Model picker — which AI brain you're using

Top of the topbar in /interview shows the current model as a small pill. Click it to open the dropdown. Six models, split into free and Pro tiers.

Free tier (3 models)

Model	Strength	Speed
Claude Haiku 4.5	Fastest, smart enough for most behavioral + light-technical	~300 ms first token
GPT-4o Mini	Different "voice" than Claude — useful for variety	~400 ms first token
Llama 3.3 70B	Open-source, no vendor lock-in. Solid generalist	~500 ms first token

All three free models handle: behavioral STAR answers, "tell me about yourself," resume-grounded responses, basic system design verbal walkthroughs.

Pro tier (3 models) — Pro / Pro+ / Lifetime only

Model	Strength	When to use
Claude Sonnet 4.6	Best balance of accuracy + speed. Excellent at synthesis	Default for most paid sessions
Claude Opus 4.6	Highest reasoning quality available. Slower (~800 ms first token)	FAANG / quant / hard-systems-design rounds
GPT-4o	Strong code reasoning, slightly different style from Claude	When you want a non-Claude voice for variety

How to pick

Default rule of thumb:

• Free tier, want fastest → Haiku 4.5
• Free tier, want a non-Claude voice → GPT-4o Mini or Llama 3.3
• Paid, want default → Sonnet 4.6
• Paid, hardest question of the round → Opus 4.6 (worth the ~500 ms extra)

Pro models auto-downgrade for free users

If you're on free tier and the dropdown shows a Pro model (e.g., Sonnet) — that's a stale selection. The system silently falls back to your tier's free default (Haiku 4.5) before sending the request. No error, no broken session, just quieter quality. Upgrade to unlock the actual pick.

Pro+ exclusive: "Always Pro" guarantee

Pro+ subscribers get a flag (features.alwaysProModel) that prevents the system from auto-downgrading the requested model under load. When the rest of the world is hitting Anthropic's rate limits and we fall back to Haiku, Pro+ stays on Opus. Marketed as "your model never auto-downgrades when it matters most" — true; see entitlements policies.

Switching mid-session

Yes — switch any time. The next AI answer uses the new model. Already-streamed answers don't change. There's no penalty.

Cost-aware switching tip

Opus is ~5× the per-token cost of Sonnet. If your session is mostly easy questions, Sonnet handles them fine. Save Opus for the 1-2 hardest questions of the round (system design, complex tradeoff questions, niche-domain technical drills).

Why no Gemini

Dropped 2026-04-22 to shrink the support matrix to 3 providers (Anthropic + OpenAI + Groq/Llama). Fewer providers = fewer 5xx modes, fewer "which one is down?" support tickets. Will re-add when Google reliability is measurable AND the free-tier differentiates enough to matter for conversion.

Still stuck?

Help & Feedback → Contact us with category "How do I…?" if a model selection isn't behaving as expected.

Video walkthrough

Coming soon.