Google Releases Stable Gemini 1.5 Pro-002 and Flash-002, Pro Cut in Price by More Than Half
Google has released stable Gemini 1.5 Pro-002 and Flash-002, updated builds that can be deployed in a paid product without an "experimental" label. According to Google's measurements, math benchmarks MATH and HiddenMath improved by about 20%, and Pro for requests up to 128 thousand tokens is more than twice as cheap. If your service runs on Gemini, recalculate your budget and test -002 on your own requests.
AI-processed from X: @GoogleDeepMind; edited by Hamidun News
Google DeepMind has released Gemini 1.5 Pro-002 and Flash-002. The "002" suffix denotes the second, updated build of each model. "Stable" means the models are no longer labeled experimental: you can build a service with real users and payments on them without worrying that the model will be changed or shut down unexpectedly. The second piece of news concerns money: Pro for requests up to 128 thousand tokens (long documents, contracts, reports, chunks of a codebase) is now more than twice as cheap. Below we break down what has changed and what to do about it.
What Changed in the Models Themselves
The new versions build on the experimental builds Google has shown over the past few weeks. According to the company, the main improvements concern three areas: understanding long context, working with images, and math. In practice, this means the model loses details in large documents less often, reads screenshots, diagrams and tables better, and solves calculation tasks more confidently.
Google cites figures from its own measurements. On the difficult general-knowledge benchmark MMLU-Pro, the score rose by about 7%, and on the math benchmarks MATH and HiddenMath by about 20%. Gains on vision tasks and code generation are more modest, around 2–7%. This is a noticeable step, but not a revolution. All numbers come from Google's internal comparisons, and on your own data the difference may turn out to be different.
There is one more detail: the answers have become shorter. For question-and-answer tasks, length decreased by about 5–20%. For chatbots this is a plus: the model rambles less and spends fewer tokens. But if you need detailed, elaborate texts, check whether the model has become too terse and, if necessary, adjust your prompt.
Prices, Limits and Speed
For businesses, the most welcome part is the price cut. For Gemini 1.5 Pro on requests of up to 128 thousand tokens, the rate is now $1.25 per million input tokens and $5 per million output tokens. Previously it cost $3.50 and $10.50 respectively, which is where "more than twice as cheap" comes from. The discount also applies to context caching, that is, scenarios where the same large document is sent to the model many times. Google tied the new prices to the beginning of October 2024.
Limits have also increased. Flash now accepts up to 2,000 requests per minute, and Pro up to 1,000. For high-traffic services, this means they hit rate limits noticeably later. In addition, the company reports roughly a twofold increase in generation speed and a threefold reduction in time to first token. For chat interfaces and voice assistants, this matters: users see the start of a response sooner.
What Developers Should Do
First, check which version your product runs on. If you call the model via the "latest" alias, its behavior may change on its own. If the version is pinned explicitly, the move to -002 should be made deliberately, after running your own test requests.
Second, recalculate costs. For applications that send large documents to Pro (contract, report and code analysis), the savings can be substantial. If you previously chose Flash for its price, compare again: the gap between Flash and Pro has narrowed.
Third, keep the limitations in mind. The math and code gains have so far been confirmed only by Google's own measurements, and how the model compares with the best competitor models remains an open question. The reduced rates apply to requests of up to 128 thousand tokens; for longer prompts, prices remain higher. Before moving production traffic, run your own set of tasks and compare quality, cost and latency between the old and new versions.
Overall, the release shows where the market is heading: labs compete not only on quality but also on price per token. For those building products on LLMs, this is good news: a model of the same tier now costs less, and the new limits make scaling easier.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What Changed in the Models Themselves?
The new versions build on the experimental builds Google has shown over the past few weeks. According to the company, the main improvements concern three areas: understanding long context, working with images, and math. In practice, this means the model loses details in large documents less often, reads screenshots, diagrams and tables better, and solves calculation tasks more confidently. Google cites figures from its own measurements.
What are the key details on Prices, Limits and Speed?
For businesses, the most welcome part is the price cut. For Gemini 1.5 Pro on requests of up to 128 thousand tokens, the rate is now $1.25 per million input tokens and $5 per million output tokens. Previously it cost $3.50 and $10.50 respectively, which is where "more than twice as cheap" comes from. The discount also applies to context caching, that is, scenarios where the same large document is sent to the model many times. Google tied the new prices to the beginning of October 2024. Limits have also increased.
What Developers Should Do?
First, check which version your product runs on. If you call the model via the "latest" alias, its behavior may change on its own. If the version is pinned explicitly, the move to -002 should be made deliberately, after running your own test requests. Second, recalculate costs. For applications that send large documents to Pro (contract, report and code analysis), the savings can be substantial. If you previously chose Flash for its price, compare again: the gap between Flash and Pro has narrowed.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.