Gemini 2.5 Pro hands-on: a very strong model with 1M input tokens, 64K output tokens, a January 2025 knowledge cut-off, and very, very impressive coding skills
There's a new release from Google Gemini this morning: the first in the Gemini 2.5 series. Google call it “a thinking model …
Simon Willison's WeblogSimon Willison
Context & Ripple Effects
Gemini’s long-context strategy predates this release: Google introduced Gemini 1.5 to developers with a 1M-token context window and later put Gemini 1.5 Pro into Vertex AI preview, including planned Code Assist use. Gemini 2.5 Pro extends that arc with a much larger stated output allowance and a newer knowledge cutoff.
The release also follows positive early reception for Gemini 2.0 Flash’s spatial reasoning and streaming interface. The important change is not context length alone, but the combination of large-context handling with a hands-on report of notably strong coding performance.
First-order effects
Developers evaluating Gemini now have a model positioned for tasks that require both very large inputs and long generated outputs, with coding a particularly salient use case.
Google gains a stronger flagship for its Gemini 2.5 line, while the model’s January 2025 cutoff makes recency a defined constraint for users assessing its answers.
Second-order effects
Coding-assistant and enterprise buyers will increasingly test models on end-to-end repository and document workflows rather than isolated benchmark prompts, favoring products that can manage large supplied context reliably.
Rival frontier-model providers face added pressure to pair long-context claims with credible coding quality and usable output limits; context size by itself has already become less differentiating.
Third-order effects
If these capabilities hold up in production, frontier competition will shift toward allocating inference and product design around long-running, context-heavy knowledge work—not merely maximizing a model’s headline context window.
The likely enduring constraint is economic: larger inputs and outputs make context a compute budget, so providers will need to distinguish premium, high-context workflows from broadly available AI features.
The trend: Frontier AI products are converging on long-context, reasoning-oriented coding systems whose practical value depends on how efficiently they turn large private corpora into usable work.
I wrote up some first impressions here, I'm particularly impressed by audio transcriptions and bounding box recognition on images simonwillison.net/2025/Mar/25/ ...
@artificialanlys
@artificialanlys on x
Google's new Gemini 2.5 Pro Experimental takes the #1 position across a range of our evaluations that we have run independently Gemini 2.5 Pro is a reasoning model, it ‘thinks’ before answering questions. Google has released it as an experimental API in AI Studio only, and has [i…
@maxwinebach
Max Weinbach on x
Google now has the best model, hands down. Assuming they keep the same pricing as Gemini 1.5 Pro, this will be another Deepseek R1 moment. Cheap, available, fast, and intelligent. Also, it's all TPUs!
@beinghamed
Hamed on x
Gemini 2.5 pro vs DeepSeek-V3-0324 Prompt : make a fully working chess game in html in one file Gemini 2.5 pro generated 570 lines of code. Deepseek V3 (guess how much).. a whooping 2372 lines of code 🔥 [video]
@emollick
Ethan Mollick on x
Gemini 2.5 Pro is a very good model, seems like a real step forward, in both metrics and practical use. I think because it is labelled 2.5 and was sort of quietly rolled out, people may miss how big a jump it is, but discussions are making me think others are feeling similarly.
@maxwinebach
Max Weinbach on x
It's sorta wild Gemini 2.5 Pro is the first model that's actually able to coherently look at long context and actually understand it without it just being all needle in a haystack tests
@levie
Aaron Levie on x
Google's Gemini 2.5 update is not screwing around. Incredible how quickly we're seeing new levels of capability be achieved in AI. Nothing seems to be slowing down. [image]
@simonw
Simon Willison on x
Notes on putting the new Gemini 2.5 Pro through its paces - I'm impressed: it did great on image recognition, audio transcription and returning bounding boxes for creatures in a complex photograph, plus it rendered a solid SVG of a pelican on a bicycle! https://simonwillison.net/…
@simonw
Simon Willison on x
Added a new section about Gemini 2.5 Pro for code: I'm VERY impressed after my coding experiments so far, it's tackling some sophisticated problems that start with 100,000+ tokens of existing codebase https://simonwillison.net/...
@simonw
Simon Willison on x
Updated my blog post to note a couple of extra things I had missed: the output token limit of 64,000 is a huge upgrade from the 8,192 of previous Geminis, and the knowledge cut-off date of January 2025 is an improvement on Gemini 2.0's cut-off date of August 2024
@amperlycom
@amperlycom on x
Yay! I made a thing. 3D game with python and Gemini 2.5 pro [video]
@alexanderchen
Alexander Chen on x
Gemini 2.5 Pro codes a TV. 📺 With just one prompt, it came up with its own ideas for 10 channels and coded each one! Link + details in 🧵 [video]
@officiallogank
Logan Kilpatrick on x
So, what have you done with Gemini 2.5 Pro today? Please show impressive things only 🧵
@ai_for_success
AshutoshShrivastava on x
Gemini 2.5 Pro Experimental is the only mode capable of performing multiple reasoning (thinking) steps within the same response cycle. I hadn't seen this before. This is definitely new in Gemini 2.5 Pro Experimental. Anyone else noticed this? You can see two distinct thinking [vi…
@mrdoob
@mrdoob on x
Gemini 2.5 Pro really is impressive... “Can you create a simple 3d car simulator with Three.js in a single HTML? Please add clouds, mountains, a road, some trees and a train going around. Make sure it works on mobile.” [video]
@kellynewso
Newman on x
Honestly, I haven't seen this kind of reasoning pattern before. Gemini 2.5 Pro Experimental is thinking in stages — analyzing the input, then extending it with deeper logic. It's starting to behave more like a thinker than a responder. [video]
@matthewberman
@matthewberman on x
Gemini 2.5 Pro is insane at coding. It's far better than anything else I've tested. Below are ONE SHOT demos 🧵 [image]
@_philschmid
Philipp Schmid on x
Combining the Youtube link feature + Gemini 2.5 Pro is insane. 🔥 Response 83s for a 30 min long video (500 000 tokens). Still watched it, nice video @theo! [image]