Gemini 3.1 Pro versus Opus 4.6: inspect two real coding tasks

Prepare the two jobs before comparing models

The 20 February 2026 recording compares Gemini 3.1 Pro with Claude Code using Opus 4.6. Its tasks are a one-page flight simulator and a SaaS feature for Luke’s YouTuber Systems platform. Those are different tests: one asks for a self-contained interactive result, while the other must fit an existing product and its behaviour.

1. Follow setup at 0:35 and the feature brief at 2:02, recording the requirements before looking at the outcome.
2. Keep each model’s starting code, instructions and changes identifiable so one attempt does not inherit another’s fixes.
3. Define separate acceptance checks for the simulator and the integrated feature rather than choosing a winner from the first attractive preview.

Test the visible result and investigate blocked routes

The simulator starts at 3:35, with results at 4:29. SaaS work begins at 6:55 and a Gemini-struggles segment at 7:39. The description reports that one run could not get past a 404, but it does not provide a complete reproduction or code listing. Treat a blocked route as a failure to diagnose, not as evidence of a usable completed feature.

1. Open the simulator and try its intended interaction rather than inspecting the code alone.
2. For the SaaS feature, follow the existing product journey and note the first point where the implementation diverges from the brief.
3. Record the failing URL or action and visible response, then compare any proposed repair with the same case.
4. Use the side-by-side comparison at 10:44 and packaging feature demo at 12:26 to inspect what actually became usable.

Keep the historical spending argument tied to this workload

API costs are discussed at 3:50 and the verdict at 14:46. Luke argues that subscriptions such as Claude Max make more sense for his daily use. That opinion concerns his workload and the prices available then; it is not a current rate comparison or a guarantee that a subscription suits every usage pattern.

1. Record usage, follow-up attempts and unresolved failures alongside the accepted output.
2. Compare the cost of finishing each task, not only a quoted per-request price.
3. Check present access and billing terms separately before choosing a model route for your own daily work.

The takeaway

Evaluate both the standalone build and the integrated feature, preserving failures and total effort in the comparison.