Gemini 2.5 Pro: Full Spec Sheet (June 2026)

Gemini 2.5 Pro is Google DeepMind's flagship model in the Gemini 2.x family, released in March 2025. It succeeded Gemini 2.0 Pro (which itself replaced Gemini 1.5 Pro in late 2024) and brought three step-changes: native thinking mode (configurable reasoning budget per call), tier-2-quality vision matching GPT-5's vision benchmarks, and stable 1M-token context behavior with recall that holds across the full window.

What makes Gemini 2.5 Pro structurally different from GPT-5 or Claude Opus: it is natively multimodal across more modalities than either. Text, image, audio, video, and PDF inputs all flow through the same `contents` array. Pass an MP4 video file, an audio recording, a stack of PDFs and a free-form text question — Gemini accepts all of it in one call and reasons across them. GPT-5 supports text + image. Claude supports text + image. Only Gemini 2.5 Pro (and its Flash sibling) support video and audio natively in production.

Thinking mode (Google's name for configurable reasoning) is enabled by default on Gemini 2.5 Pro with a model-decided budget. Force a specific budget with `thinking_config: {thinking_budget: 5000}`; disable thinking entirely with `thinking_budget: 0` for fastest possible response. Thinking tokens bill at the output rate like reasoning tokens on GPT-5 and thinking tokens on Claude.

Gemini 2.5 Pro uses a unique tiered pricing model among frontier providers. Below 200,000 input tokens per call: $1.25/M input, $10/M output. Above 200,000 input tokens: $2.50/M input, $15/M output. The tier applies to the entire call — if you send 250K input tokens, the full 250K bills at the higher tier, not just the portion above 200K.

Worked: a 100K-token input + 1K output call bills `(0.100 × $1.25) + (0.001 × $10) = $0.125 + $0.01 = $0.135`. The same input as 250K + 1K output bills `(0.250 × $2.50) + (0.001 × $15) = $0.625 + $0.015 = $0.640`. Crossing the 200K threshold is a 2× input price + 1.5× output price step function, not a smooth ramp.

Implication: keep calls under 200K input where you can. If you're at 195K, padding to 205K to fit one more chunk is a 5× price increase. If you're going to cross 200K, go all the way — 250K and 500K bill at the same per-token rate.

Caching: explicit, via the `cachedContents` API endpoint. Pre-create a cached content block (1-hour default TTL, configurable up to 24 hours), reference it by ID in subsequent calls. Cached portion reads at 75% off ($0.31/M in the ≤200K tier, $0.625/M in the >200K tier). Largest cost lever on long-context workloads. Worked $ across providers: GPT/Claude/Gemini cost calculator.

Gemini 2.5 Pro accepts 1,000,000 tokens in a single call. For reference: a full feature-length screenplay is ~30K tokens, a 300-page novel is ~150K tokens, the codebase of a medium SaaS application is 200-500K tokens, the full Lord of the Rings trilogy is ~600K tokens. Gemini 2.5 Pro fits any of these in one call.

Recall holds across the full 1M window — Google's needle-in-haystack benchmarks show >99% recall through ~1M tokens for the Pro model. The practical bottleneck is cost and latency, not recall. A 1M-token call bills `1.0 × $2.50 + (output × $15) = $2.50+` per call before output costs and runs 30-60 seconds end-to-end at typical streaming rates.

When 1M actually matters: full-codebase reasoning (refactor planning across an entire repo), full-document Q&A on books or legal contracts, long-form audio/video understanding (transcribe and analyze a 1-hour meeting in one shot), multi-document research synthesis.

When 1M doesn't matter: classification, extraction, chat, structured-data tasks, anything that fits in 50K tokens with RAG. For most production workloads, the smaller ≤200K tier on Gemini 2.5 Pro (or even Gemini 2.5 Flash at $0.30/M) is the right pick.

Pass a video file as an inline base64 blob or via the Google File API (recommended for files >20MB). Gemini extracts frames at 1 FPS by default, transcribes audio, and reasons across the combined stream. Token accounting: video bills at ~258 tokens per second of footage (compressed across video + audio).

A 5-minute video bills `300 seconds × 258 tokens = 77,400 tokens` — still under the 200K tier. A 30-minute meeting recording is ~465K tokens — over the 200K threshold, into the >200K tier. Plan accordingly.

PDFs are processed page-by-page with image + text extraction. A typical text-heavy PDF page is ~258 tokens; image-heavy pages can be substantially more. Pass via the File API for documents over 20MB.

Audio (without video) bills at ~32 tokens per second. A 10-minute audio recording is ~19,200 tokens. Use for meeting transcription + summarization, podcast analysis, voice-note structuring.

Practical caveat: multimodal inputs inflate token counts fast. A naive 'summarize this hour-long meeting' call can easily exceed 200K tokens (1 hour video = ~930K tokens). Cache the video via the File API and reference it across multiple analytical calls instead of re-uploading.

Gemini 2.5 Pro supports JSON Schema function calling: declare functions in the `tools` parameter, the model picks one (or several in parallel) and returns the arguments. Parallel function calling is supported and on by default for the Pro model.

Structured outputs are first-class: pass a `responseSchema` (JSON Schema subset) in `generationConfig` and Google guarantees the model's output validates against that schema. Supports nested objects, arrays, enums — comparable to OpenAI's structured outputs and Anthropic's tool-use-as-output pattern.

Built-in tools that you don't have to implement yourself: **code execution** (the model writes and runs Python in a sandbox, sees the output, iterates), **Google Search grounding** (the model issues a Google Search and cites the results in its response, with attribution), **URL context** (the model fetches and reads URLs in the conversation). Built-in tools are unique to Gemini in the frontier-model menu and dramatically reduce orchestration code for agentic workflows.

Thinking mode is enabled by default on Gemini 2.5 Pro with a model-decided thinking budget. Override with `thinking_config: {thinking_budget: N}` where N is the maximum thinking tokens for the call. Set N=0 to disable thinking entirely; set N=-1 (or very high) to let the model decide dynamically.

Thinking tokens bill at the output rate (same as reasoning tokens on GPT-5 and thinking tokens on Claude). On Gemini 2.5 Pro: `$10/M` in the ≤200K tier, `$15/M` in the >200K tier. A 3,000-token thinking budget on a ≤200K call adds $0.03 to the call.

When to set explicit thinking budgets: cost control (cap at 1,000 for routine tasks), quality control (boost to 5,000-10,000 for complex reasoning), latency control (set to 0 for fastest possible response on simple tasks). When to leave thinking on auto: general-purpose chat where Gemini's calibration is well-tuned for the task mix.

**Pick Gemini 2.5 Pro** when you need native multimodal across video/audio/PDF, when you need 1M context in general availability, when built-in tools (code execution, Search grounding) replace custom orchestration, or when you're already in the Google Cloud / Workspace ecosystem and Vertex AI billing simplifies procurement.

**Pick Claude Opus 4.7** when long-form writing voice, refusal-calibration discipline, or hard reasoning is the bottleneck. Opus is $15/$75 vs Gemini 2.5 Pro's $1.25/$10 (≤200K) — Gemini is dramatically cheaper for everything except the narrow tasks where Opus's quality premium pencils.

**Pick GPT-5** when you need 400K context without the >200K tier bump on Gemini, when you're in the OpenAI tooling ecosystem (Responses API, Assistants, ChatGPT Pro), or when structured outputs with the most-mature JSON Schema enforcement matter.

Cross-vendor head-to-head: GPT-4o vs Gemini 2.5 Pro.

Every number on this page was verified against Google's live documentation on 2026-06-20. Sources: ai.google.dev/gemini-api/docs/models/gemini for context, modalities, and feature support; ai.google.dev/pricing for AI Studio direct pricing; cloud.google.com/vertex-ai/generative-ai/pricing for Vertex AI pricing (currently identical to AI Studio direct).

Google's pricing updates are announced via the Vertex AI release notes and the ai.google.dev changelog. Prices have moved twice on Gemini 2.5 Pro since launch (both downward). Re-verify quarterly if your bill is significant.

Methodology: when a number could not be cross-confirmed against an official Google page on the verification date, it was omitted from this card rather than guessed.