DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5

I previously compared DeepSeek v3.2 and Claude Opus 4.5 across two tests: writing a PRD and then building a space dashboard app from that PRD. Opus came out on top, and the version it built looked professional, nailed the brief, and only had one or two minor bugs. A viewer suggested testing DeepSeek again using the full Opus PRD instead of its own shorter one.

That made sense, so I gave DeepSeek another chance using the full Opus PRD. I also added a third model to the mix, GPT-5.1 Codex Max. Both DeepSeek and Codex received the exact same PRD, and I compared their builds against the Opus reference.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - context

I wanted to see if a richer PRD would help DeepSeek build a better space dashboard. I also wanted to see how GPT-5.1 Codex Max handled the same brief. For a focused writing-tool comparison between Opus and GPT-5.1 in another domain, see this direct head-to-head on writing tasks.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - PRD we used

It specified feature requirements, architecture design, UI and UX requirements, and technical specifications. It listed all required API endpoints and a recommended tech stack. Both models used the same stack: React 18 with TypeScript.

It also included a development roadmap that broke the project into tasks. It covered risks and mitigations, future considerations, and an appendix. In short, a complete PRD built for production.

For context on how a newer Codex model compares against Opus in a broader sense, check this Codex vs Opus analysis.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - setup

I pasted the full Opus PRD into both workspaces. I let each model handle the build process. I then ran both dashboards in dev mode for testing.

Build times and stability

DeepSeek took about 40 minutes and got stuck a few times. I had to cancel and prompt it to continue with the task.

I am not sure if that was an API connection issue or a client issue. It did not happen in the first DeepSeek test. Either way, it took longer to complete the build this time.

If you care about throughput and operating costs on newer variants, see these notes on speed, cost, and quality for GPT-5.2 vs Opus 4.5.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - DeepSeek v3.2 results

Some font and formatting could be improved. The font was too large in places and made text hard to read. That caused clear formatting issues that look like a CSS problem to fix.

The ISS tracker was still loading at the time of testing. The astronomy picture of the day widget broke the image. Upcoming events, data sources, and about sections were present.

Comparing this to the original DeepSeek v3.2 dashboard built from the shorter DeepSeek PRD, there were no formatting issues in the earlier one. The layout was very similar across both versions, with the map and picture of the day at the top and three widgets below. The main difference was the new design and the formatting problems here.

It might be that the detailed PRD contributed to the formatting problem. One possible workaround is to break the PRD into tasks. You could ask Opus to break it into modules, then ask DeepSeek to build module by module.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - GPT-5.1 Codex Max results

The astronomy picture of the day widget appeared to have broken images. That could be an API or CORS issue to fix. The widget itself looked clean from a UI perspective.

Further down, there was a space events calendar, a near-Earth object monitor, and space weather. There was a clear formatting issue in the middle section where content was squeezed and text wrapped to one word per line. There was a lot of empty space caused by poor layout in that row.

It did not nail how those three sections should work together. The ISS tracker section was solid. The mission control section at the top was also good.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - Opus 4.5 reference

Opus 4.5 again comes out on top when compared with the DeepSeek v3.2 version using the detailed PRD and the GPT-5.1 Codex Max build. That held true across both design quality and functional reliability.

For estimates on usage at scale, you can review this breakdown of token costs for GPT-5.2 vs Opus 4.5.

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - comparison overview

For more three-way perspectives that include other families, this broader comparison with GLM is helpful: GLM 4.7 vs Opus 4.5 vs GPT-5.2.

Use cases

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - DeepSeek v3.2

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - GPT-5.1 Codex Max

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - Opus 4.5

If you are comparing newer Codex variants to Opus on throughput and quality, here is another practical read on speed, cost, and quality trade-offs.

Pros and cons

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - DeepSeek v3.2

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - GPT-5.1 Codex Max

DeepSeek V3.2 vs GPT-5.1 Codex MAX vs Opus 4.5 - Opus 4.5

If you are weighing ongoing usage, see this breakdown on token costs for GPT-5.2 vs Opus 4.5 to plan budgets.

Step-by-step - reproduce the test

Set reasoning to extra high in Codex.
Paste the full Opus PRD into both sides.
Prompt each model to build the space dashboard app.

Wait for builds to complete.
Switch to dev mode for each project.
Test the ISS tracker, APOD, events, NEO, and space weather widgets.

Note build times and any required prompts.
Record formatting issues, broken images, and layout problems.
Compare each build to the Opus 4.5 reference for quality.

For more context on Codex vs Opus across versions, here is a focused look at Codex vs Opus.

Quick CSS fix - font and layout

:root {
  --font-base: 14px;
  --font-scale: 1.1;
}

body {
  font-size: var(--font-base);
  line-height: 1.5;
  -webkit-font-smoothing: antialiased;
  -moz-osx-font-smoothing: grayscale;
}

.widget h2, .widget h3 {
  font-size: calc(var(--font-base) * var(--font-scale));
  margin: 0 0 8px;
  line-height: 1.2;
  word-break: keep-all;
  overflow-wrap: anywhere;
}

.dashboard-row {
  display: grid;
  grid-template-columns: repeat(3, minmax(0, 1fr));
  gap: 16px;
  align-items: start;
}

.widget {
  min-width: 0;
  overflow: hidden;
  padding: 12px;
  background: rgba(255, 255, 255, 0.06);
  border: 1px solid rgba(255, 255, 255, 0.1);
  border-radius: 10px;
}

.widget p {
  margin: 0 0 6px;
  font-size: 0.95rem;
}

/* Fix for one-word-per-line issues in tight columns */
.widget .content {
  white-space: normal;
  word-break: break-word;
}

Final thoughts

This test also suggests a practical tip for DeepSeek v3.2: break a large PRD into modules and build step by step. That can reduce stalls and formatting regressions and improve outcomes with the same input. For a broader lens that includes GLM, this three-way comparison is a good companion read: GLM 4.7 vs Opus 4.5 vs GPT-5.2.