# Gemini 4 Release Date: What We Know So Far > Source: > Published: 2026-08-24 00:00:00+00:00 # Gemini 4 Release Date: What We Know So Far Google confirms Gemini 4's pre-training run is underway. Here's the leaked timeline, DeepMind leadership shakeup, and what it means. ## Gemini 4 doesn’t have a confirmed release date yet, but the signs point to late 2025. Google has publicly confirmed that Gemini 4’s pre-training run is underway, and prediction markets currently give it roughly an 85% chance of shipping by November 30. Sundar Pichai has said the model needs to be “significantly larger” to compete with rivals like Claude and GPT-5, and the announcement lands alongside a major leadership shakeup at DeepMind, including Demis Hassabis stepping back from day-to-day CEO duties. Nothing is official beyond Google’s own statements, so treat specific dates and leaked benchmarks as unconfirmed until Google says otherwise. ## TL;DR - Google confirmed on July 21 that it has started its **most ambitious pre-training run yet** for Gemini 4, a rare public acknowledgment mid-training rather than at launch. - Sundar Pichai told investors on the **Q2 earnings call**(July 23) that Gemini 4 will use a significantly larger base model, with coding and autonomous agents named as top priorities. **Demis Hassabis stepped down as CEO of DeepMind**, moving into a chair and Alphabet chief scientist role, amid a wave of departures including Noam Shazeer, Jonas Adler, and Alexander Pritzel to OpenAI and Anthropic.- Leaked, unverified evals suggest Gemini 4 may carry a **very large context window** and could match or beat Claude Opus and GPT-5 class models on coding, though none of this is confirmed. - A Google SDK reference reportedly mapped a **Gemini 4 Flash preview** to a Gemma 4 tokenizer family, which points to active internal testing. **Polymarket odds** place a 68% chance of release by October 31 and 85% by November 30, roughly matching analyst expectations of a November or December launch.- A mystery model called **“Horizon Alpha”** appeared on OpenRouter and in coding tools around the same time, fueling speculation it could be an early Gemini 4 test, though this is unconfirmed and it may belong to another lab entirely. ## Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. ## What has Google actually confirmed about Gemini 4? Google’s confirmation is narrow but notable. On July 21, the company stated it had begun its most ambitious pre-training run to date for Gemini 4. That kind of mid-training announcement is unusual. Labs typically stay quiet until a model is closer to release, so the disclosure reads as a signal to investors and competitors that Google is scaling up hard after a rocky stretch. Two days later, on the Q2 earnings call, Pichai added detail: Gemini 4 requires a substantially larger base model to keep pace with the rest of the frontier field, and the company is prioritizing coding and autonomous agent capability. That framing matters. Gemini 3.5 Pro missed its expected release window and landed behind competing models when it finally shipped, and a next-generation model that isn’t at least competitive on coding would be a real problem for Google’s positioning against Anthropic and OpenAI. Beyond the earnings call, there’s indirect technical evidence of internal testing. A reference in a Google SDK reportedly linked a “Gemini 4 Flash preview” build to a Gemma 4 tokenizer family, an implementation detail that would only exist if a Gemini 4 variant were already running through Google’s internal pipelines. ## Why is Google’s DeepMind leadership shakeup relevant to Gemini 4’s direction? The timing of Gemini 4’s development overlaps with a significant leadership transition at DeepMind. Demis Hassabis has stepped down as CEO, moving into a role as DeepMind chair and Alphabet chief scientist, a shift that reportedly frees him to spend more time on Isomorphic Labs, DeepMind’s drug discovery spinout. The context behind that move matters for anyone tracking where Gemini 4 is headed. Hassabis was reportedly less committed than rival labs to the industry’s dominant strategy of pushing coding and agentic models toward recursive self-improvement (RSI), the idea that AI systems become capable enough to meaningfully accelerate their own research and development. Hassabis has instead been associated with a world-models approach to advancing capability, a different bet than the coding-and-agents sprint that OpenAI, Anthropic, and xAI have all leaned into. That divergence has coincided with a wave of departures from DeepMind: Noam Shazeer left for OpenAI, and John Jumper, Jonas Adler, and Alexander Pritzel moved to Anthropic. Reported reasons include low morale, leadership described as absent, sustained 60-hour work weeks, and internal friction over DeepMind’s Pentagon contract work. Notably, Google co-founder Sergey Brin has reportedly been back in the office working directly on coding models, which reinforces the sense that Gemini 4 represents a real strategic pivot toward the RSI-style approach the rest of the industry has already committed to. ## What do the leaked Gemini 4 benchmarks actually show? Some unverified evals have circulated suggesting Gemini 4 could ship with a very large context window and coding performance competitive with, or ahead of, Claude Opus and GPT-5 class models. None of this is confirmed by Google, and leaked benchmarks for unreleased frontier models have a poor track record of matching what actually ships. Treat these numbers as directional rumor rather than fact. ## Plans first. *Then code.* Remy writes the spec, manages the build, and ships the app. The reason to take the coding claims seriously at all is strategic logic rather than the leak itself: Pichai explicitly named coding and autonomous agents as priorities, and Google has a clear incentive to close the gap after Gemini 3.5 Pro underdelivered. If Gemini 4 doesn’t compete credibly on coding benchmarks at launch, that would represent a real setback for Google’s positioning in the frontier model race, not just a missed marketing beat. ## What is Horizon Alpha, and is it related to Gemini 4? A stealth model nicknamed “Horizon Alpha” appeared anonymously on OpenRouter and inside coding tools, triggering speculation across AI communities about which lab built it. Some early testers report strong results on certain benchmarks; others describe it as unremarkable compared to current frontier models. Reports suggest it may be an early multimodal model accepting text, image, and video input, and on at least one broad evaluation suite it scored in the mid-range, roughly comparable to GPT-5 class models rather than clearly ahead of them. Speculation has connected Horizon Alpha to several possibilities: an early Gemini 4 checkpoint, a Chinese lab’s stealth release, or the first public model from Safe Superintelligence (SSI), Ilya Sutskever’s company. Investor Gavin Baker reportedly said in a podcast interview that SSI was targeting an August release. None of these connections are confirmed, and the model’s true origin remains unknown. It’s a useful data point for the general acceleration happening across labs right now, but it shouldn’t be treated as evidence for a specific Gemini 4 release date. ## When will Gemini 4 actually be released? Nobody outside Google has a confirmed date. The best available signals are market-based and pattern-based rather than official. Prediction markets on Polymarket have priced in a 68% chance of release by October 31 and an 85% chance by November 30. Independent analysts following Google’s historical release cadence generally expect a November or December launch, which would put Gemini 4 roughly a year after Gemini 2.0’s debut and would give Google a full pre-training and safety evaluation cycle after its July confirmation. Given that Google explicitly confirmed the pre-training run was still underway in July, a release before October seems unlikely purely on the mechanics of training, fine-tuning, and safety testing large models. A Q4 2025 window is the most defensible estimate based on public information, but it remains an estimate, not a commitment from Google. ## Frequently Asked Questions ### Has Google officially announced a Gemini 4 release date? No. Google has confirmed that Gemini 4’s pre-training run is underway and that it will be a significantly larger model, but it has not announced a specific release date. ### Why did Demis Hassabis step down as DeepMind CEO? Hassabis moved into a chair and Alphabet chief scientist role rather than leaving the company. Reported context includes internal disagreement over strategy (world models versus the industry’s coding-and-agents push toward recursive self-improvement) alongside broader departures of senior researchers to OpenAI and Anthropic. ### Is Horizon Alpha the same model as Gemini 4? It’s unconfirmed. Horizon Alpha is a stealth model that appeared on OpenRouter and in coding tools, and its lab of origin hasn’t been verified. Theories range from an early Gemini 4 test to a Chinese lab release to Ilya Sutskever’s Safe Superintelligence, but none are confirmed. ### What will Gemini 4 be optimized for? ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. Sundar Pichai said coding and autonomous agent capability are priorities for Gemini 4, alongside a significantly larger base model overall. This aligns with the broader industry focus on agentic and coding performance as a path toward more autonomous AI systems. ### How reliable are the leaked Gemini 4 benchmarks? Not very, at this stage. They’re unverified leaks, not confirmed by Google, and past leaked benchmarks for unreleased frontier models have often failed to match final shipped performance. They’re worth watching as a directional signal, not treating as fact.