Marauder Bot

Marauder Bot runs three frontier models in production — OpenAI's GPT-5.6 Luna, Anthropic's Claude Opus 4.8, and Google's Gemini 3.5 Flash — against a steady stream of real interview questions: algorithm problems, multiple-choice sections, system design prompts. Users can switch between them per analysis, which means we get to watch how the current frontier actually behaves on this specific workload. Here's what that looks like in 2026, and how to choose.

First, the honest baseline

On standard interview material, all three models are past the point where raw correctness separates them. A typical LeetCode-style medium, a CS-fundamentals multiple-choice question, a "design a URL shortener" prompt — each of the three handles these reliably. The frontier moved; interviews are still catching up to that fact.

The differences that matter in practice are speed, depth on hard problems, and style — and they're real enough that we keep all three models in the picker rather than crowning one.

GPT-5.6 Luna: the default, for a reason

OpenAI's latest is the fastest of the three in our lineup and the one we ship as the default. On the core loop — parse a page, produce working code with tests, or pick the right option with a tight explanation — it delivers frontier-quality answers with the least waiting. When you're reviewing a 20-question practice section and running an analysis per question batch, latency is the difference between a review loop you sustain and one you abandon, so "excellent and fastest" wins the default slot.

Reach for it when: you want the best speed-to-quality ratio, which is most of the time.

Claude Opus 4.8: the deep read

Anthropic's Opus 4.8 is the most capable model we serve, and the slowest. Where it earns its keep is the material where the extra depth shows: gnarly dynamic-programming problems, algorithm questions with subtle edge-case traps, and especially system design, where its writeups tend to be the most thorough — fuller tradeoff discussions, more careful requirement analysis. Its code answers also skew toward more complete test coverage.

Reach for it when: the question is genuinely hard, the answer will be studied rather than skimmed, or it's a system design prompt.

Gemini 3.5 Flash: the sprinter

Google's 3.5 Flash (which replaced Gemini 3.1 Pro in our lineup) is built for speed, and it shows: it's the model you want on high-volume, lower-depth work. Blasting through a long multiple-choice section, getting a quick sanity-check on a straightforward problem, triaging which questions on a page deserve a deeper look — Flash turns these around with minimal wait while staying solidly accurate on standard material.

Reach for it when: volume and turnaround matter more than maximum depth.

The pattern we designed around

One production detail worth sharing: for interview-style questions, we run every model with reasoning minimized. Extended "thinking" happens before any output streams, so heavy reasoning modes mean long silent waits — the opposite of what a real-time answer needs — and for this class of question, the quality difference rarely justifies it. It's a good reminder that "which model" matters less than how a model is run for a given job.

So which should you pick?

Our honest recommendation, matching how we see people use the picker:

Default to GPT-5.6 Luna and don't overthink it.
Escalate to Opus 4.8 for hard problems and design questions worth a slow, thorough answer.
Drop to Gemini 3.5 Flash for rapid-fire volume.
Disagreement is information. When an answer looks off, re-run the question on a second model (one click in the extension, or use regenerate). Two models agreeing is strong evidence; two models disagreeing tells you the question deserves your own careful read — the verification habit matters more than the model choice.

The lineup will keep evolving — we retire and add checkpoints as the vendors ship — but the shape (a fast default, a deep specialist, a sprinter) has been stable. Try all three against the same question on the algorithm demo, or start the free trial and run the comparison on your own practice material.