Gemini’s API Grew 14× in Six Months. Google’s PMs Explain the Catch-Up.

Nathan Labenz’s guests are Shrestha Basu Mallick and Logan Kilpatrick, product managers for Google’s Gemini API and AI Studio. The occasion is a new grounding feature that lets Gemini pull live web information from Google Search at runtime. It arrives a couple of days after an earnings call in which CEO Sundar Pichai said Gemini API usage had grown by a factor of 14 over six months, and after what Labenz estimates as a cool hundred billion dollars added to Alphabet’s market cap after hours.

Kilpatrick, on his third Cognitive Revolution appearance and openly chasing the show’s guest record, starts with the door rather than the model. AI Studio is the entry point for developers building with Gemini: every latest model is there, and the intended experience is as close to zero friction as they can imagine. Three or four clicks should get you an API key, a chance to poke the models, and enough generated code to leave and build something. Labenz, who had spent the morning wiring Gemini into an app he is developing, does not dispute the click count.

The 14× is not, in their telling, a single launch. Independent of search grounding, the last six months stacked experimental models, production models, a new Flash 8B class, code execution, context caching, and structured outputs, with a running effort to clean up error codes, rate limits, and the rest of the platform. Kilpatrick’s working theory is that developers are being trained, implicitly, to gravitate toward whichever platform keeps shipping. Basu Mallick adds that the first versions of those features are not the last: they keep taking developer feedback on the features and on the overall experience.

Labenz has watched people try Gemini, wander into Vertex, and get lost, then blame Google. He puts himself in the other line: for prototypes and low-stakes proofs of concept, AI Studio is on par with OpenAI’s and Anthropic’s developer products, and in some ways ahead. The workflow he actually used was choose a template, tinker in the sidebar, generate code, paste it into a coding assistant, and let it land in the app. End to end, that took about 90 minutes with Vercel’s AI SDK, and Gemini was still the third provider he added, after OpenAI and then Claude Sonnet.

He went looking for a product reason and did not find one. LMSYS arena scores put Gemini models at or near the top; ease of use, structured outputs, and the rest of the feature list felt like a dead heat. So is the ranking just a hangover from before Kilpatrick’s era? Kilpatrick calls the answer complicated, then offers a simple slice of it: Gemini has not even been out a full year, with the first model announced in December 2023, while OpenAI had models and an API dating to 2021 or 2022. The last year was a mission to reach parity with the ecosystem, then start delivering differentiated capabilities.

Some of that differentiation, he says, was there from the beginning: Gemini was the first large-scale frontier model to go natively multimodal, launched with a million-token context window, then 1.5 Pro’s two million at I/O. The breakout was 1.5 Flash. You can track the API-volume climb back about six months, to a model that balanced the quality, price, and speed developers actually care about. The remaining problem is mindshare. People already chose an LLM when the technology first arrived, and Google is still getting over that. The strategy now is to solve problems uniquely suited to Google — search grounding, and long context as an infrastructure challenge — on top of models that already look strong on public metrics.

Basu Mallick is blunt about the rest. When they talk to developers, they hear that OpenAI and Anthropic have built excellent models with great buzz, and a busy founder prototyping does not have time to evaluate everything. Under those conditions, going with the loudest name is directionally a safe choice, and they have empathy for that. Google’s counter is to make prompting guidance and quick evals easy enough that someone can see the quality-price-latency tradeoff in minutes. Kilpatrick thinks perception is shifting — the joke on Twitter is that the vibe is shifting — but you earn a reception change over a very long time. This is a multi-year endeavor. The last six months make him think they will actually do it, if they keep earning it.

The earnings call and accompanying blog named Snap, and Hiscox, an insurer using Gemini to cut quotes on complicated risks from days to minutes. Labenz wants the more useful category: teams that evaluated the field and chose Gemini for a reason. Kilpatrick’s first answer is a class of work, not a logo. Gemini excels, and developers get most excited, around multimodal problems, because native multimodality has been in the model’s DNA from the first iteration. Those are also the use cases with the fewest products and companies so far: native audio, native video understanding, native image understanding, scaled into production, still in the early innings.

Tldraw is the example he loves, a company pushing on what the next LLM user experience might look like and powering it with Gemini’s image, audio, and video work. From a founder’s seat, he would rather sit on the frontier than in text, where a thousand and one startups already cover every plausible use case. Basu Mallick adds the rest of the current scoreboard. New Computer built an AI companion called Dot whose job is to know a person; they use Flash’s long context and low price to compress daily conversations into retrievable memory objects. Envision helps people with low vision understand their surroundings, a setting where Flash’s output speed is the point. Locofy turns Figma files into code with its own models, and uses Gemini 1.5’s million-token window to reformat and rename front-end designs that can contain hundreds of layers before handing them to a proprietary generator.

Kilpatrick has been sending people to Artificial Analysis, which goes a few levels deeper than LMSYS ELO by folding in price and latency. The chart that keeps putting a smile on his face is Flash sitting in a quadrant by itself. LMSYS is a decent proxy for how good a model is, he says, and it does not quantify many of the characteristics developers actually use to pick a production model. Flash also leads on time to first token and tokens per second. His favorite circulating tweet was that Google is making it impossible not to build with Flash because of how good the value proposition is. Flash 8B is another step down the same price-performance-latency curve.

Labenz had once challenged himself to spend a dollar a day on Flash, processed five years of email — on the order of 12 million tokens — and spent nothing, because it fell into the free tier. He wants more people thinking about information processing at previously unimaginable scale. The phrase he keeps using is luxury software: an app that reads every new arXiv paper against 50,000 or even 100,000 tokens of personal context, a hundred papers a day, still lands around a dollar. Kilpatrick notes that a static block like that belongs in context caching, so you pay a tiny storage fee and the incremental tokens; Google shipped caching first, and Anthropic and OpenAI now have their own versions.

The free tier is, for Google, a strategy rather than a promo. Kilpatrick’s view is that the fundamental limit on building with AI is still economic: the cost to experiment and put something in users’ hands is generally high. Gemini’s free tier needs no credit card, works in 200-plus countries, and on 1.5 Flash allows 1,500 requests a day, each of which can be a long-context call, or about 1.5 billion tokens every day. Search grounding and code execution are available to try for free in AI Studio. Swyx had floated a live video feed at one frame per second still fitting in the free tier; Basu Mallick walks the arithmetic: a million tokens is roughly an hour of video, each frame about 258 tokens plus metadata, or roughly 300, which is about one frame per second.

Labenz separates two cost conversations. As friction reduction, a free taste is obvious, and YouTube making roughly a billion dollars a week can afford it. As a project budget, he tells people they should save 90 percent versus human labor, then spend 90 percent of the remaining 10 percent on R&D and people, leaving about 1 percent for inference. He sees developers far too worried about that last percent. Kilpatrick blames a memory: the first time many developers looked at LLM prices was two years ago, at something like $30 per million tokens, and that number stuck even though costs have fallen about 99.9 percent. Ask a thousand developers if AI is expensive, he bets, and most will say yes. Only about a fifth of developers are actively building with AI; he posits cost as a big reason, and Labenz half-jokes that obstinance may be harder to fix.

Ongoing camera monitoring is, for Labenz, an obvious Flash job. He has been in conversations with companies that already have cameras in warehouses and factories and are not actually that good at knowing what is going on. His unwelcome answer is to use a foundation model. The same pattern shows up in senior living: people hate fall-monitor wristbands, take them off, and then fall, and a camera plus a fast model that notices within a minute is the kind of thing that saves lives, privacy fights included.

Multimodal fine-tuning is on the roadmap, Kilpatrick says, and a request they get often. He spent the early part of his career training domain-specific computer-vision models, including a year on data curation and iteration for a single problem, and his assumption now is that those problems probably just work out of the box with models like Gemini. Image fine-tuning is the knob he expects teams to use when they need production-scale quality to flip an existing system. The more interesting set is everything never deployed because it was cost-prohibitive: a small business owner who cannot hire a crew of ML experts for a niche problem, and who will instead throw a prompt and a couple of images at a foundation model.

Basu Mallick flags a smaller unlock that already shipped. Fine-tuned models, and the act of tuning, can now be reached with API keys. That used to require OAuth, and it was one of the requests they heard constantly. For anyone who has been blocked on that plumbing, the change is more enabling than a new model card.

The basic idea is intuitive: consult Google Search, then use what comes back to make the answer more accurate, fresher, and better on long-tail queries the model does not handle well. Labenz has built the homemade version, typically with Brave’s search API, which he calls good, fast, and cheap, then a second model to summarize. Having it in one call is an obvious win. He does not have a huge coverage benchmark, but at Waymark he searches for long-tail small businesses, and his prior is that Google still wins the long tail. Brand trust in the Google search stack, built over a long time, is now enough that he would tell people to try this first.

Testing turned up a second effect: with grounding on, answers often get richer, not just more current. Ask for the smallest horse breed in the world and the model alone might say Falabella; with search it adds average size, Argentina, and more detail. Kilpatrick and Basu Mallick disagree on a toy query, “what is the capital of Mars.” The ungrounded model says Mars has no capital, which Kilpatrick prefers. The grounded version lists science-fiction books that have given Mars a capital, which Basu Mallick prefers. Developers, she says, can pick for product taste.

Control is two-layered. First is on or off: a toggle in AI Studio, a tool parameter in the API. Second is dynamic retrieval, a 0-to-1 parameter that says, even with grounding on, how often to invoke search. Set it to zero or low and search fires on almost every query, useful if you want a rigid layer of extra detail. Set it to 0.7 or 0.8 and the model is selective, which lines up with recency — questions past the knowledge cutoff. Labenz’s first trial, at the two extremes, was the state of the World Series: without grounding the model’s latest was 2023; with search it went and got the live answer. He expects to start by always grounding when he might need it, then see how much of that decision he can delegate.

The other half of the release is receipts. Grounded answers link out to sources both inline and in a list at the bottom, plus search suggestions: example queries that would have retrieved similar content. Basu Mallick gives two reasons: they want to send traffic to the publisher ecosystem, and in testing users liked being able to validate or go deeper by clicking through. Kilpatrick zooms out. Early generative AI asked people to put a little trust in a model they had been told not to trust, with a murky line between those modes. The next inning, he says, is putting citations in developers’ hands so products can show that the material came not just from weights but from reputable sources.

Labenz’s snapshot of the field is that OpenAI has the o1 reasoning series for hard problems, Claude is alone in computer use for unstructured agentic work, Gemini still has by far the longest context windows and now native Google Search, and Meta has open weights. He is genuinely unsure whether capabilities will converge, as a platonic-representation story would predict, or keep diverging. Kilpatrick’s read of developer psychology is that someone ships a capability and the next question is when every other provider will have it. Google still has to keep the base layer developers expect. The net-different work is whatever leverages what Google is uniquely good at: large-scale infrastructure, which you already see in long context, and search.

Basu Mallick splits the roadmap into three buckets. First, learn from excellent work other players are doing. Second, bring what only Google can: search grounding, and in Vertex the ability to pass YouTube links, which they hope to bring to AI Studio. Third is the boring, important API work — error codes, support, rate limits, and a price such that no developer should be deterred from building. A tactical lesson from that first bucket is code execution. Other providers expose it as a toggle on the default endpoint, not a special API variant, and developers do not want a different version of the API for one feature. Search grounding follows the same rule: stay on the basic request most people already send, and turn it on by passing a tool.

They also choose, sometimes, not to match. Structured outputs let a developer specify a response schema. OpenAI’s version has all fields required by default, so every field comes back even if empty. Gemini’s treats fields as optional unless tagged required. That was the only outright confusion Labenz hit in an otherwise two-hour integration. Kilpatrick calls this one of the toughest balances: the best outcome is being different in a way that matches a developer’s mental model, and optional-plus-required, in his view, is closer to how JSON actually works. Basu Mallick remembers Michelle Pokrass of OpenAI describing a long internal debate on the other side, aimed at making sure developers got complete output. Gemini bet on the language’s default. Who is right, they both say, is something they will learn in the open.

Context caching is the mirror image: Google shipped first, Anthropic and OpenAI followed, and Google is now taking feedback on where its version should improve. The closing invitation is the same as the opening strategy. Keep sending feedback. Keep showing up. Kilpatrick hopes to catch Labenz at I/O next year and see what has landed. Perception, he has already said, is earned in years. The last six months are the first evidence that the earn is working.