Gemini 2.5 bills for reasoning tokens it does not fully show

A model that reasons by default, announced 25 March 2025

Google's announcement, Gemini 2.5: Our newest Gemini model with thinking, describes Gemini 2.5 Pro as 'a thinking model, designed to tackle increasingly complex problems' by reasoning through intermediate steps before producing a final answer, shipping on day one with a 1-million-token context window and a further 2-million-token window described as coming later. The post states the model tops the LMArena leaderboard 'by a significant margin' and leads on the GPQA and AIME 2025 benchmarks, adding the specific qualification that these results were reached 'without test-time techniques that increase cost, like majority voting' — a detail that matters because majority-vote sampling can inflate a benchmark score at the cost of running the same query many times.

What 'thinking budget' actually controls

Google's documentation, Gemini thinking, current as retrieved on 16 September 2026, explains that Gemini's thinking models engage in 'dynamic thinking by default, automatically adjusting the amount of reasoning effort based on the complexity of the request,' with a budget parameter available for a developer to cap that effort directly. The documentation states plainly that 'pricing is based on the full thought tokens the model needs to generate, despite only the summary being output from the API' — meaning a developer is billed for reasoning tokens they do not see in full, only as a shortened summary. That is a direct cost consequence of the same hidden-reasoning design other reasoning models use, stated here as a pricing mechanic rather than a safety rationale.

What a single benchmark table does not resolve

The announcement cites Humanity's Last Exam (18.8%, without external tools) and SWE-bench Verified (63.8%, with what the post describes as a custom agent setup) as headline figures, but a custom scaffold built specifically for one benchmark is part of what produced that score, not a property of the base model alone; a different harness or tool configuration on the same benchmark could move the number. A fuller technical report, Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities, followed the announcement by more than three months, itself a reminder that a launch post's benchmark table is typically the marketing-facing summary of results whose full methodology arrives, if at all, in a separate and later document.

• Was a cited benchmark score produced with a custom agent scaffold, and would the same scaffold be available in a production deployment?
• How much of the per-query cost is thinking tokens the API never fully discloses, versus visible output tokens?
• Does the announcement's 'without majority voting' qualification apply to every benchmark cited, or only some of them?

A reasoning-by-default model turns the cost and the evidentiary basis of a benchmark claim into two separate things to check: what the harness around the model contributed, and what the hidden reasoning tokens are actually billed at. Both are answered, where they are answered at all, outside the launch announcement itself.