The 18 LLM leaderboards we'd actually use.

Leads both SWE-bench Verified (87.6) and SWE-bench Pro (64.3) and tops Chatbot Arena; the surest hand on a real PR.

Best generally-available τ-bench (87.5); stays on-task across long tool loops and self-corrects without prompting.

Best current BFCL (72.5) with rock-solid JSON-schema adherence and a 1M window at $5 out.

Best open-weights model we've tested: #1 LiveCodeBench (93.5), 80.6 SWE-bench, 1M context, $0.87 out.

The long-context standard — 1M tokens with cheap input that keeps the bill survivable.

$0.22 / 1M out with LiveCodeBench 91.6 and a 1M window — nothing else is this capable this cheap.

Tops Chatbot Arena (1503) and writes paragraphs you'd ship; understands tone notes and edits like a copy chief.

GPQA Diamond 94.2 (top GA) with the cleanest footnoted synthesis across many sources.

1M context plus MMLU-Pro 88.6 at $3 out — handles a 500-page transcript faithfully without breaking the budget.

Highest groundedness and graceful refusal — says "not in the docs" instead of guessing.

Best coverage of lower-resource languages with strong idiom handling and a huge context for document-level consistency.

Best text-to-SQL accuracy in production — HumanEval 94.2 and top-tier reasoning pick the right join and respect dialect quirks.

The current photoreal leader — brand-consistent, with the best text rendering and hands in the open ecosystem.

Best shot-to-shot continuity with native synchronized audio.

Most expressive voices with the lowest artifact rate — the studio standard for production VO.

Lowest WER on noisy real-world audio with the broadest language coverage; cheap to self-host.

Most production-ready output; stems are usable and the vocals are believable.

Best inpaint preservation — brand colors and untouched regions stay put across edits.