Published: 2026-09-16
Summary

DeepSeek V4.1 Flash vs GPT-6 Astra on real builds: 4–6× cheaper, 3–5× slower

Chapters / key moments (click to jump — plays here on the page)

Cost per token doesn't tell you what a build costs. What matters is total spend and wall-clock time to a working result, and this test measures exactly that with the paid OpenRouter API, not a free tier. The trade-off is clear. DeepSeek V4.1 Flash cost roughly 4–6× less and took roughly 3–5× longer, because it needed more check-and-fix passes to reach the same result.

Source video

"I Tested DeepSeek V4.1 Flash vs GPT-6 Astra (RAW RESULTS)" by Bart Slodyczka — Watch on YouTube →

Results as measured

Build	GPT-6 Astra	DeepSeek V4.1 Flash	Notes
Age of Empires replica (2–3 iterations)	$12.25 · 47 min	$3.30 · 4h14m	Astra's first version had fog-of-war and better UI. DeepSeek's farm and upgrades only worked after re-prompting.
Scroll-scrub animated website	$48.63 · 2h10m	$7.74 · about 7.5h	Both used Higgsfield for images and video. Media generation dominated the wait. DeepSeek dropped frames in blocks and got the hover images wrong until its fourth pass.
Barber booking app (SQLite, staff area)	$17.88 · 41 min	(not captured in the excerpt)	Both blocked double-bookings and supported cancellation. Neither was mobile-ready without a follow-up prompt.

Figures as stated in the video. The fourth test (a Blender truck model) came after the part of the transcript we reviewed.

How to reproduce the method

Use the paid OpenRouter endpoint for both models, so cost and token logs come from the same billing source.
Give identical prompts, then identical feedback rounds. Score quality, functionality, cost, build time, and how well each model takes feedback.
Test mobile view separately. Neither model handled it unprompted.
If a build uses an image or video MCP, confirm it is connected in both sessions. His first Astra site had no video because the Higgsfield MCP wasn't set up for that session.

Key Takeaways

DeepSeek's lower price is partly paid back in time: it produced good results but kept iterating.
For client or agency work, his view is to price Astra's API cost into the job and ship faster.
For personal projects where waiting is fine, DeepSeek V4.1 Flash is a legitimate choice.

Model your own costs in the cost calculator, which includes DeepSeek and GPT-6 Astra rates. Contains a sponsor segment (Hostinger).