BLUF: Google’s Gemini app crossed 1 billion monthly users this week, matching ChatGPT’s milestone from June and making Gemini the 14th Google product to hit that mark, according to TechCrunch. The same week, Anthropic took a different bet entirely: an unreleased Claude model made measurable progress on the Riemann hypothesis, one of mathematics’ most famous unsolved problems, per TechCrunch. Read together, the two stories map out three distinct strategies the frontier labs are now running in parallel — and which one matters most depends on what you’re actually building.
Gemini joins the billion-user club
Google confirmed this week that its Gemini app has surpassed 1 billion monthly active users, according to TechCrunch, making it the 14th Google product to reach that scale. The figure refers specifically to the standalone Gemini app — not the broader “AI Mode” in Google Search, which TechCrunch reports has separately crossed 1 billion monthly users of its own. ChatGPT hit the same 1-billion mark back in June, and The Verge reported this week that both chatbots have now cleared the threshold, framing OpenAI as “still the chatbot leader” but noting “the race appears to be tightening fast.”
The usage details TechCrunch reported are specific enough to matter for anyone building on top of these platforms: 63% of Gemini users now talk to the assistant directly using its voice feature, the app generates more than 150 million images per day, and Gemini has passed 100 million active users on iOS alone. The milestone lands right after Google’s Q2 2026 earnings call, where the company reported more than 950 million monthly users with daily active users tripling over the past year, and just ahead of a Made by Google event where more Gemini-powered Pixel features are expected. Google has also shipped Gemini 3.5 Flash, which TechCrunch reports is built specifically to improve coding and autonomous agent tasks — a signal that Google is not just chasing raw user count but trying to close the developer-workflow gap with OpenAI’s Codex and Anthropic’s Claude Code.
Anthropic is playing a different game
While OpenAI and Google trade blows over monthly active users, Anthropic spent the same week on a story with almost no user-facing component at all. According to TechCrunch, an unreleased Anthropic model made significant progress on the Riemann hypothesis, a problem so consequential that the Clay Mathematics Institute has a $1 million bounty on a full proof — one that remains unclaimed.
What makes the result notable is how it happened. An Anthropic staff member without significant mathematical training prompted the model to “take a real stab” at the hypothesis, then let it run largely unsupervised across roughly a day and a half. TechCrunch reports the model tested 650 different approaches, coordinating across 60 subagents and spending 31 million output tokens in total. A footnote in Anthropic’s paper, cited by TechCrunch, breaks down the division of labor in granular detail: of the 60 subagents, two developed the key mathematical ideas, 13 contributed supporting ideas, 30 attempted but failed to produce new ideas, 13 served as validators checking correctness, and the final two helped write up the paper. Two of Anthropic’s in-house mathematicians confirmed the result, and it was formalized using the open-source proof assistant Lean.
TechCrunch frames this as part of a broader pattern this year — LLMs solving a growing number of Erdos problems, and OpenAI separately publishing 10 major results proved by its internal “Astra” model. In other words, Anthropic is not sitting out the scale race by accident. It is spending its research and compute budget on a different kind of proof point: that its models can do original mathematical work, not just serve more users faster.
The detail that should stand out to anyone building agentic systems is the workflow, not just the result. TechCrunch’s account describes a human prompting the model once, with minimal mathematical background, then stepping back for a day and a half while the system self-organized 60 subagents into distinct roles — idea generation, validation, and writeup — without a human re-directing the process at each stage. Most production agent deployments today still require a human to check in every few steps or every few hours. A model that can sustain a self-coordinated, multi-agent research process for that long, on a problem hard enough to carry a $1 million unclaimed bounty, is a more useful capability signal for long-horizon automation than almost any benchmark score would be.
What this means for you
It is tempting to read a billion-user milestone as the headline and the math result as a footnote, mostly because user counts are easier to compare at a glance. That instinct is worth resisting. A monthly-active-user figure tells you how many people opened an app; it does not tell you whether the underlying model can be trusted to run an unsupervised multi-step process on your behalf. Both numbers matter, but they answer different questions, and conflating them is how procurement decisions end up optimizing for the wrong thing.
These two stories are not really about who “won” the week. They describe three different bets that are now running simultaneously across the frontier labs, and each one implies something different depending on what you’re building:
- If you’re building consumer or prosumer products, the Gemini and ChatGPT numbers are the more relevant signal. Both platforms are now default-scale distribution channels with hundreds of millions of daily touchpoints — voice, image generation, and mobile are where the growth is concentrated, per TechCrunch’s usage breakdown, so weight your integration decisions toward those surfaces rather than pure text chat.
- If you’re building on model capability for research, coding, or complex reasoning workflows, Anthropic’s math result is the more useful data point, even though it will never show up in a monthly-active-user chart. A model that can coordinate 60 subagents on an open problem for a day and a half without human intervention is a stronger signal for agentic-workflow reliability than a user-count milestone.
- Don’t assume “biggest” means “best for your task.” Our own testing in the ChatGPT vs. Claude vs. Gemini shootout found the highest-scale model was not the strongest performer on every task — distribution and capability are separate axes, and this week’s news is a clean illustration of why.
- Watch Anthropic’s roadmap for downstream effects. Research investments like the Riemann hypothesis work are frequently funded by the same infrastructure buildout covered in our report on Anthropic’s $10 billion Volta deal and new chip team — if compute gets tight, expect research-heavy bets like this one to compete directly with Claude API capacity.
- Check which surfaces actually drove Gemini’s growth. TechCrunch’s breakdown shows voice and image generation, not text chat, are where Gemini usage concentrated. If your integration plan assumes a text-first interface is the primary way users will reach an assistant, this milestone is a reason to re-test that assumption against how people are actually using the market leaders.
Three strategies, one takeaway for buyers
Zoom out and the pattern across both stories is the same: none of the three labs is trying to win the same race. OpenAI is defending its distribution lead while fighting a legal battle with Apple and managing its own leadership churn. Google is converting its existing product surfaces — Search, Android, Workspace, Pixel — into Gemini distribution at a pace no independent lab can match, which is why 14 Google products have now crossed a billion users apiece. Anthropic, without a comparable consumer distribution channel, is instead spending its compute budget proving that its models can do work no other lab has publicly demonstrated. None of these strategies is objectively “winning” in August 2026 — they are optimized for different outcomes, and the lab that looks best six months from now will depend on which axis turns out to matter most for your specific workload.
For a team choosing a primary model vendor, the practical move is to stop treating “biggest” as a proxy for “best,” and instead map your own workload against the axis each lab is actually optimizing for — distribution and reach for Google and OpenAI, or raw capability ceiling for Anthropic.
Key takeaways
- Gemini crossed 1 billion monthly users this week, becoming the 14th Google product to hit that scale and matching ChatGPT’s June milestone, per TechCrunch and The Verge.
- Google reported 63% voice usage, 150 million-plus daily image generations, and 100 million-plus iOS users for Gemini.
- The same week, an unreleased Anthropic model made verified progress on the Riemann hypothesis using 60 coordinated subagents and 31 million output tokens, confirmed by in-house mathematicians and formalized in Lean.
- The two stories represent different strategic bets — distribution scale versus research capability — and buyers should evaluate vendors against the axis that matches their actual use case.
Next step: Before your next model-vendor review, separate your evaluation criteria into “distribution fit” (voice, mobile, image volume) and “capability depth” (complex reasoning, agentic coordination) — and score each vendor on both, rather than defaulting to whichever has the bigger headline user count.
