Pareto 26.10 Preview
About this model
Pareto 26.10 Preview is Unbiased's preview release of its Pareto composite language system, available via the Unbiased platform and aggregators such as OpenRouter under the model ID unbiased/pareto-26.10-preview. Rather than a single open-weight checkpoint, Pareto routes each request through a blend of frontier and open-source models and returns one unified answer, targeting research, software engineering, and agentic workflows while keeping API pricing aligned with open-model tiers.
The preview expands context to 1,048,576 input tokens with up to 131,072 completion tokens per request, accepts text and image inputs with text output, and supports tool calling. Unbiased positions 26.10 as faster and cheaper than Pareto 26.9, with attested zero-data-retention on OpenRouter and Cloudflare traffic. The release date is October 1, 2026; behavior may change until the preview graduates to a stable slug.
In preliminary October 1, 2026 vendor evaluations, Unbiased reports strong scores on GPQA-Diamond (92.4%), Humanity's Last Exam text-only (49.9%), and DeepSWE v1.1 agentic coding (69.9%). These figures are self-reported, may be revised before final publication, and have not been independently reproduced. No public Hugging Face weights or parameter count are published; deployment is API-only.
Benchmark Scores
Technical Specs
- Architecture: Composite routing
- Context Window: 1,048,576 tokens
- Input Modalities: text, image
Hardware Requirements
- Compute: API only
Pricing
| Input | Output | Currency |
|---|---|---|
| 0.80 / 1M tokens | 3.20 / 1M tokens | USD |