Gemini 2.5 Pro arrives as a thinking model on a staged platform rollout

What Google shipped and what it withheld

Gemini 2.5 Pro Experimental launched on March 25, 2025, as an experimental release. It is available immediately in Google AI Studio and in the Gemini app for Gemini Advanced users, who can pick it from a model dropdown on desktop and mobile.

Two things are explicitly not ready yet. Vertex AI availability is "coming weeks," and pricing — along with the higher rate limits needed for scaled production use — is also "coming weeks." That means the model you can benchmark today is not yet the model you can budget and deploy against at scale.

The framing matters: this is an experimental version of 2.5 Pro. Google is inviting feedback while iterating, which is a different commitment than a stable, priced production endpoint.

The benchmark claims, read carefully

Google leads with LMArena, which measures human preference rankings, where 2.5 Pro debuts at #1 by what the company calls a significant margin. That signals style and response quality more than raw task accuracy.

On harder reasoning evals, Google is careful to note conditions. It claims leads on GPQA and AIME 2025 "without test-time techniques that increase cost, like majority voting" — a meaningful qualifier, since majority voting inflates scores at the expense of inference cost. On Humanity's Last Exam it reports 18.8% "across models without tool use."

For coding, the headline number is 63.8% on SWE-Bench Verified, but with an important caveat: it was achieved "with a custom agent setup." That is not the base model in isolation; it reflects an agent scaffold Google built around it. Teams reproducing agentic results will need their own harness.

Reasoning moves into the base model

The technical claim underneath the benchmarks is that Gemini 2.5 combines a significantly enhanced base model with improved post-training, rather than bolting reasoning on as a separate mode. Google's earlier Gemini 2.0 Flash Thinking was its first thinking model; 2.5 is the follow-through.

Going forward, we're building these thinking capabilities directly into all of our models, so they can handle more complex problems and support even more capable, context-aware agents.Montana Labs

That is a platform-level decision. If thinking becomes default across the Gemini line, developers lose the clean distinction between fast, cheap models and slow, reasoning models — and will need to reason about latency and cost differently for every model, not just a designated "thinking" variant.

Context window and multimodality as the platform differentiators

2.5 Pro ships with a 1 million token context window, with 2 million described as coming soon, and retains native multimodality across text, audio, images, video and entire code repositories. Google positions this as the ability to comprehend vast datasets and pull from multiple information sources in one pass.

For applied work, this is the more durable claim than any single leaderboard position. Whole-repository ingestion and mixed-media inputs change what you can attempt without building retrieval plumbing yourself.

The gap between demo and deployment is the thing to manage

The specific implication of this release is timing. The model is strong enough to evaluate now in AI Studio and the Gemini app, but the two levers production teams depend on — Vertex AI hosting and published pricing with production rate limits — are both deferred to an unspecified "coming weeks."

Treat this window as an evaluation phase, not a migration phase. Validate the reasoning and coding claims against your own tasks now, note that the SWE-Bench figure assumes a custom agent you would have to build, and hold deployment decisions until pricing and Vertex availability let you model real cost and reliability.