Google folds thinking budgets, MCP, and computer use into the Gemini API

What the 2.5 Pro and Flash numbers actually claim

Google says the updated 2.5 Pro now leads the WebDev Arena coding leaderboard with an ELO of 1415 and tops the LMArena human-preference boards, while retaining its 1 million-token context window. It also says 2.5 Pro, after incorporating the LearnLM education models, was preferred by educators over other models and outperformed top models on all five of the learning-science principles Google uses to build tutoring systems.

The Flash story is about cost, not leadership. Google describes 2.5 Flash as its efficiency workhorse and says the new version uses 20-30% fewer tokens in its own evaluations while improving on reasoning, multimodality, code, and long-context benchmarks. For teams that pay per token, a same-or-better model that consumes fewer tokens is a direct margin change.

Deep Think ships to trusted testers, not everyone

Deep Think is an experimental reasoning mode for 2.5 Pro that Google says considers multiple hypotheses before responding. It cites a score on the 2025 USAMO — described as one of the hardest math benchmarks — a lead on LiveCodeBench for competition-level coding, and 84.0% on MMMU for multimodal reasoning.

Notably, Google is gating access. Rather than a broad rollout, Deep Think goes first to trusted testers via the Gemini API while Google runs additional frontier safety evaluations.

Because we're defining the frontier with 2.5 Pro DeepThink, we're taking extra time to conduct more frontier safety evaluations and get further input from safety experts.Montana Labs

For applied teams, that means the highest-reasoning tier is not something to plan production work around yet — it is a preview channel with an explicit safety hold.

The developer-experience changes carry more practical weight

Three API-level additions matter more day-to-day than the leaderboard positions. Google is extending thinking budgets — already on Flash — to 2.5 Pro, letting developers cap the tokens a model spends reasoning, or turn thinking off entirely. That converts reasoning from a fixed behavior into a cost-and-latency dial.

Thought summaries now appear in the Gemini API and Vertex AI, organizing the model's raw thoughts into headers, key details, and notes on when it used tools. Google frames this as an aid for debugging, which is the honest use: understanding why an agentic call went the way it did.

And native SDK support for Model Context Protocol definitions lands in the Gemini API, easing integration with open-source tools. Adopting MCP means Google is building against a shared tool-calling standard rather than a Gemini-only interface — a meaningful choice for teams that don't want to rewire integrations per vendor.

Agentic capabilities arrive alongside a prompt-injection defense

Google is bringing Project Mariner's computer use into the Gemini API and Vertex AI, with Automation Anywhere, UiPath, Browserbase, Autotab, The Interaction Company, and Cartwheel named as early explorers, and a broader developer rollout planned for the summer. The Live API also gains audio-visual input and native audio-out dialogue, plus features like Proactive Audio and Thinking in the Live API.

Paired with those agentic features is a security claim that belongs in the same conversation: Google says its new approach significantly raised Gemini's protection rate against indirect prompt injection during tool use, calling 2.5 its most secure model family to date. That pairing is the point — computer use and tool retrieval are exactly where injected instructions become dangerous.

The implication: Gemini 2.5 is being packaged as a controllable agent platform

Read together, thinking budgets, thought summaries, MCP support, computer use, and the injection defense describe a platform being tuned for developers building agents, not just chat. The controls answer the practical questions of running reasoning models in production: how much do I pay to think, can I see why it acted, does it speak a standard tool protocol, and will retrieved data hijack it?

The availability timeline reinforces where Google wants attention. Updated 2.5 Flash reaches general availability in Google AI Studio and Vertex AI in early June, with 2.5 Pro following soon after — while Deep Think stays behind a safety gate. The frontier reasoning mode is the headline, but the shippable substance is the cheaper, more controllable Flash and the API plumbing around it.