What Is Gemini 4? Google's Next-Gen AI Model (October 2026)
What Is Gemini 4? Google’s Next-Gen AI Model (October 2026)
Gemini 4 AI is Google DeepMind’s next frontier generation. Its first model, Gemini 4 Argon, was announced on September 30, 2026 and is in limited release to trusted cyber defenders. It was not launched at Google I/O 2026 as this page originally reported — I/O (May 19-20, 2026) introduced the Gemini 3.5 series and Antigravity 2.0.
Last verified: October 6, 2026. (First published May 19, 2026.)
What changed as of October 2026
- Correction: the May 19, 2026 version of this page said Gemini 4 was unveiled at the Google I/O keynote and went live across the Gemini app, Search AI Mode and Vertex AI that day. That was wrong. Google’s I/O announcements were the Gemini 3.5 models (led by Gemini 3.5 Flash) and Antigravity 2.0; Gemini 4 was not announced there. Google DeepMind teased Gemini 4 pre-training in July 2026 (see Is Gemini 4 out yet?).
- Gemini 4 Argon was announced on September 30, 2026. It rolls out first to trusted cyber defenders through Google’s Fairwind Program, without cyber guardrails for those defenders; general access follows “as soon as possible”, starting with paid API customers and Google AI Ultra subscribers.
- Pricing is set: introductory $2 per million input tokens and $10 per million output tokens (cached input 95% off), rising to $4 / $20 after the introductory period.
- The leaked specs were not confirmed. Google’s announcement does not claim a 10M-token context window or an ARC-AGI2 score of ~84.6%; what it does confirm is a 1M-token output limit (up from 64K).
- Our newer pages: How to get Gemini 4 Argon access · Gemini 4 Argon vs Gemini 3.8 Flash · Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 Astra.
Quick facts
| Property | Value (as of October 6, 2026) |
|---|---|
| Announced | September 30, 2026 (Gemini 4 Argon, Google blog) — not at Google I/O in May |
| Vendor | Google DeepMind |
| Availability | Limited: Fairwind Program cyber defenders, internal Google teams, U.S. government pre-release testing |
| Next to get access | Paid Gemini API customers and Google AI Ultra subscribers (no date) |
| Output limit | 1M tokens (up from 64K) |
| API price | $2 input / $10 output per 1M tokens (introductory) → $4 / $20 later |
| Modalities | Coding, reasoning and multimodal input incl. charts, documents and long video |
| Successor to | Gemini 3.x line (current public Flash model: Gemini 3.8 Flash) |
| Earlier leaked claims | 10M-token context, ~84.6% ARC-AGI2, Ironwood-optimized — not confirmed in Google’s announcement |
What’s actually new
What Google confirmed on September 30, 2026:
- 1M-token output limit — up from 64K. Google says the headroom lets the model reason through hundreds of thousands of tokens in one trajectory.
- Long-horizon engineering — Google says Argon agents are migrating C/C++ codebases to Rust inside Google (up to 800K+ lines for the Fuchsia Zircon kernel) and produced a Rust video decoder 2.7x faster than the earlier port.
- Enterprise knowledge work — leading results on the Vals Index (finance, coding, legal, tax), Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.
- Cyber defense — able to find, validate and patch vulnerabilities; released without cyber guardrails to trusted defenders only.
- Safeguards before broad release — misuse refusals per Google’s Frontier Safety Framework, prompt-injection hardening (leading on Gray Swan’s indirect prompt injection benchmark), and chain-of-thought misalignment monitoring.
The May 2026 version of this section listed a 10M-token context, native video/music generation, Android “Gemini Intelligence” workflows and Ironwood-optimized inference as Gemini 4 features. Those came from pre-I/O leaks and are not in Google’s Gemini 4 announcement.
Benchmarks (publicly cited)
| Benchmark | Gemini 4 Argon | Notes |
|---|---|---|
| DeepSWE v1.1 | 77.9% | Long-horizon software engineering — state of the art per Google |
| AutomationBench (Zapier) | 51.3% | #1 per Google |
| LVBench | 91.7% | Long video understanding — state of the art per Google |
| Vals Index | #1 | Economic-impact index across finance, coding, legal, tax |
| ARC-AGI2 | — | The ~84.6% figure was a pre-I/O leak; Google has not published an ARC-AGI2 score |
Benchmarks where Argon trails rivals (from Google’s own comparison table, as recorded in our pricing reference on October 5, 2026): FrontierSWE v2 (GPT-6 Astra ahead) and Terminal-Bench 4.0 (Claude Opus 5.5 ahead).
What Gemini 4 powers
| Surface | Status as of October 6, 2026 |
|---|---|
| Fairwind Program | Available — trusted cyber defenders, without cyber guardrails |
| Internal Google work | Available — Google says thousands of Googlers use it daily |
| Gemini API (paid tier) | Not yet — first in line when access widens |
| Google AI Ultra subscription | Not yet — first in line on the consumer side |
| Gemini app, AI Mode in Search, Android | Not announced for Gemini 4 Argon; these run on Gemini 3.x models today |
| Googlebook, Android XR glasses | Products announced around I/O 2026 — not tied to Gemini 4 in Google’s announcement |
How it compares (October 2026 frontier models)
| Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | |
|---|---|---|---|
| Vendor | OpenAI | Anthropic | |
| Released | Sep 30, 2026 (limited) | Sep 3, 2026 | Sep 22, 2026 |
| Public API | Not yet | Yes | Yes |
| API price (in / out per 1M) | $2 / $10 intro → $4 / $20 | $10 / $50 | $4 / $20 |
| Strength | Long-horizon engineering, enterprise workflows, long video | FrontierSWE, Terminal-Bench Science | Terminal-Bench 4.0, agentic coding |
Rule of thumb in October 2026:
- Need it today? Gemini 4 Argon is not an option unless you are in Fairwind — use Gemini 3.8 Flash, GPT-6 Astra or Claude Opus 5.5.
- Planning for when it opens? Budget at the post-intro $4 / $20 rate.
- Pure agentic coding in a terminal harness? Claude Opus 5.5 still leads Terminal-Bench 4.0.
The May 2026 version of this table compared “Gemini 4” with GPT-5.5 and Claude Opus 4.7; both have since been superseded.
Where Gemini 4 wins
- Very long outputs — a 1M-token output limit is far above typical frontier limits (128K for GPT-6 Astra and Claude Opus 5.5).
- Enterprise knowledge work — #1 on the Vals Index across finance, coding, legal and tax.
- Long video and document understanding — state of the art on LVBench.
- Price at launch — the $2 / $10 introductory rate is below both main rivals.
- Google product graph — Gmail, Calendar, Drive, Maps and YouTube context once it reaches Google’s consumer surfaces (no date yet).
Where it might not be the right call
- Anything you need to ship this quarter — there is no public access date.
- Terminal-heavy agentic coding — Opus 5.5 leads Terminal-Bench 4.0 and GPT-6 Astra leads FrontierSWE v2.
- Enterprise data residency outside Google Cloud — Bedrock-managed Anthropic / OpenAI may suit better.
What ships when
- May 19-20, 2026 (Google I/O) — Gemini 3.5 series and Antigravity 2.0; not Gemini 4.
- July 2026 — Google DeepMind teases Gemini 4 pre-training.
- September 30, 2026 — Gemini 4 Argon announced; Fairwind Program access begins.
- Next (no date) — paid Gemini API customers and Google AI Ultra subscribers, then developers, enterprises and consumers more broadly.
TL;DR
Gemini 4 is real but not yet public: Gemini 4 Argon was announced September 30, 2026, is limited to Fairwind cyber defenders and trusted testers, and will cost $2 / $10 per million tokens at launch ($4 / $20 later). It was not launched at Google I/O in May 2026, and the 10M-context and ARC-AGI2 figures that circulated then were never confirmed.
Sources
- Google — Gemini 4 Argon: our next era of frontier intelligence (September 30, 2026)
- Google Developers Blog — All the news from the Google I/O 2026 Developer keynote (May 19, 2026)
- Google — I/O 2026: Sundar Pichai’s opening keynote