A field guide to 2026's LLM benchmarks (and the ones you should ignore)

New benchmarks are a cottage industry. Twelve dropped this quarter alone. Only about half tell you anything useful.

The keepers

• SWE-bench Verified 2026 — still the gold standard for code tasks, the "verified" subset eliminates the ambiguous cases that inflated earlier scores.
• TAU-bench — tool-use evaluation that actually grades multi-step planning. Hard to game.
• GPQA Diamond — graduate-level reasoning, contamination-resistant because the questions require synthesis.
• LiveCodeBench — rolling window of recent competitive programming problems; harder to memorize.

The skippables

• Anything with "AGI" in the name
• Any benchmark whose test set has been on arXiv for more than six months
• Single-shot QA benchmarks on public Wikipedia dumps
• Anything from a vendor evaluating their own model

How to actually choose

Pick two benchmarks that match your use case, not your vibe. A customer-support bot should be evaluated on conversation-length tasks, not on IMO problems. Then build one internal eval that's closer to your production distribution than any public benchmark.

The model that wins on your eval is the one that matters.