Gemini 3 Pro, reviewed

Brilliant. That is the word for Gemini 3 Pro on one specific job, and it overstates the model badly everywhere else. Google's own positioning, the public benchmark record, and the consistent community discussion all point the same direction.

Gemini 3 Pro is best understood as a complement to Claude Opus 4.7, not a swap for it. The natural fit is anything that touches an image or a Google Workspace document, especially when the rest of the stack already runs Anthropic or OpenAI. Google's own product material and pricing structure both point to a vision-first role in a multi-model stack.

The headline result is narrow and consistent. On tasks that combine vision and reasoning — read a dashboard screenshot and explain what is broken, parse a hand-annotated PDF, turn a whiteboard sketch into a structured description — Gemini 3 Pro leads the field by a clear margin, not a hair. Almost every other category is more even, and a refusal pattern runs across persona-taking and speculative prompts that no amount of prompt engineering fully fixes.

The vision-first architecture Google emphasizes in DeepMind's Gemini overview shows up in practice exactly where you would expect. On text-only work it doesn't try to beat Claude; it earns its place on a different pass of the pipeline entirely.

Where vision-plus-reasoning lands

The category Google sells the hardest is image-plus-reasoning, and the public benchmark record backs the positioning. A common reference test in the community: a screenshot of a dense administrative settings panel, roughly forty controls in three tabs, several of them grayed out or sitting in indeterminate states, a few visually inconsistent with their neighbors. The public discussion of how each frontier model handles that kind of test is consistent. Gemini reads every visible control accurately, names the state of each toggle, and flags the visual inconsistencies a design review would care about. Claude produces a competent description but misses some of those inconsistencies. GPT-5 sometimes hallucinates controls that are not present at all, which is the classic vision-model failure.

The same gap shows up on hand-drawn whiteboard parsing, on photo OCR, and on Arabic-script document images. Google built the model around vision and it shows. If image work matters to your stack at all, Gemini 3 Pro is the right pass for that part of the pipeline, regardless of what runs everywhere else. For the full image-side comparison across four models, see the multimodal ranking.

Workspace integration, finally

Google has spent two years promising Gemini integration into Workspace and shipping versions that ranged from useless to actively counterproductive. The version that ships with Gemini 3 Pro is the first one worth keeping turned on. Pulling structured data out of a Sheet into a written summary in a Doc works, and so does drafting a reply with full thread context. The search layer over Workspace documents is more useful than Google's search has been in years.

All of this only matters if Workspace is where your work lives. Write in Markdown and code in a serious editor and the integration becomes a nice-to-have that rarely fires for you. For an organization that runs most operational work through Docs and Sheets, it changes daily work in real, measurable ways. The pricing case for the consumer plan holds either way: $20 a month is roughly the cost of two lunches, and the integration earns it on a single workday where you save a structured-extraction round trip.

The refusal pattern

Gemini 3 Pro refuses prompts that the other frontier models answer without comment. The refusals aren't aligned to the obvious safety categories. The community discussion across Google's developer forum, the Gemini subreddit, and the broader research community is consistent: the refusals cluster around persona-taking, speculative business predictions, and tasks the model classifies as potentially unfair to a category of people.

Ask the model to roleplay as a tough editor giving feedback on a piece of copy and it often refuses, citing reluctance to take on personas that might come across as critical. Ask for a realistic three-year success probability for a startup concept and you tend to get a refusal about making speculative business predictions. A request for a sarcastic monologue from a fictional grumpy mechanic in a video script gets turned down too, this time over concern about negative stereotypes of working-class characters.

None of those refusals is wrong in the abstract, and each one has a reasonable justification behind it. The problem is that Claude and GPT-5 both engage with the same prompts, so the friction of working around Gemini's refusals piles up across a working session into a usability cost you'll feel. For workloads that depend on persona-taking or speculative reasoning, plan around the pattern.

Any single refusal is easy to shrug off. They wear on you the way a small piece of grit in a shoe does, fine until you have walked a mile.

1. Mar 2023 Bard launches

   Google's first public LLM chat product. Not great.

2. Dec 2023 Gemini 1

   First model branded as Gemini. Ultra, Pro, Nano tiers.

3. Feb 2024 Gemini 1.5 Pro

   First million-token context window in production.

4. Dec 2024 Gemini 2

   Better multimodal, faster inference, lower price.

5. Nov 2025 Gemini 3 Pro

   1M context, vision lead, Workspace integration that finally works.

What it costs

Gemini 3 Pro through the AI Studio API costs $2 per million input tokens and $12 per million output, per Google's Gemini API documentation. That input price sits below Anthropic's Opus 4.7 ($5) and just above OpenAI's GPT-5 ($1.25) (verified against Google Cloud's Vertex AI pricing for enterprise use). For a vision-heavy workload at scale, the price advantage is meaningful: thousands of images a day add up fast on any model.

Frontier-tier API pricing, May 2026, per provider docs

Model	Input ($/M tokens)	Output ($/M tokens)	Best at
Gemini 3 Pro	$2	$12	Vision, Workspace
Claude Opus 4.7	$5	$25	Code, long context, honest hedging
GPT-5	$1.25	$10	Visual design, conversational warmth

The Gemini Advanced consumer plan at $20 a month is a clean call if you live in Workspace. If you only open Workspace a couple of times a week for shared documents, treat the integration as a bonus rather than the reason to subscribe. Technical users will get more out of the API tier, and its pricing math is the easier of the two decisions. For the cost picture across all the frontier and mid-tier models, see price per use case.

The role you should put it in

Gemini 3 Pro is the right tool for one specific job: anything that pairs an image with a question. The gap to the alternatives on screenshot understanding, hand-drawn diagrams, photo OCR, and Arabic-language document images is large and consistent across the public record. For that work, this is the only correct pick in early 2026.

For general-purpose work (writing, coding, long-form reasoning) Gemini is competent without pulling ahead of the alternatives, and the refusals levy a friction tax on top of that. The session-to-session variance that turns up in the community discussion is the kind of defect Google will presumably fix in subsequent releases. If you can only run one model, Opus 4.7 stays the better default.

If you can run more than one, put Gemini 3 Pro in your stack as the vision pass: route screenshots, scanned PDFs, and Arabic-script documents to it, and leave everything else to Claude or GPT-5. That's the design Google's product surface seems to assume, and for a vision-heavy workload it's also the cheapest split you can set up this quarter.