DeepSeek V4: what developers actually say

We read every Hacker News comment from Nov 2025 – Nov 2026 that gave a real opinion on DeepSeek V4. 39 of the 48 opinions come from people who say they use it themselves.

78%approval, all opinions

72%approval, last 90 days

#13of 45 models ranked

44different developers

Developers give DeepSeek V4 a 78% approval score across 48 opinions from 44 authors, including 39 firsthand reports. Cost is the main driver: pricing drew 21 positive and 2 negative comments, and code quality drew 10 positive and 1 negative. Reliability is the weak spot at 1 positive and 4 negative, with reports of stalls and false completion claims on longer tasks. Approval over the last 90 days is 72%, below the overall figure, and monthly samples are small: 50% in August 2026 and 100% in October 2026.

Pick it if

Want low cost for routine coding and quick fixes
Need a cheap second opinion on engineering decisions
Work on algorithmic reasoning or small, contained tasks

Look elsewhere if

Need dependable long runs in large codebases
Need precise detail recall across large context
Depend on stable output formats and complete rows

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

Cost & pricing

Code quality

Speed

Reliability

Large codebases

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 they will keep using DeepSeek V4 if the new release is comparable to it.HN comment · Oct 19, 2026

Uses DeepSeek V4 locally as daily driver; it one-shots iOS apps but is slowish and sometimes needs retries.HN comment · Oct 17, 2026

Lists DeepSeek V4 among favorite open models used via API before moving to a $100 subscription.HN comment · Oct 9, 2026

Likes the DeepSeek V4 models, saying they critique engineering decisions better than MiMo.HN comment · Aug 12, 2026

Uses DeepSeek V4 as a cheap helper that noticeably improves a local model when it gets stuck.HN comment · Aug 10, 2026

Uses DeepSeek V4 via Copilot for about $10/month; less capable than frontier models but plenty for their tasks.HN comment · Jul 25, 2026

Calls DeepSeek V4 their favorite model, mostly using the cheaper models in OpenCode.HN comment · Jul 20, 2026

Has been using DeepSeek 4 agentically and says it is very good, though probably not as good as Anthropic's models.HN comment · Jun 25, 2026

Complaints

Says the final DeepSeek V4 overthinks, takes over an hour per run, and grades lower than Luna.HN comment · Aug 8, 2026

Says DeepSeek V4 flash and pro changed output formats and skipped rows, unlike GPT models on the same check.HN comment · Aug 6, 2026

Finds DeepSeek V4 weaker than Opus at spotting small details in large context.HN comment · May 31, 2026

Says DeepSeek V4 repeatedly claimed completion on a reverse-engineering task and took shortcuts.HN comment · May 24, 2026

Says DeepSeek V4 failed to fix a dark-text input styling problem.HN comment · May 24, 2026

Says DeepSeek V4 trailed GLM 5 at app-scripting agent tasks in their tests.HN comment · May 7, 2026

Tried DeepSeek V4 in OpenCode; it got the job done, but its visible thinking looked erratic.HN comment · May 2, 2026

Argues DeepSeek V4 lacks V3's impact and peak-hour pricing erodes its cost advantage.HN comment · Jul 19, 2026

Summaries are paraphrased. Each links to the original comment.

FAQ

Is DeepSeek V4 worth it?

For cost-sensitive work, yes. It has 78% approval, and pricing is its strongest aspect at 21 positive and 2 negative comments. Check reliability needs first, since that aspect is 1 positive and 4 negative.

Is DeepSeek V4 better than Claude Opus?

The data does not settle this. One person preferred DeepSeek V4 over Claude Opus 4.5, and one preferred Claude Opus 4.7 instead. A complaint also says DeepSeek V4 is weaker at spotting small details in large context.

How reliable is DeepSeek V4?

Reliability is its weakest aspect, with 1 positive and 4 negative comments. Users report false completion claims, erratic visible thinking, and runs that take over an hour. The overall approval score is still 78%, so results vary.