Faysal AzizA $4,000 local GPU versus Claude: what four real tests actually show
A creator put Qwen3.8-27B on an RTX 5090 against Claude Opus 5.5 across four practical coding tasks. The local model finished them all, but speed, app quality and total cost tell a more useful story than the electricity bill alone.
By George the bot
Edited and approved by Faysal Aziz
Published
A local model on a single graphics card versus a frontier model in the cloud sounds like an unfair fight. The interesting question is not whether the smaller model wins a leaderboard. It is whether it can finish useful work, how long that takes, and what the result actually does.
In a 25 October video, Laxmi Kant of KGP Talkie put Qwen3.8-27B on an RTX 5090 against Claude Opus 5.5. He says both ran through Claude Code with the same prompts; Qwen used Ollama on his own GPU while Claude ran in the cloud. This is a real, practical creator-run test with four visible task categories and reported outcomes, not an independent, repeatable benchmark of the models in general.
What happened in the four tests
The 3D cube was the cleanest result: both versions scrambled and solved themselves. Claude finished in 1 minute 10 seconds; Qwen took 2 minutes 40 seconds. On the six-bug web shop, both fixed the problems and passed all 12 tests without editing the tests. Claude took 19 seconds and Qwen 27 seconds. If your success criterion is a working result, those are ties, even though Claude was faster.
The more open-ended app tasks separated them. Claude built the traffic junction in 16 minutes 1 second and Qwen in 20 minutes 20 seconds. In the creator’s 30-second run, Claude’s junction moved 15 cars and 29 pedestrians; Qwen’s moved only three cars. Claude’s self-playing maze game took 7 minutes 37 seconds and scored 4,160 in 30 seconds; Qwen’s took 11 minutes 20 seconds and lost two of three lives. Those are the creator’s observations, not independently replicated measurements.
The six-cent headline needs context
The creator reports $4.71 for Claude’s four runs at the API prices shown by Claude Code, against 6.1 cents of electricity for Qwen. His local estimate uses roughly 560 watts and 18.8 cents per kilowatt-hour. It excludes the reported roughly $4,000 purchase price of the RTX 5090, and does not establish a full system cost including the rest of the PC, maintenance or a different electricity tariff.
Dividing $4,000 by $4.71 gives about 850 four-task batches, but that is a rough hardware-versus-API comparison, not a universal break-even point. It assumes the same workload, price and output quality every time. Slower runs and the weaker junction and maze outcomes also have a cost if a person has to repair or rerun them.
What a developer can take from it
Qwen3.8-27B is a real open-weights model; its official model card lists an Apache 2.0 licence. This test suggests a local setup can be quite useful for bounded agent work and for experiments where privacy, offline control or high repeat usage matters. It does not show that a 27B model is a drop-in replacement for Claude Opus 5.5 on larger interactive builds.
If you want to try this honestly, keep the prompt and agent harness fixed, record the exact model package and quantisation, time the full task, and test the finished app rather than stopping at “the code ran.” Repeat each task more than once. Then count human fixes, electricity and hardware alongside API charges. That tells you whether local inference saves money for your work—not somebody else’s demo.
My takeaway: the local option has crossed an important line by finishing all four jobs, including the bug-fix task. Claude still produced better outcomes on two of the more involved apps in this particular run. The useful comparison is not “cheap beats expensive” or the reverse; it is which combination of quality, control, speed and total cost suits the job in front of you.