Excelrs
Gemini 2.5 Pro: think, then use a million tokens
Launched 25 March 2025. Thinking plus a long multimodal desk. Arena rank is taste. Three chunks still beat a dump.
Sandeep Jaiswal·25 March 2025 · 5 min read
On 25 March 2025 Google DeepMind launched Gemini 2.5 Pro Experimental. The specialty is a thinking model that still carries Gemini’s old strengths: native pictures, audio, video, and a one-million-token window.
This post is what they shipped that day, which scores they published, and when a million tokens is a tool — and when it is stuffing.
What launched
They called 2.5 a thinking family: the model reasons through thoughts before it answers. 2.5 Pro Experimental was the first. It was in Google AI Studio and in the Gemini app for Gemini Advanced that day. Stable 2.5 Pro came later, on 17 June 2025. Date this post to the first public drop.
They said it debuted at number one on LMArena, a human-preference board. Preference is not an exam. It means people liked the style that week.
They kept native multimodality and shipped a 1 million token context (they said 2 million was coming). That is a long desk: a repo, a lecture set, a mixed pack of slides and audio.
A million tokens is a desk, not a brain. You still choose what sits on it.
What the specialty is
Two things at once: think, and hold a lot.
If the job is “this whole course folder, then a short cited answer,” this is the launch that matches. If the job is “one photo of a sheet,” GPT-4o already did that. If the job is “one contest item,” o1 and R1 are the closer cousins.
Their coding story at launch: visually heavy web apps, edits, and 63.8% on SWE-bench Verified with a custom agent setup. Custom means their scaffold, not a raw box. Treat it that way.
They also said 18.8% on Humanity’s Last Exam without tool use — a hard set where most models sit low. Leading a hard set at 19% still means “mostly miss.” Do not sell it as a tutor that knows everything.
Numbers they published
| Claim (Google, 25 March 2025) | How to read it |
|---|---|
| #1 on LMArena at debut | People preferred the writing that week |
| Leads AIME 2025 and GPQA in their write-up, no majority vote | Thinking helped exams |
| 18.8% Humanity’s Last Exam, no tools | Hard set; still mostly wrong |
| 63.8% SWE-bench Verified, custom agent | Scaffold included |
| 1 million token window | You can pass a repo if you must |
Your five cases still win.
Walk through once
You have a 40-page attendance handbook and one question: “What is the 75% rule?”
Wrong use: paste all 40 pages into 2.5 Pro and ask for a friendly essay. You will get a fluent page and a weak cite.
Right use: chunk, retrieve two cards, then let the thinking model draft from those cards. The million-token window is a backup when retrieve is unsure and you must stuff a neighbour chapter — not the default.
A second job: “Here is the whole repo. Where is attendance calculated?” A million tokens can take the repo. You still open the file it names. If it names a file that is not there, fail.
A bait: empty retrieve plus “think harder.” Thinking on silence invents a Thursday. Stop first.
What good looks like
The long window is in the spec as a ceiling, not the default payload.
A lecture plus slides is one job. A random film is not in the call.
You can point at the file the model named. If you cannot, you fail the run, even if the prose is nice.
LMArena is a footnote. Your five cases are the score.
The tape is space. The bubble is extra spend. Fill neither by habit.
When the window hurts
A million tokens costs money and attention. The model can “see” a line and still miss it. That is why we already wrote about stuffing.
If three chunks answer the question, send three. Save the window for “I truly cannot cut this video plus this repo today.”
Audio and video in the same call are the Gemini habit. A lecture recording plus the slide PDF is on-topic. A random film is not.
Common mistakes
Pasting the handbook because the window allows it. You will not know which line won.
Reading LMArena as an exam. It is a taste vote.
Quoting 63.8% without “custom agent.” The scaffold is part of the score.
Treating 18.8% as a product. Most hard items still fail.
Skipping retrieve because thinking feels serious. Serious and grounded are different.
Practise
- Same question, three chunks, thinking on. Cite must open.
- Same question, whole handbook pasted. Compare. Keep the shorter input if both pass.
- A repo question. Demand a path you can open.
- Empty input. Confirm stop.
- Write the token count in the trace.
If the thin tray passes, the stuffed tray is a luxury you can drop.
Checklist
- You can date it: 25 March 2025 for the experimental Pro.
- Specialty is think plus a long multimodal desk.
- Arena rank is preference, not an exam.
- A million tokens is optional, not the plan.
- Cites still open.
The next step is three chunks with thinking on, not a dump of the course zip.