Pareto is a composite model from Unbiased: instead of one set of weights, it runs a mix of frontier and open-source models against the same request and returns the best result. The model is chosen per request and does not switch mid-conversation, which is what lets prompt caching keep working across turns.
26.10 is a preview of the next Pareto release. It changes two things that matter: the context window grows from 262,144 to 1,048,576 tokens, and the price falls to roughly a third of the September release.
Unbiased publishes four figures for this release, from runs dated 1 October 2026:
Unbiased also reports the cost of each run: $0.24 per task on DeepSWE v1.1, $0.48 on Terminal-Bench 4.0, $0.008 on Humanity's Last Exam and $0.004 on GPQA-Diamond.
These are the vendor's own numbers and Unbiased marks them preliminary — “results are from October 1, 2026 runs and may change before final publication.” They have not been independently reproduced.
Four times the context
1,048,576 tokens against 262,144 on the September release — enough for a whole repository, a large document set or a long agent transcript in one prompt.
A lower rate
Input drops from $3.4385 to $1.10032 per 1M tokens, output from $10.3155 to $4.40128.
Preview status
Unbiased states the preview may change without notice and points to the stable Pareto release for behaviour you can pin.
Repository-scale coding
Read whole codebases in one prompt and work across files without chunking.
Agentic workflows
Tool calling and streaming suit agent loops that plan, call tools and iterate; the two agentic benchmarks Unbiased reports are the ones it leads with.
Research over long sources
Papers, contracts and report sets fit inside the context window intact.
Multimodal analysis
Charts, screenshots and scanned pages go into the prompt alongside the text around them.
Sources: https://unbiased.ai/ · https://openrouter.ai/unbiased/pareto-26.10-preview
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
import requests r = requests.post( "https://api.aimlapi.com/v1/chat/completions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "unbiased/pareto-26.10-preview", "messages": [ { "role": "user", "content": "Hello!" } ] }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "unbiased/pareto-26.10-preview", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"unbiased/pareto-26.10-preview","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Cached input |
Preview release — Unbiased states behaviour may change without notice. Cached input is billed separately.
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| DeepSWE | 69.9% | Agentic software engineering: resolving real repository tasks end to end | Source | October 1, 2026 |
| Terminal-Bench | 50.8% | Autonomous shell/terminal task completion | Source | October 1, 2026 |
| Humanity's Last Exam | 49.9% | Expert-level questions across many domains | Source | October 1, 2026 |
| GPQA Diamond | 92.4% | Google-proof graduate science questions (hardest subset) | Source | October 1, 2026 |
Published by Unbiased from runs dated 1 October 2026 and marked preliminary; Terminal-Bench 4.0 and DeepSWE v1.1. Not independently reproduced.
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Pareto 26.10 Preview This page | Long-context research, coding & agentic workflows | |||
| Research, coding & agentic workflows |
It is a preview release of Unbiased's composite model. Rather than one set of weights, Pareto runs a mix of frontier and open-source models against the same request and returns the best result. The 26.10 preview raises the context window to 1,048,576 tokens, four times the 262,144 of the September release.
Two things: the context window grows from 262,144 to 1,048,576 tokens, and the price drops — $1.10032 per 1M input tokens against $3.4385 for the September release. It is a preview, so behaviour can change; the stable Pareto card remains the one to build against if you need that.
Unbiased marks this release as a preview that may change without notice, and points to the stable Pareto release for behaviour you can pin. Use the preview to evaluate the larger context window and the lower price; keep production on the stable model until Unbiased promotes this one.
On AI/ML API it is $1.10032 per 1M input tokens and $4.40128 per 1M output tokens, with cached input at $0.041262 per 1M — about a twenty-sixth of the input rate, which makes repeated context cheap to re-send.
Unbiased reports 69.9% on DeepSWE v1.1 (agentic coding), 50.8% on Terminal-Bench 4.0 (agentic terminal use), 49.9% on Humanity's Last Exam text-only, and 92.4% on GPQA-Diamond. The runs are dated 1 October 2026 and Unbiased marks them preliminary — they may change before final publication, and they have not been independently reproduced.
Text and images in, text out. Screenshots, diagrams and scanned pages can go straight into the prompt alongside the surrounding text. It supports streaming, tool calling and vision.
No. Pareto picks a model per request and does not switch mid-conversation, which is what lets prompt caching keep working across turns. Cached input is billed separately at $0.041262 per 1M tokens.
Through the standard chat completions endpoint at https://api.aimlapi.com/v1/chat/completions with the model id unbiased/pareto-26.10-preview. Any OpenAI-compatible client works — point the base URL at AI/ML API and use your AI/ML API key.
