Gemini 3: what developers actually say

We read every Hacker News comment from Nov 2025 – Nov 2026 that gave a real opinion on Gemini 3. 58 of the 70 opinions come from people who say they use it themselves.

56%approval, all opinions

–approval, last 90 days

#37of 45 models ranked

67different developers

Developers are split on Gemini 3 for software work. Approval is 56% across 70 opinions from 67 authors, and 58 of those opinions come from firsthand use. Code quality mentions lean positive at 15 to 9, while reliability is negative at 2 to 9. Volume fell from 22 opinions in November 2025 to 1 or 2 opinions in each month with data since March 2026, and the recent 0%, 0%, 100% and 25% readings rest on one or two opinions each.

Pick it if

Want fast UI and app scaffolding from short prompts
Need to pull data from PDFs, drawings or scans
Have short, clear tasks and will review the output

Look elsewhere if

Need dependable long-running agent sessions
Rely on stable GitHub Copilot integration
Need strong multi-step reasoning or loop-free tool calls

Strengths and weaknesses

Each row counts opinions that mentioned the topic. Bars to the right of the line are praise, bars to the left are complaints.

PraiseComplaints

Code quality

Reliability

Speed

Autonomy

How opinion has moved

Monthly approval (line) and number of opinions (bars). Months with fewer than 3 opinions are left out.

What developers say

Praise

Says Gemini 3 greatly helped build a Redis-compatible cache, including reasoning through Raft edge cases and generating tests.HN comment · Feb 27, 2026

Says recent models including Gemini 3 now handle complex compiler and hardware tasks.HN comment · Feb 14, 2026

Vibe coded a retro emulator and assembler with Gemini 3 using minimal prompts, with great results.HN comment · Feb 9, 2026

Switched to Gemini 3 and found the agent much smoother than with Kimi 2.5.HN comment · Feb 2, 2026

Chose Gemini 3 over GPT-5 and Claude 4.5 for agent orchestration, saying it handles React code better.HN comment · Jan 21, 2026

Says Gemini 3 feels better than its predecessors for vibecoding small, non-critical apps.HN comment · Jan 14, 2026

Says Gemini 3 Pro and Flash make a huge difference extracting data from PDFs like footprints and drawings.HN comment · Jan 12, 2026

Says Gemini 3 changed their workflow by quickly producing working UI and app code.HN comment · Jan 9, 2026

Complaints

Says Gemini 3.x hung, looped, and hallucinated on longer tasks, making it unusable for serious work.HN comment · Jul 17, 2026

Says Gemini 3 kept sending them from one post-mortem report to another.HN comment · Mar 17, 2026

Reports Gemini 3 was unreliable in GitHub Copilot with frequent server errors, and unusable for two months.HN comment · Feb 19, 2026

Finds Gemini 3 useless and says it hallucinates more than Gemini 2.5.HN comment · Feb 6, 2026

Says Gemini 3 also failed a simple Ruby class refactor, alongside Codex.HN comment · Jan 19, 2026

Says Gemini 3 is consistently the worst, getting stuck in loops and failing tool calls.HN comment · Jan 8, 2026

Calls Gemini 3 benchmark-driven and a flop; says it codes worse than OpenAI and Anthropic models, though its image output is good.HN comment · Jan 8, 2026

Finds Gemini 3 far behind Opus and GPT-5.2 in wit and multi-step problem solving.HN comment · Jan 7, 2026

Summaries are paraphrased. Each links to the original comment.

Gemini 3 compared

Gemini 3 vs Claude Opus 4.5

FAQ

Is Gemini 3 worth it?

The data is mixed. Approval is 56% across 70 opinions, and code quality mentions run 15 positive to 9 negative. Reliability mentions run 2 positive to 9 negative, so a trial on your own tasks is the sensible step.

Is Gemini 3 better than Claude Opus 4.5?

Most preference data favors Claude Opus 4.5. Five people preferred Opus 4.5 instead, and one preferred Gemini 3 over it. Complaints also place Gemini 3 behind Opus and GPT-5.2 on multi-step problem solving.

Is Gemini 3 reliable for coding?

Reliability is the weakest aspect in the data, with 2 positive and 9 negative mentions. Users report hangs, loops, failed tool calls and frequent Copilot server errors. Short, clear tasks get better reports than long agent runs.