Gemini 4: $200 Billion In, Nobody Knows What It Does

Link to our newsletter: https://bitbiased.ai/ Google has confirmed Gemini 4 is training — but almost everything else being said about it is speculation. Sundar Pichai described Gemini 4 as “very ambitious” during Alphabet’s July earnings call, while Google’s AI operation was dealing with a wave of high-profile researcher departures, a major DeepMind leadership reshuffle, and enormous pressure to justify one of the biggest infrastructure spending pushes in the company’s history. So instead of inventing Gemini 4 specs, release dates, or benchmark scores, this video starts with what we can actually measure: Gemini 3.8 Flash, the current foundation Gemini 4 has to improve on. :contentReference[oaicite:0]{index=0} Gemini 3.8 Flash ships with a 1,048,576-token context window, up to 65,536 output tokens, native support for text, images, audio, video, and PDFs, and three reasoning levels. But the most revealing part is its benchmark profile. Terminal-Bench 2.1 jumped from 81.6% on Gemini 3.7 Flash to 90.8% on 3.8 Flash. SWE-Bench Pro improved to 61.6%. Meanwhile, Humanity’s Last Exam barely moved, from 54.6% to 54.9%. That split matters because Pichai has specifically highlighted coding and agentic coding as areas where Google needs to improve. Gemini’s recent progress appears strongest exactly where Google knows it needs to compete — while harder, general reasoning remains a much bigger question. And the competition isn’t standing still. The relevant comparison now includes OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Sonnet 5 and Claude Fable 5.1, not older models that still appear in many benchmark charts. Pricing, coding, reasoning, and agentic performance all tell different stories depending on which benchmark you examine. Underneath all of this is Google’s massive infrastructure bet. Ironwood TPU v7 delivers 4,614 teraflops of FP8 compute with 192GB of high-bandwidth memory per chip. Google says a superpod containing more than 9,000 Ironwood TPUs can reach 42.5 exaflops. At the same time, Anthropic has committed to buying up to one million of these chips — meaning Google is supplying frontier-scale infrastructure to one of its most important AI competitors. Alphabet’s 2026 capital-spending guidance has also climbed to $195–205 billion, while Google is securing future energy through agreements including up to 500 megawatts of nuclear capacity from Kairos Power. Then there’s the talent story. Noam Shazeer left for OpenAI. AlphaFold co-creator John Jumper went to Anthropic. Jonas Adler and Alexander Pritzel also departed for Anthropic. Jeff Dean left after 27 years at Google to build a new AI research company. And on August 5, Demis Hassabis moved from CEO of Google DeepMind to chairman and Alphabet chief scientist, leaving Koray Kavukcuoglu in charge of day-to-day operations and the Gemini roadmap. All of that makes Gemini 4 more than another model release. The real question is whether Google’s “much larger base models,” enormous TPU deployment, infrastructure spending, and organizational changes can translate into the reasoning gains its current models haven’t yet demonstrated. There’s also the million-token context ceiling. Gemini 3.5 through 3.8 have stayed at roughly the same maximum context size. Google previously made a generational leap when Gemini 1.5 jumped from 32,000 tokens to one million. Could Gemini 4 break another context barrier? It’s a possibility based on Google’s history — not a confirmed feature. And despite rumors, there is still no confirmed Gemini 4 release date, parameter count, context window, or official benchmark sheet. What we do have is enough evidence to understand what Google needs Gemini 4 to accomplish — and why this release carries considerably more pressure than the generations before it. CHAPTERS 00:00 Google Is Already Training Gemini 4 01:53 Who Google Lost While Building It 03:17 The Reshuffle Nobody Expected 04:09 The Model Gemini 4 Actually Has to Beat 06:41 Measured Against Who’s Actually Competing 07:37 The Infrastructure Bet Underneath It 09:48 What Gemini 4 Is Actually Likely Built to Fix 11:47 Where This Actually Leaves Google #Gemini4 #GoogleGemini #GoogleDeepMind #ArtificialIntelligence #AI