Frontier Models on Benchmarks 2026: Sonnet 4.7, GPT-5, Gemini 3 Cited

One table, twelve benchmarks, five models. Every number sourced. No single overall ranking.

Numbers below are the most recent vendor or leaderboard score we could source for each model on each benchmark. Where the vendor reported a higher headline with a substantially weaker methodology, we use the methodology-disclosed number and note the gap in the per-benchmark page linked below.

Three patterns to call out. First, the frontier is tight on most benchmarks: the gap between the leader and second place is within 2 to 4 points on SWE-bench, MMLU-Pro, BFCL, and Tau-Bench. Second, reasoning-tuned models (GPT-5, Sonnet 4.7) lead on AIME and HLE by larger margins (4 to 8 points) because the benchmarks reward extended-thinking budgets. Third, open-weight models trail by a meaningful margin on agentic benchmarks (SWE-bench, Tau-Bench, GAIA) and a smaller margin on knowledge benchmarks (MMLU-Pro). The agentic gap is the most consequential for practitioners choosing a deployment model.

Cost. Sonnet 4.7 input is $3 per million tokens; Llama 4 70B can be self-hosted at roughly $0.40 per million tokens through a hosted endpoint. The cost spread is roughly 8x. A 5-point benchmark gap is rarely worth an 8x cost multiplier in production; a 20-point gap usually is. Read the table alongside the cost calculator linked below.