Sonus AI drops Sonus-1 LLMs with 91.8% on MATH benchmark

The frontier benchmark race used to belong strictly to a handful of well-funded tech giants, but smaller contenders are finding ways to get on the board.

Sonus AI kicked off the new year by unveiling Sonus-1, its new family of proprietary large language models designed to cover everything from rapid lightweight tasks to heavyweight reasoning.

Instead of dropping a single model, Sonus launched four distinct tiers. Sonus-1 Mini is built for fast responses and tight budgets. Sonus-1 Air sits in the middle, balancing performance with resource usage. Sonus-1 Pro serves as the high-horsepower tier for complex problems, while the top-shelf Sonus-1 Pro (w/ Reasoning) bakes in chain-of-thought processing to compete with the top proprietary labs.

Flexing big numbers on math and code

Sonus AI put most of the spotlight on the reasoning-enabled Pro model, rolling out a barrage of benchmark scores focused on quantitative problem solving and software development.

On standard math evaluations, Sonus-1 Pro with reasoning hit 97.0% on GSM-8k and a notable 91.8% on the MATH-500 benchmark.

For general reasoning and hard sciences, the flagship posted 90.15% on MMLU, 73.1% on MMLU-Pro, and 67.3% on the notoriously difficult GPQA-Diamond benchmark. It also scored 88.9% on the DROP reading and reasoning test.

Coding benchmarks tell a similar story. The model notched a 90.0% pass rate on HumanEval, reached 51.9% on LiveCodeBench, and posted 72.6% on the Aider-Edit code-editing harness. You can currently test the lineup directly at chat.sonus.ai.

Why it matters

Putting up a 90%+ score on both HumanEval and MATH-500 is no small feat. If these metrics hold up in real-world workflows, developers tackling algorithmic tasks and code maintenance might have an interesting alternative to the default big-name providers.

That said, we think it pays to stay cautious. Sonus AI has kept technical details under wraps: there are no whitepapers detailing parameter sizes, training regimes, or API pricing tiers just yet. Benchmark claims from fresh startups often face a gap when confronted with messy edge cases and extended conversations, so taking the web interface for a test drive is the best first step.