Pick a task. Watch the models race. Replays are recorded and free; live runs are real, capped, and behind a bot-check. Not an Elo leaderboard — a priced race: the cheapest correct answer wins, and every cost is computed from the provider's own usage counts (estimates are marked ≈).
The cheapest lane was right. The one costing 74× as much never finished.
reasoningOne sentence, one letter, one number. Nobody's safe.
witThree lines, must contain "cell". Whether it's any good is your call.
extractionA coffee run becomes JSON. "July 5th" and "two flat whites" are the traps.
extractionThree addresses buried in office chatter — one of them spelled out in words.
instruction-followingSummarize a paragraph in exactly 12 words. Count along on your fingers.
instruction-followingEight words, alphabetized by their second letter. Auditable mid-stream.
instruction-followingEvery "a" becomes a 4. One missed letter loses the lane.
reasoningName the third Wednesday of September. Showing your work costs money here.
witOne line, under 12 words, three banned words. The tripwire is live.