Claude vs Gemini, Tested: Writing, Analysis and Docs

Which should you choose for writing, analysis and documents?

Choose the plan that provides the model access you need, then compare its output against your source documents. The official documentation establishes pricing and access, not which product writes the best report for your practice.

Claude Pro includes Anthropic’s recommended starting model, Opus 5.5. It excludes Fable 5.1, which Anthropic describes as intended for demanding reasoning and long-horizon agentic work. Max 5x lists Fable access at 50% of weekly limits, from $100 per month.

Google AI Pro lists flagship access and bundled storage. Confirm that Gemini 4 Argon is available in your app before subscribing specifically for it, because the verified documentation describes an ongoing rollout.

For writing, compare factual fidelity, tone and required edits. For analysis, check whether conclusions follow from the supplied evidence. For documents, ask for traceable source references and flag omissions. Use the same sanitized material and instructions in each product; these are suggested evaluation steps, not findings from a Gemini test.

Our guides to Claude for professionals and Gemini for professionals cover each product separately.

What models and plans are available?

The current lineups include flagship and faster models, but API availability and subscription access are different things. An API price does not establish what a consumer subscription includes.

Model data checked 2026-10-19 on the vendors’ official pages. API prices are USD per million tokens, input/output.

Claude Mythos 5.1, released 2026-09-01, is restricted to vetted cyber-defense and life-science organizations, not a consumer option. Gemini 3.8 Flash replaced the older Gemini 3.7 Flash, deprecated 2026-10-08. Gemini 4 Argon is not on the API free tier.

Prices and plans checked 2026-10-19 on official pricing pages. Figures are US USD list prices, excluding tax.

Readers in the UK, Australia, India and elsewhere should check locally: prices, taxes and rules differ, and local-currency prices are not verified here.

Google’s subscriptions page markets AI Pro and Ultra as providing “access to Gemini 3.1 Pro.” Its model documentation lists that model as a preview in the API and Gemini 4 Argon as rolling out to AI Pro and Ultra in the app. Confirm the answering model in the app rather than assuming API and subscription availability match. No Gemini team price was recorded for this page.

What does our October 2026 round show?

Our October 2026 round on 2026-10-19 included Claude Opus 5.5, the older Claude Sonnet 5 and GPT-6 Astra, but excluded Claude Fable 5.1, Gemini 4 Argon and the current Claude Sonnet 5.5. Claude Fable 5.1 and Gemini 4 Argon were added on 2026-10-20 on all 80 profession prompts across both rounds, with the same prompts and rubrics; Claude Sonnet 5.5 remains untested.

The round ran on 2026-10-19: 40 prompts, including 30 English and 10 Spanish prompts, across 10 professions. Each prompt had 5 pre-written rubric criteria and 3 runs per model, totaling 360 runs.

For the English prompts in our October 2026 round, Claude Opus 5.5 passed 94% of rubric criteria on average, with 64 of 90 runs meeting every criterion. Claude Sonnet 5 passed 91%, with 57 of 90 runs meeting every criterion. Claude Fable 5.1 and Gemini 4 Argon passed 90% and 92% of criteria on the same English prompts, with every criterion met in 54 of 90 and 60 of 90 runs, respectively; both shared the Claude models’ judge, gpt-6-astra. Claude ran through a command-line API harness with tools, web and memory off, while Gemini ran through Google’s Antigravity CLI with a tool-less agent; none were consumer-app sessions, and app results may differ. The judges were not equally strict, so the figures do not establish overall superiority.

Claude outputs were graded by gpt-6-astra; GPT outputs were graded by claude-opus-5-5. In the October 2026 calibration, re-grading 30 sampled Claude runs changed the criteria pass rate from 93% under the published judge to 98% under the swapped judge. For 30 sampled GPT runs, both judges passed 97%. These calibration figures concern the same API-harness setup, not app performance, and demonstrate judge asymmetry.

An editor agent, not a human, independently re-read 20 verdicts and agreed with 18. Part of the Claude-versus-GPT gap therefore comes from the grader. Opus and Sonnet shared a judge, making that comparison like-for-like, but neither can be called better overall from these numbers.

In our October 2026 round, most lost criteria involved added content: warnings, dates, benefits or legal assumptions absent from the input. Word limits also cost the Claude models points. Check summaries for additions and count words when a limit matters.

“Tested” here means at least 2 runs completed and graded, not professional sign-off. Our ChatGPT vs Gemini comparison also separates documented plan fit from benchmark evidence.

How can you compare them on your own documents?

Try the same sanitized source material and prompt in each product, then review both outputs against the original. The templates below are suggested workflows, not verified Gemini results.

Record the plan and model shown, keep the input unchanged, and note missing facts, invented details and corrections required. Do not treat fluent prose as evidence that a calculation or contractual interpretation is sound. For additional templates, see our prompt library.

Engineers can start with a site observation draft. Use fictional or sanitized details in every placeholder. The site observation report skill offers a dedicated workflow.

Help an engineer draft a site observation report using only these sanitized notes. Do not add observations, dates, code references, measurements or conclusions. Put ambiguous information under "Needs clarification."

Project: [SANITIZED PROJECT LABEL]
Visit date: [DATE]
Observer role: [ROLE]
Field notes: [SANITIZED FIELD NOTES]

Output in US English:
- Brief summary.
- Numbered observations, each with its location and supporting source note.
- Needs clarification.
- Items for the engineer of record to confirm before issue.
Include recommendations only if they appear in the notes.

Interior designers can organize a brief before confirming scope with a client.

Organize this sanitized interior-design brief using only the supplied text. Do not invent rooms, budgets, dates, materials or preferences. Mark uncertainty as "To confirm with client."

Brief: [SANITIZED BRIEF, WITHOUT PERSONAL IDENTIFIERS]

Output in US English:
- Spaces in scope.
- Stated preferences and constraints, grouped by space.
- Stated budget and timeline, quoted as written if present.
- Items to confirm with the client.
- Questions for the next meeting, ordered by their effect on the design.

Event planners can draft a run-of-show while keeping unresolved gaps visible.

Draft a run-of-show using only this sanitized schedule. Do not invent times, speakers, vendors, rooms or cues. Flag overlaps and gaps without assigning new times or owners.

Event: [SANITIZED EVENT LABEL]
Date: [DATE]
Venue: [SANITIZED VENUE LABEL]
Schedule and vendor roles: [SANITIZED SOURCE MATERIAL]

Output in US English:
- Table with time, item, owner, location and notes.
- Conflicts and gaps.
- Vendor contact checklist using supplied role labels only.
- Open questions for the venue.

What must you check before using the output?

Check data permissions before entering information, then verify technical and contractual content before issuing the document. Neither a paid plan nor a benchmark result provides professional sign-off.

Limits:

• Data terms: We have not verified either vendor’s data retention, training or business-protection terms for this page. Read the plan’s own data/privacy terms and your professional confidentiality rules before pasting client, patient or employee data.
• US health information: Under HIPAA Privacy Rule 45 CFR 164.502(e), a covered entity may disclose protected health information to a business associate only with satisfactory assurance, through a business associate agreement, that it will safeguard the information. An AI vendor receiving PHI is a business associate. Confirm any agreement with the vendor; none is verified here.
• Technical review: Have the appropriately licensed professional in your jurisdiction check code references, calculations, dimensions, fire ratings, accessibility requirements and specifications before relying on them.
• Contract review: Compare vendor quotes, cancellation clauses, deposits and dates against source documents. Seek qualified review where legal or tax interpretation is needed.

This article is not legal, medical or tax advice. Follow the sign-off requirements applicable to your work. The guides for engineers, interior designers and event planners provide profession-specific checks.

FAQ