Pareto 26.10
Pareto 26.10 (unbiased/pareto-26.10) is Unbiased AI's multimodal composite model — a multi-model ensemble that runs several constituent models on one request and synthesizes a single answer. It offers a 1,048,576-token context window, 131,072-token max output, image input, tool calling, and aggressive $0.80/$3.20 per-million-token pricing, positioned as near-frontier capability at a fraction of flagship cost.
Empirical Evaluation Results
Architectural Profile & Capabilities
Pareto 26.10 is the generally-available (GA) release of Unbiased AI's "Pareto" line. It was declared out of preview on 15 October 2026; the October 1, 2026 `pareto-26.10-preview` build preceded it under the same underlying model string. Unbiased states that the GA build retains the same four published benchmark scores, the same 1M-token context window, the same 131,072-token maximum output, and the same underlying composite architecture as the October 1 preview — i.e., the preview-to-GA change was commercial (label and pricing) rather than a model swap. Architecture and hosting: Pareto is explicitly described by its vendor as a *composite* model — several underlying models process the same request and the system produces one synthesized response. It is therefore best classified as a multimodal mixture-of-models / multi-agent ensemble with a synthesis layer, NOT a sparse mixture-of-experts (MoE) transformer with token-level expert gating, and not a diffusion LM. Unbiased does not publish the component-model roster, weighting, router policy, or active-parameter count, and no source retrieved identifies the constituent base models by name or version. Importantly, OpenRouter does not perform model routing for this slug: the listing has a single provider (Unbiased) and requests are forwarded directly to it, so all composition happens provider-side. Practically, this means the ensemble's behavior — including latency profile, token accounting, and failure modes — is opaque to the API caller. Commercials and limits: $0.80 per 1M input tokens, $3.20 per 1M output tokens, with cached input at roughly $0.03 per 1M tokens. Context 1,048,576 tokens; max output 131,072 tokens; text + image input, text output; tool calling supported. Vendor-reported per-task economics: ~$0.24/task on DeepSWE v1.1, ~$0.48/task on Terminal-Bench 4.0, ~$0.008/task on HLE (text-only), ~$0.004/task on GPQA Diamond. Benchmark status: Unbiased's model card publishes four scores explicitly labeled as *preliminary, vendor-run* measurements dated October 1, 2026 — GPQA Diamond 92.4%, HLE (text-only / no tools) 49.9%, DeepSWE v1.1 69.9%, and Terminal-Bench 4.0 50.8% — and the GA page states these remain preliminary preview measurements, not final GA results. Third-party aggregators (BenchLM, Puter, AI-on-Mac, OrcaRouter) reproduce these same four numbers and no independent evaluator runs were found. No published scores exist for Pareto 26.10 on SWE-bench Verified/Pro, AIME 2026, ARC-AGI-2, FrontierMath, BrowseComp, OSWorld, MCP-based TAU suites, MRCR long-context, or AA Intelligence Index; those are therefore omitted rather than proxied. Note also that these scores are NOT the September `Pareto 26.9` figures (DeepSWE 74, Terminal-Bench 4.0 51, MMMU-Pro 78, HLE no-tools 49, ArXivMath 88) — 26.9 is a distinct predecessor variant and must not be substituted. The headline positioning is cost-efficiency: the vendor contrasts 69.9% on DeepSWE v1.1 at $0.24/task against Fable 5 at 70.0% for $13.50/task, i.e. near-parity coding performance at roughly 1/56th the per-task cost. Operator guidance: treat the four published scores as vendor self-reported point estimates with no disclosed harness details (prompt scaffolding, retry policy, tool versions, judge model), and re-validate on your own task distribution before committing.
Pareto 26.10 is not a single autoregressive dense transformer checkpoint and not a sparse MoE with token-level expert gating; it is a composite inference system in which several constituent models are run against the same request and a synthesis layer merges their outputs into one answer. Unbiased explicitly calls it a 'multimodal composite model' and confirms the composition happens inside Pareto rather than at the OpenRouter routing layer (OpenRouter lists exactly one provider and forwards requests directly). Component models, parameter counts, active-parameter ratios, attention variant (MLA/GQA), and reasoning-token budget are all undisclosed — no retrieved source names the underlying base models, their versions, their weights in the ensemble, or the router/planner/judge policy. Externally observable interface: text + image input, text output, tool calling supported, 1,048,576-token context, 131,072-token maximum output, $0.80/$3.20 per 1M input/output tokens with ~$0.03/1M cached input. Because synthesis requires multiple internal generations, expect multi-pass token consumption and elevated first-token latency relative to a single-model deployment of equivalent context.
Recommended Workloads & Primary Use Cases
- •Vendor-run, preliminary numbers only: all four published scores (GPQA Diamond 92.4%, HLE no-tools 49.9%, DeepSWE v1.1 69.9%, Terminal-Bench 4.0 50.8%) are self-reported October 1, 2026 preview measurements that the GA page still flags as non-final. Harness details — prompt scaffolding, retry/repair policy, tool versions, grading/judge model — are undisclosed, and no independent (Artificial Analysis, LMArena, SWE-bench leaderboard) verification was located. Re-baseline in-house before stake-bearing deployment.
- •Opaque ensemble = unattributable failures: because Pareto composes multiple undisclosed models provider-side, you cannot tell which constituent produced a given answer, cannot pin or A/B a specific component, and cannot reason about its token budget. Debugging regressions is effectively black-box; expect inconsistent output style and formatting across consecutive identical prompts.
- •Long-context degradation and TTFT: the 1,048,576-token window is nominal and unpublished for retrieval accuracy — no MRCR v2 / RULER / needle-in-haystack results exist for this variant. Combined with multi-pass synthesis, TTFT on large prompts can dominate wall-clock; chunk and summarize rather than assuming full-window fidelity, and cap max output well below 131,072 unless genuinely needed.
- •Reasoning headroom is the weak axis, not coding: 49.9% on HLE (no tools) is materially behind frontier reasoning flagships, so do not delegate hardest open-ended scientific/olympiad-style problems here. There are zero published results for AIME 2026, FrontierMath Tier 4, ARC-AGI-2, BrowseComp, OSWorld Verified, SWE-bench Pro/Verified, or AA Intelligence Index — treat those capabilities as uncharacterized rather than inferred from sibling or predecessor variants (Pareto 26.9 scores are a different model and must not be substituted).
- •Version-string and tool-semantics hazards: OpenRouter exposes several overlapping slugs (`unbiased/pareto-26.10`, dated `pareto-26.10-20261015`, `pareto-26.10-preview`, legacy `pareto-latest`, predecessor `26.9`). Unbiased states the preview and GA builds are the same weights, but pin the dated GA string in production to avoid silent substitution. Also confirm tool-call schema conformance and parallel-call behavior empirically — composite systems frequently normalize or drop malformed tool payloads differently from single-model providers, and 'HLE (text-only)' vs 'HLE (with tools)' framing must be tracked explicitly in any internal eval suite.
Calls to Pareto 26.10 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Unbiased.