The Frontier Without a Leaderboard: Why Google Gemini Is My Biggest AI Bill

On July 21 Google shipped three models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. No Pro. No Ultra. The flagship 3.5 Pro is still “testing with partners” with no ship date, and the one specialist model that sounds exciting, the cyber model tuned for vulnerability detection, is locked to “governments and trusted partners” via a limited pilot. So: a mid-tier release, a withheld frontier, and a security model nobody outside a gov contract gets to touch.

The internet reached its verdict within the hour. Google lost the frontier.

On the thing they’re measuring, they’re right. And I’m about to tell you why it doesn’t matter to me, because I spend more money on Google every month than on the tools that actually beat it.

Yes, Google Lost the Coding Race

Let me concede the whole case before I argue with it, because the case is real.

Google’s own model card puts 3.6 Flash behind GPT-5.6 Luna, Grok 4.5, and Claude Sonnet 5 across the agentic-coding benchmarks: SWE-Bench Pro, DeepSWE, Terminal-Bench. Not close on some of them. The community was less polite than the card.

This model is not for builders and engineers. From the outside it seems as though Google cannot keep up with other frontier models.

— Hacker News commenter on the Gemini 3.6 Flash release

One tester on X: “Another flop from Google. They shipped a speed patch and called it a version. Six months of silence for THIS?” Another noted it’s “less intelligent and more expensive than GLM-5.2, while being closed weight,” which is a genuinely brutal thing to be in mid-2026. If your job is agentic coding, none of this release is for you, and I wouldn’t argue otherwise.

So that’s settled. Now the part nobody benchmarks.

The Frontier I Actually Pay For

Here is my AI spend, honestly. I run five coding subscriptions. Together they cost me roughly a thousand US dollars a month. And my Google inference bill, metered per token across a handful of products in production, beats all five combined. Gemini is my single largest AI line item, and it isn’t close.

That should be impossible if Google “lost.” It’s only impossible if you think there’s one frontier. There are two, and only one has a leaderboard.

The benchmarked frontier is agentic coding, because coding is what the people who post benchmarks do all day. The other frontier is product inference: generating images, moderating user content, reading a photo and deciding if it’s real, writing landing copy grounded in live search, a million times, at a per-unit cost that has to survive contact with an actual P&L. Nobody runs a Frontend Code Arena for that. But it’s where the volume is, and quietly, it’s Google’s.

Google isn’t even hiding the strategy. Logan Kilpatrick, who runs the Gemini API, sold 3.6 Flash as “higher intelligence, more token efficient, and with a new lower price, based directly on developer feedback… deeply usable in real world scenarios.” DeepMind’s own line was “faster, smarter, and cheaper at scale.” Read that again: not smartest, cheaper at scale. That’s not a frontier lab that missed. That’s a frontier lab that picked a different frontier, one with margins instead of headlines.

What That Spend Actually Buys

That bill isn’t one runaway app. It’s several products, each metering Gemini at volume in production. The clearest one to point at is famecake.com, a mobile billboard-advertising platform where AI is the feature, and it makes the argument concrete: it runs entirely on Gemini. No Anthropic, no OpenAI, nowhere in the codebase. Every AI call is Google.

What famecake does with Gemini, and what the others do in their own shapes:

• Image generation with Nano Banana (the gemini-flash-image line) for every billboard creative users make, plus reframing and outpainting to fit different screens.
• Content moderation on every single edit, a vision call with structured JSON output against a category taxonomy, cheap enough to run on flash-lite so it can run constantly.
• Photo verification, a multimodal call that looks at a user’s proof photo and judges whether it really shows their ad on a real screen, routing anything uncertain to a human.
• Grounded growth copy using Google Search as a live tool, and landing-page SEO at scale.

Every one of those is a job Claude and GPT are either worse at, can’t do, or can’t do at a price that works per-unit. Native image generation especially: that’s not a thing you route to Claude at all. The division of labor isn’t loyalty or laziness. It’s that multimodal-at-volume is a different sport, and Google is winning it.

And you might think that’s the catch: sure, Google wins the image stuff, but that’s a narrow niche. So take a second product I run, a product market intelligence tool, that looks nothing like famecake. No images at all. It’s a pure-text batch pipeline that pulls structured attributes out of tens of thousands of messy listings, resolves which company owns which brand, and writes clean descriptions, roughly 90,000 records a pass. The workhorse is Gemini Flash-Lite. I benchmarked the pricier Gemini tier against it and killed it in my own docs: 46 times the cost for no accuracy gain. OpenAI’s GPT-5-mini is in there too, but only as a decorrelated second opinion that votes on the hard calls and judges the output. The bulk tokens, the ones that actually move the bill, are all Google.

The brand-resolution step is the tell. It leans on grounded search, Google Search wired straight into the model so it can look up who really owns a brand mid-inference, and that is another axis, like multimodal, where nobody seriously competes with Google. Both products lean on it: famecake grounds its growth copy the same way. Opposite use cases, same verdict: at volume, and anywhere search-grounding matters, nothing else pencils out.

And the punchline writes itself: I built famecake, a Gemini-only product, using Claude and OpenAI. The tools that wrote the code and the model the code calls sit at opposite ends of the industry’s supposed leaderboard. I reach for Anthropic and OpenAI to write the software. I reach for Google to run it.

Where This Doesn’t Hold

I’m not telling you Gemini is what you build everything on. Both my examples are high-volume inference where “good enough at a fraction of the price” is the whole game. That is not every workload.

The honest caveats:

• The frontier ranking matters when the task is genuinely hard. For a single difficult reasoning call, or an agent looping in a terminal, you want the top of the leaderboard, and that isn’t Gemini. My case is inference at volume, not a claim that Gemini wins the hard single-shot problems. Different question, different answer.
• It’s closed-weight and pricier per token than GLM-5.2. For the privacy-and-cost crowd I’ve written about before, self-hosted open weights still win. Google’s pitch is managed convenience at scale, not sovereignty.
• The gated cyber model is a real miss, not a hidden strength. Shipping a security model only governments can use, in the exact week defenders are reaching for open models that don’t refuse them, is Google reading the room backwards.

The Takeaway

• “Lost the frontier” is a category error when there’s more than one frontier. Google lost the benchmarked one and is quietly running the metered one. Both are true at once.
• The leaderboard measures what its authors do for a living. Coding benchmarks dominate because coders write the benchmarks. The largest real AI workloads, product inference at volume, have no comparable scoreboard, so they look like nothing is happening there. Something very much is.
• Watch where the money goes, not where the tweets go. My single biggest AI expense is the vendor everyone spent this week writing off. If that’s true for one indie product, it is very much true at the scale of companies who never post a benchmark in their lives.

There’s a leaderboard for the frontier Google lost. There’s an invoice for the one it won. The whole industry spent this week reading the leaderboard out loud, and not one of those posts mentioned that the company at the bottom of it is the largest number on my books, and probably on the books of everyone shipping AI instead of screenshotting benchmarks. Google didn’t miss the frontier. It looked at the race with a scoreboard and the race with a cash register, and went to stand by the register. The scoreboard is louder. The register is the one that rings.