6月24日晚,一张截图在X上疯狂刷屏。
画面极具冲击力——深色界面中央,一行浅紫色大字赫然写着「Gemini 2.5 — FREE 1M Tokens/Min」。下方五张卡片一字排开:1M TOKENS PER MINUTE、FREE、NO CREDIT CARD、∞ MULTIMODAL INPUT、0× COST MULTIPLIER。
发帖者@nahid_pro09的文案更猛:
"1 million tokens/min completely free on gemini 2.5 flash and pro - official google api, zero strings attached😳"
「每分钟100万tokens,Google官方API,零附加条件。」
428个赞,554次收藏,5万+浏览。评论区很快热闹起来。




▲ @nahid_pro09 的病毒推文:Gemini 2.5 Flash/Pro 每分钟100万tokens免费,配图极具官方感
这条帖子踩中了开发者的集体痛点。
OpenAI的API烧钱。Anthropic的API更烧钱。个人开发者、独立黑客、学生党——每个月在API调用上花掉几百美元是常态。突然有人说Google白送每分钟100万tokens,还不要信用卡?
收藏。转发。生成API Key。先薅为敬。
但问题来了:这个数字,是真的吗?
Grok冲进来泼了盆冷水
推文发出几小时后,Grok官方账号直接在评论区回复了。
"Yes, the Gemini free tier in Google AI Studio is real—no credit card needed and solid for testing. But the '1M tokens/min on 2.5 Flash & Pro' claim is overstated now. Limits got reduced (often lower RPM/TPM + daily caps); check your exact quotas in AI Studio after grabbing a key."
「免费阶层确实存在,无需信用卡,测试够用。但"1M tokens/min"的说法被夸大了。限额已经被调低(通常是更低的RPM/TPM加上每日上限)。拿到Key之后,自己去AI Studio查看实时配额。」

▲ Grok官方回复:免费阶层真实存在,但1M/min的数字夸大了,建议开发者自己去查实时配额
Grok说得相当克制。但本质就一个意思:截图里的数字,不代表你能用到。
评论区另一个用户@voy_aj更直接:
"Stop spreading fake news, you only get 20 request per day for 2.5 flash"
「别传假消息了,2.5 Flash每天只有20次请求。」
一场围绕"截图vs现实"的争论就此展开。
官方文档里到底写了什么?
翻遍Google AI的官方Rate Limits页面(https://ai.google.dev/gemini-api/docs/rate-limits),你会发现一个关键事实:
Google没有在任何地方承诺Free Tier的固定RPM/TPM数字。
官方页面上写得明明白白:速率限制的计算单位是项目(per project),和API Key没有直接绑定关系;而且——
"Specified rate limits are not guaranteed and actual capacity may vary."
「标称的速率限制不做保证,实际容量可能变化。」
▲ Google官方Rate Limits文档:限额按项目计算,不做保证,指向AI Studio实时查看
整个页面反复强调的只有一点:去AI Studio查看你的实时配额。
也就是说,@nahid_pro09截图里那个"Rate Limit: 1M/min",可能是某个特定账户在某个特定时刻抓到的数字——可能是老账户、区域差异、预览模型、甚至是UI历史状态残留。它没有被官方文档背书,也不代表新注册用户的体验。
社区追踪的数字差了多少?
独立指南网站Pecollective长期追踪Gemini免费阶层的实际限额。他们给出了一组完全不同的数据:
- Gemini 2.5 Flash 免费:10 RPM、250,000 TPM、1,500 RPD
- Gemini 2.5 Pro 免费:5 RPM、150,000 TPM、50 RPD
- Gemini 2.0 Flash(上一代):15 RPM、1,000,000 TPM
▲ Pecollective的社区追踪数据:2.5 Flash实际免费TPM为25万,与热传截图差距明显
注意看:2.5 Flash的真实TPM只有25万,距离热传的100万差了整整四倍。2.5 Pro更是只有15万。
而且Pecollective特别强调了一句:"10 RPM是大多数人首先撞到的天花板。"每分钟只能发10次请求——这意味着即使每个请求用满,实际吞吐也远到不了百万量级。
病毒截图里的"1M/min",很可能是把上下文窗口(Context Window)的1M tokens和速率限额(Rate Limit)的TPM混为一谈了。上下文窗口决定单次请求能塞多少内容,速率限额决定每分钟能跑多少总量。两者不是一个东西。
免费就是免费——但别指望"无限吞吐"
Google免费阶层本身仍有价值。2026年,它依然是开发者能拿到的最大方的AI API福利之一。
不需要信用卡。1M上下文窗口。多模态原生支持(图像、视频、音频、PDF直接喂)。30秒生成API Key。
对于个人项目、每日定时任务、原型验证、学习实验,这已经远超够用。做个简单的算术:10 RPM × 每次5k tokens × 60分钟 × 24小时,每天的理论上限接近7000万tokens——前提是你能把RPM打满且不触发RPD限流。
现实中的瓶颈在哪?
RPD(每日请求上限)才是最容易被忽略的杀手。Flash免费阶层的1500 RPD,意味着你一天最多只能发起1500次请求。很多开发者兴冲冲写好循环脚本,跑了几分钟就撞墙。
并发限制也很关键。Agent场景下,一个任务可能fan-out成5-10次模型调用。10 RPM在burst时瞬间打满,后面的请求全部排队。
更好的用法是分层调度:日常高频走Flash或Flash-Lite,真正需要深度推理时才切Pro。配合指数退避(1s→2s→4s)、结果缓存、并发控制,把有限的免费额度用到极致。
路由器的"乘数效应"才是隐藏大招
@nahid_pro09的上一帖其实更有信息量——他展示了一个叫FreeLLMAPI的开源路由器。




▲ FreeLLMAPI路由器仪表盘:聚合多家免费模型Key,月token预算超13亿
原理很直白:把Gemini、DeepSeek、Groq、Qwen、Llama等多家的免费API Key聚合到一个OpenAI兼容端点里,自动做智能路由和fallback。
单个Gemini Key的限额一般,但堆上五六个不同提供商的免费Key,每月轻松过亿tokens。对个人开发者来说,这已经接近"无限使用"的体验。
Grok在回复里也肯定了这一点:"The free LLM router repo is legit. It stacks keys from many providers into one OpenAI-compatible endpoint with smart routing and fallbacks, easily hitting 1B+ tokens/month total for personal use."
谷歌在下一盘更大的棋。免费阶层越慷慨,开发者越依赖Gemini SDK和生态。这和当年Firebase的策略高度一致——先用免费拉你入坑,等你产品做大了,迁移成本也高了。
对于想认真做产品的开发者,记住三件事:
- 拿到Key后马上去 https://aistudio.google.com/rate-limit 查实时配额,别信任何人的截图。
- 日常用Flash,难题用Pro。Pro免费阶层的RPD极低(50次/天),别浪费在简单任务上。
- 别把所有鸡蛋放一个篮子里。免费限额随时可能调整,关键路径至少准备一个付费降级方案。
一张截图点燃了开发者的贪便宜本能。Google免费阶层的价值货真价实——前提是你知道它真正的边界在哪。
On the evening of June 24, a screenshot went viral on X.
The image was striking—in the center of a dark interface, a line of light purple text read 「Gemini 2.5 — FREE 1M Tokens/Min」. Below it, five cards were laid out in a row: 1M TOKENS PER MINUTE, FREE, NO CREDIT CARD, ∞ MULTIMODAL INPUT, 0× COST MULTIPLIER.
The poster @nahid_pro09's caption was even bolder:
"1 million tokens/min completely free on gemini 2.5 flash and pro - official google api, zero strings attached😳"
"1 million tokens per minute, official Google API, zero strings attached."
428 likes, 554 bookmarks, 50K+ views. The comments section heated up quickly.
▲ @nahid_pro09's viral tweet: Gemini 2.5 Flash/Pro at 1 million tokens per minute for free, with an image that looked extremely official
This post hit a collective pain point for developers.
OpenAI's API burns money. Anthropic's API burns even more. For independent developers, indie hackers, and students, spending hundreds of dollars a month on API calls is the norm. Then suddenly someone says Google is giving away 1 million tokens per minute for free, no credit card required?
Bookmark. Share. Generate an API Key. Get the free stuff first.
But here's the question: is that number even real?
Grok jumped in and poured cold water on it
A few hours after the tweet went out, Grok's official account replied directly in the comments.
"Yes, the Gemini free tier in Google AI Studio is real—no credit card needed and solid for testing. But the '1M tokens/min on 2.5 Flash & Pro' claim is overstated now. Limits got reduced (often lower RPM/TPM + daily caps); check your exact quotas in AI Studio after grabbing a key."
"The free tier does exist, no credit card needed, and it's good enough for testing. But the '1M tokens/min' claim is exaggerated. The limits have been lowered (usually lower RPM/TPM plus daily caps). After you get your key, go check your real-time quota in AI Studio yourself."
▲ Grok's official reply: the free tier is real, but the 1M/min figure is exaggerated, and developers are advised to check their real-time quotas themselves
Grok put it quite diplomatically. But the essence was clear: the numbers in the screenshot don't represent what you'll actually get.
Another user in the comments, @voy_aj, was more direct:
"Stop spreading fake news, you only get 20 request per day for 2.5 flash"
"Stop spreading fake news—2.5 Flash only gets 20 requests per day."
A debate about "screenshot vs. reality" broke out from there.
What exactly does the official documentation say?
Going through Google AI's official Rate Limits page (https://ai.google.dev/gemini-api/docs/rate-limits), you'll find one key fact:
Google has not promised a fixed RPM/TPM number for the Free Tier anywhere.
The official page states it plainly: rate limits are calculated per project, not directly tied to an API Key; and—
"Specified rate limits are not guaranteed and actual capacity may vary."
"The stated rate limits are not guaranteed, and actual capacity may vary."
▲ Google's official Rate Limits documentation: limits are calculated per project, not guaranteed, and users are directed to check real-time quotas in AI Studio
The entire page repeats only one point over and over: go check your real-time quota in AI Studio.
In other words, the "Rate Limit: 1M/min" in @nahid_pro09's screenshot may have been a number captured from a specific account at a specific moment—possibly an old account, regional differences, a preview model, or even a leftover state in the UI history. It is not backed by official documentation, nor does it represent the experience of newly registered users.
How far off are the community-tracked numbers?
The independent guide site Pecollective has been tracking the actual limits of Gemini's free tier for a long time. They've published a completely different set of numbers:
Gemini 2.5 Flash free: 10 RPM, 250,000 TPM, 1,500 RPD
Gemini 2.5 Pro free: 5 RPM, 150,000 TPM, 50 RPD
Gemini 2.0 Flash (previous generation): 15 RPM, 1,000,000 TPM
▲ Pecollective's community-tracked data: 2.5 Flash's actual free TPM is 250K, a clear gap from the widely shared screenshot
Note this: 2.5 Flash's real TPM is only 250K—a full four times less than the viral 1M. 2.5 Pro is even lower, at just 150K.
And Pecollective specifically emphasized: "10 RPM is the ceiling most people hit first." You can only send 10 requests per minute—which means even if every request is maxed out, actual throughput will never reach the million scale.
The "1M/min" in the viral screenshot was very likely a mix-up between the 1M-token context window and the TPM rate limit. The context window determines how much content a single request can hold; the rate limit determines how much total volume can run per minute. The two are not the same thing.
Free is free—but don't expect "unlimited throughput"
Google's free tier still has real value. In 2026, it remains one of the most generous AI API benefits a developer can get.
No credit card required. 1M context window. Native multimodal support (images, video, audio, and PDFs can be fed directly). API key generated in 30 seconds.
For personal projects, daily scheduled tasks, prototype validation, and learning experiments, this is already far more than enough. Do some simple math: 10 RPM × 5k tokens per request × 60 minutes × 24 hours gives a theoretical daily ceiling of nearly 70 million tokens—provided you can saturate the RPM without triggering the RPD limit.
Where are the real-world bottlenecks?
RPD (daily request limit) is the killer that's easiest to overlook. The Flash free tier's 1,500 RPD means you can make at most 1,500 requests per day. Many developers excitedly write their loop scripts, only to hit a wall after a few minutes of running.
Concurrency limits are also critical. In agent scenarios, a single task can fan out into 5–10 model calls. At 10 RPM, the limit is hit instantly during bursts, and all subsequent requests are queued.
A better approach is tiered scheduling: use Flash or Flash-Lite for high-frequency daily tasks, and only switch to Pro when you truly need deep reasoning. Combine exponential backoff (1s→2s→4s), result caching, and concurrency control to squeeze the most out of the limited free quota.
The router's "multiplier effect" is the real hidden ace
@nahid_pro09's previous post was actually more informative—he showcased an open-source router called FreeLLMAPI.
▲ FreeLLMAPI router dashboard: aggregates free model keys from multiple providers, with a monthly token budget exceeding 1.3 billion
The principle is straightforward: aggregate free API keys from Gemini, DeepSeek, Groq, Qwen, Llama, and others into a single OpenAI-compatible endpoint, with automatic smart routing and fallback.
A single Gemini key's limits are mediocre, but stack up five or six free keys from different providers and you easily clear 100 million tokens per month. For an individual developer, that's close to an "unlimited usage" experience.
Grok also confirmed this in its reply: "The free LLM router repo is legit. It stacks keys from many providers into one OpenAI-compatible endpoint with smart routing and fallbacks, easily hitting 1B+ tokens/month total for personal use."
Google is playing a much bigger game. The more generous the free tier, the more developers come to rely on the Gemini SDK and ecosystem. This is highly consistent with the Firebase strategy of years past—hook them with free first, and once their product scales, the switching cost is already high.
For developers who seriously want to build products, remember three things:
After getting your key, immediately check your real-time quota at https://aistudio.google.com/rate-limit —don't trust anyone's screenshot.
Use Flash for daily tasks, and Pro for hard problems. The Pro free tier's RPD is extremely low (50/day), so don't waste it on simple tasks.
Don't put all your eggs in one basket. Free limits can be adjusted at any time; for critical paths, have at least one paid fallback plan ready.
A single screenshot ignited developers' instinct to grab a bargain. The value of Google's free tier is genuinely real—provided you know where its true boundaries lie.