Gemini 3.7 Flash vs Claude Sonnet 5: which fits your work?

What actually differs?

Google documents text, image, video, audio and PDF inputs for Gemini 3.7 Flash, with text output. It is not an image-generation model. Anthropic presents Sonnet 5 around coding, reasoning and tool-assisted work. These are provider capabilities and positioning, not results from our own head-to-head test.

A chat interface exposes only part of a model’s capabilities. An API feature list does not mean that the same file types, tools or context allowance are available in GO AI. For this comparison, start with ordinary text tasks that the web chat supports.

Google: Gemini 3.7 Flash · Anthropic: Claude Sonnet 5 · Anthropic: Sonnet 5.5, 28 September 2026 · Anthropic: Sonnet 5 specifications

Test 1: a client message with fixed facts

Use a short draft containing a date, a name, an amount and one uncertainty. Ask for a rewrite, then check those four details before judging style. A polished answer that turns an estimate into a promise fails the test.

Try a second tone only after both models have answered the original brief. Otherwise you are comparing different instructions. Keep the original draft next to the answers so you can verify every change.

Rewrite this message for [audience] in a [tone] tone. Keep all names, amounts and dates unchanged. Preserve uncertainty; do not add commitments. Use no more than 120 words. Then list any facts that need clarification. Draft: [text].

Test 2: a summary that can be checked

Give each model the same meeting notes. Count invented decisions, missing actions and incorrectly assigned owners. This distinguishes an attractive summary from a dependable one.

Include an unresolved proposal in the notes. A useful answer keeps it separate from approved decisions. Do not assume either model has read a linked document unless you actually supplied its contents.

Use only these notes: [notes]. Return three sections: decisions, actions with owners and deadlines, and unresolved questions. For each decision, quote the supporting phrase. Write “not stated” for missing owners or dates. Do not turn proposals into decisions.

Test 3: a small code review

Provide a short function, the intended behaviour and one failing example. Compare whether the answer identifies a reproducible problem, proposes a focused fix and explains a meaningful test.

A generated test is not a test result. Run the proposed checks in your own development environment before accepting a fix. Neither model should receive credit for claiming to have executed code in an ordinary chat.

Review this code: [code]. Expected behaviour: [requirements]. Failing example: [input and observed output]. Identify the likely cause, propose the smallest fix and give a test that fails before the fix. State what you have not executed.

How to choose after the comparison

For each task, record factual errors, missed instructions and the minutes you spent correcting the answer. A simple pass/fail checklist is more useful than an invented overall intelligence score.

Repeat an important task before making one model your default. If Gemini handles your summaries better and Claude needs less correction on your code, keep both roles. Our recommendation is this testing method; we have not run a controlled benchmark for this article.

Common questions

Sources and scope

• Google: Gemini 3.7 Flash
• Anthropic: Claude Sonnet 5
• Anthropic: Sonnet 5.5, 28 September 2026
• Anthropic: Sonnet 5 specifications