Gemini for Professionals (2026): Tested, Plans, Limits

What is Gemini, and which models does it offer?

Gemini is Google’s family of AI models, offered through the Gemini app and Google’s API. The verified current lineup is:

• Gemini 4 Argon: flagship, released 2026-10-13 and generally available in the API that day; rolling out in the app to AI Pro and Ultra.
• Gemini 3.8 Flash: fast model.
• Gemini 3.6 Flash: fast model.
• Gemini 3.5 Flash-Lite: lite model.
• Gemini 3.1 Pro: preview in the API; the subscriptions page markets AI Pro and Ultra as providing access to it.

Gemini 4 Argon is the only Gemini model in our test round so far. The other current models listed above are not in this test round yet.

For API users, Gemini 4 Argon costs $2/$10 per MTok through 2026-11-30 and $4/$20 from 2026-12-01. It is not on the API free tier; do not confuse API pricing with app subscriptions.

Which Gemini plan fits your work?

Google AI Pro is the lowest-priced app plan listing Gemini 4 Argon. Try Free or Plus for occasional, low-stakes work; consider Ultra if AI Pro’s limits are a recurring obstacle.

Prices checked 2026-10-19 on the official pricing page.

These are US list prices excluding tax. For readers in the UK, Australia and India, local prices, currency and tax differ; check Google’s pricing page for your country.

Choose by the work you can verify, not your job title alone. An accountant drafting an email and a lawyer reviewing a legal argument need different checks. See our guides for accountants and lawyers for profession-specific context.

How did Gemini do in our tests?

Gemini 4 Argon’s results support trying it with review, not treating it as a substitute for professional judgment. We added it on 2026-10-20 using the same locked prompts, inputs and rubrics as the other models; our testing methodology explains the setup.

The harness matters. This was not the Gemini app or a bare API request. We used Google’s Antigravity CLI with a custom agent loading no tools, skills or rules, a one-line system prompt, and a fresh empty folder in sandbox mode. The client still exposed one built-in tool for listing or stopping background tasks. Any tool call would have voided the run; none occurred: 0 tool calls and 1 turn per run. We used gemini-4-argon-medium, the medium reasoning setting. App results may differ.

Read the scores with the judge caveat. All 240 Gemini runs were completed and graded by gpt-6-astra. The judges are not equally strict: on the same 10 sampled Gemini runs per round, gpt-6-astra passed 92% of criteria versus claude-opus-5-5 at 98% in round 1, and 86% versus 88% in round 2. Part of any Gemini-vs-GPT gap therefore comes from the grader, not the model.

• Round 1: 30 English and 10 Spanish prompts. Gemini passed 91% of rubric criteria; 75 of 120 runs passed every criterion.
• Round 2: 20 French and 20 Brazilian Portuguese prompts. Gemini passed 80% of criteria; 36 of 120 runs passed every criterion.

Round 2 included deliberate traps, including a wrong total and unit conversion. Scores were lower for every model; do not average the rounds.

Claude comparisons use the same judge. With gpt-6-astra grading both families, round 1 results were Claude Opus 5.5 92%, Claude Sonnet 5 91% and Claude Fable 5.1 90%. Round 2 results were Claude Opus 5.5 79%, Claude Fable 5.1 78% and Claude Sonnet 5 70%. These are like-for-like grading comparisons, not evidence that Gemini is better overall.

Spot check: an editor agent, not a human, re-read 10 Gemini verdicts and agreed with 7. In all 3 disagreements, the judge was stricter.

Excel / Microsoft 365: in the separate 2026-10-20 round of 10 task prompts, Gemini passed 91% of criteria; 19 of 30 runs passed every criterion. Judge asymmetry applies here too: on 6 sampled runs, gpt-6-astra passed 93% versus claude-opus-5-5 at 97%. These were text-only answers: formulas were judged by their logic and stated result, not executed in Excel, not Copilot in Excel or Microsoft 365. Chat models answered pasted tasks. Do not pool this round with the others.

Spanish round 2: on 2026-10-21, Gemini passed 74% of criteria, with 4 of 30 runs passing every criterion. This separate set of 10 new Spanish prompts is not pooled with rounds 1 and 2 or the Excel round. Gemini shares its judge (gpt-6-astra) with the Claude models, making those grading comparisons like-for-like, unlike comparisons with GPT: on 6 sampled Gemini runs, gpt-6-astra passed 73% versus claude-opus-5-5 at 90%. In the spot check, 12 verdicts were re-read, with agreement on 8.

Most lost points came from safety criteria involving added warnings, dates, benefits or legal assumptions absent from the input. Claude Skills were not run on Gemini. Our test-round report provides further context.

What should you check before pasting confidential data?

We have not verified Google’s data terms for this page. We cannot confirm retention policies, training practices or business data protections for any Gemini plan.

Before entering client, patient or employee data, read your specific plan’s data and privacy terms and your professional confidentiality rules. A paid subscription alone is not evidence that a particular use is permitted.

When should you avoid relying on Gemini?

Do not rely on unchecked output where an error could have legal, financial or clinical consequences. A rubric score does not establish that your particular answer is correct.

Recalculate totals, check unit conversions and remove unsupported assumptions. If your workflow depends on Gemini 4 Argon in the app, account for its ongoing rollout; if it depends on Gemini 3.1 Pro through the API, note its preview status.

Tax, legal, health and property work must be checked by a licensed professional in your jurisdiction.

FAQ