Unbiased previews Pareto 26.10 with lower API pricing
Pareto 26.10 Preview lowers API prices, with defined retention rules and important limits on version selection and benchmark comparisons.
Unbiased made Pareto 26.10 Preview available on October 1, 2026, with lower token prices and a new retention tier. The announcement lists access through its own API and OpenRouter. Its serving stack remains changeable during preview.
Pareto is a composite service. The company’s terms describe combining outputs from several underlying language models into one response. Developers send one request while Unbiased manages the underlying providers and can change the blend.
Pricing and access
The rate card lists these prices in US dollars per million tokens.
- $0.80 for input tokens
- $0.03 for cached input tokens
- $3.20 for output tokens
Direct API usage draws from prepaid credits, and new accounts need manual approval before access is enabled. The company says subscriber allowances will double when the release becomes stable. That increase does not apply during preview. The launch details explain the timing.
Check compatibility and version selection
OpenRouter lists a 1,048,576 token context window, text and image inputs, and text output. Tool calling is supported, but enforced JSON output is not. Its catalog recommends Pareto 26.9 for stable behavior.
The developer documentation describes OpenAI Chat Completions and Anthropic Messages formats, including streaming. Its quickstart still recommends the generic pareto alias and calls direct access a private beta. The announcement specifies a separate preview identifier, but ByteForward has not tested that exact endpoint.
The same docs say model discovery is unavailable and unknown model names can still be served and billed as Pareto. A successful request therefore cannot establish which version answered. Confirm version selection before comparing results or moving a workload.
Retention has a defined scope
Unbiased says its OpenRouter and Cloudflare traffic uses the new zero data retention tier. Under its retention terms, designated traffic cannot have prompt or response content retained after inference or used for training. Usage metadata may remain. Other subscription traffic and saved workspace content follow different rules. This is a company commitment, not an independent security audit.
Benchmarks remain preliminary
The model card reports 69.9 percent on DeepSWE version 1.1 at a mean cost of $0.24 per task. It also says task counts and cost methodology require confirmation. Historical DeepSWE comparators used only 30 tasks, while other comparison scores came from different evaluators and test setups.
Those limitations leave comparative performance unresolved. For developers, the useful next step is a controlled trial that checks task success, actual spending, version selection and the retention terms for the chosen access route.
