion, up from $180–$190 billion the prior quarter. The increase funds the compute infrastructure for training and serving frontier models. Q2 2026 capex alone was approximately $44 billion.
Discussion
0 comments