Pareto

Pareto is Unbiased’s proprietary composite AI offering, exposed on the Unbiased platform and third-party gateways under the model identifier pareto. Rather than routing each prompt to a single backbone, Pareto engages multiple frontier and open-weight language models in parallel on every request and synthesizes their outputs into one unified answer. Unbiased emphasizes that this design is not a conventional complexity-based router: behavior is consistent across easy and hard prompts, which helps preserve prompt-cache economics on long, multi-step workflows.

The stable Pareto release (documented as version 26.9) is positioned for research, software engineering, and agentic use cases while supporting general chat. It accepts text and image inputs and returns text. Official published evaluations on Unbiased’s model card report strong agentic coding performance on DeepSWE (tied with several frontier comparators in their table), competitive multimodal reasoning on MMMU-Pro, and expert-level reasoning on Humanity’s Last Exam without tools. Pricing is token-based (published at $2.50 per million input tokens and $7.50 per million output tokens, with discounted cached input), and the service does not ship downloadable weights.

Reproducibility-minded teams can rerun overlapping suites with Unbiased’s public pareto-evals harness, which measures accuracy and cost per task against OpenAI-compatible endpoints. Vendor-reported leaderboard numbers should be read alongside that harness and each benchmark’s own methodology; Unbiased has noted that measured per-task costs and composite scores were not published for the 26.9 comparison table.