# Unbiased

> Unbiased makes AI providers compete for every request. Pareto is our first model, not the right model for every task.

## Available now

- Pareto API access. Onboarding available, optional.
- Public API base: https://api.unbiased.ai/v1 (OpenAI-compatible chat completions; also speaks Anthropic /v1/messages). Model string: pareto. API keys are issued with a platform.unbiased.ai account.
- Credit packs: $10, $100, or $1,000 inside a platform.unbiased.ai account; the $5,000 pre-buy is set up directly (info@unbiased.ai).
- Pareto 26.8 rates: $1.25 in / $0.15 cached / $6.25 out per MTok.
- Access and use are governed by the Terms: https://unbiased.ai/terms/
- Contact: info@unbiased.ai

## Pages

- [Homepage: what Pareto is, why it exists, and how to get a key and install](https://unbiased.ai/)
- [Get started: hand setup to your agent](https://unbiased.ai/start/) — agent-facing setup instructions live at https://unbiased.ai/start/agent.md
- [How a Pareto answer gets made: mechanism, testing, latency](https://unbiased.ai/how/)
- [Pareto Model Card and final seven-benchmark comparison](https://unbiased.ai/model-card/)
- [Pareto 26.8 launch page](https://unbiased.ai/pareto-26-8/)
- [Pareto pricing: token rates and credit packs from $10 to $5,000](https://unbiased.ai/pricing/)
- [Who builds this: founders and backers](https://unbiased.ai/about/)
- [Apply for onboarding](https://unbiased.ai/apply/)
- [FAQ](https://unbiased.ai/faq/)
- [Blog: measured writing on model pricing and infrastructure](https://unbiased.ai/blog/)
- [Fable 5 pricing with real measured bills](https://unbiased.ai/blog/fable-5-pricing/)
- [OpenRouter alternatives compared with real fees and measured latency](https://unbiased.ai/blog/openrouter-alternatives/)
- [Fable 5 alternatives: what actually comes close](https://unbiased.ai/blog/fable-5-alternatives/)
- [Claude Opus 5 pricing: same rates, bigger bills](https://unbiased.ai/blog/claude-opus-5-pricing/)
- [Claude Opus 4.8 pricing: $5/$25 and why it matters](https://unbiased.ai/blog/claude-opus-4-8-pricing/)
- [Claude Sonnet 5 pricing: $2/$10 until Aug 31](https://unbiased.ai/blog/claude-sonnet-5-pricing/)
- [GPT-5.6 Sol pricing: $5/$30, the output-token catch](https://unbiased.ai/blog/gpt-5-6-sol-pricing/)
- [GPT-5.6 Luna pricing: the volume-tier bake-off](https://unbiased.ai/blog/gpt-5-6-luna-pricing/)
- [Opus 4.8 vs Opus 5: same price, 77% different bills](https://unbiased.ai/blog/opus-4-8-vs-opus-5/)
- [GPT-5.6 Sol vs Terra vs Luna: which tier, explained](https://unbiased.ai/blog/sol-vs-terra-vs-luna/)
- [LLM API pricing guide 2026: measured bills](https://unbiased.ai/blog/llm-api-pricing/)
- [LLM gateways guide 2026: types, fees, latency](https://unbiased.ai/blog/llm-gateways/)
- [LiteLLM alternatives: when self-hosting isn't free](https://unbiased.ai/blog/litellm-alternatives/)
- [Portkey alternatives: control plane, gateway, neither](https://unbiased.ai/blog/portkey-alternatives/)
- [OpenAI API alternatives 2026, priced and compared](https://unbiased.ai/blog/openai-api-alternatives/)
- [Anthropic API alternatives 2026, priced and compared](https://unbiased.ai/blog/anthropic-api-alternatives/)
- [The Verbosity Index: tokens billed for identical work](https://unbiased.ai/blog/verbosity-index/)
- [LLM API cost calculator](https://unbiased.ai/tools/llm-cost-calculator/)
- [AI gateway fee calculator](https://unbiased.ai/tools/gateway-fee-calculator/)
- [Terms & Policies (terms, privacy, data policy, analytics preferences)](https://unbiased.ai/terms/)
- [The Standings: where every AI model stands, with source and age on every number](https://unbiased.ai/delta)
- [Glossary: LLM pricing and gateway terms, defined with measured numbers](https://unbiased.ai/glossary/)
- [MTok pricing explained, with measured bills](https://unbiased.ai/glossary/mtok-pricing/)
- [Prompt caching: rates and a measured 2.6x saving](https://unbiased.ai/glossary/prompt-caching/)
- [Batch API: the 50% LLM discount, explained](https://unbiased.ai/glossary/batch-api/)
- [LLM gateway: four types and measured latency](https://unbiased.ai/glossary/llm-gateway/)
- [BYOK (bring your own key): definition and fees](https://unbiased.ai/glossary/byok/)
- [Blended model: how it differs from a router](https://unbiased.ai/glossary/blended-model/)
- [Silent quality regression: definition and evidence](https://unbiased.ai/glossary/silent-quality-regression/)
- [Shadow mode: how model comparison works](https://unbiased.ai/glossary/shadow-mode/)
- [Pareto by role: agent builders, eng leads, platform teams](https://unbiased.ai/for/)
- [Pareto for engineering leaders: a legible LLM bill](https://unbiased.ai/for/eng-leads/)
- [Pareto for platform teams: the gateway decision](https://unbiased.ai/for/platform-teams/)
- [Pareto for agent builders: context is the meter](https://unbiased.ai/for/agent-builders/)
- [Guide: choosing an LLM gateway in 2026](https://unbiased.ai/guides/best-llm-gateway-2026/)

## Evidence policy

Product performance and savings claims should identify model versions, methods, sample sizes, cost basis, latency, limitations, and places where Pareto loses.

Pareto 26.8: Beats Fable 5 and GPT 5.6 Sol at 75–85% less. Frontier performance for less. Published 26.8 slate (score; measured $ per task, same harness, vs Kimi K3, Claude Opus 5, GPT-5.6 Sol, Claude Fable 5, DeepSeek V4F-0731):

- GPQA-Diamond: 90.4; $0.0093 per task.
- Humanity's Last Exam: 47.0; $0.0261 per task.
- arXivMath: 69.2; $0.0377 per task.
- DRACO: 58.0; $0.0773 per task.
- MMMU-Pro: 77.9; $0.0039 per task.
- SWE-Bench Verified: 86.0; $0.2708 per task.
- Terminal-Bench 2.1: 86.0; $0.2690 per task.

Notes on this slate: costs are measured per-task bills from the same runs; per-benchmark methods and sample sizes are pinned in the public harness configs; SWE-Bench Verified is the one exact tie (86.0 for Pareto 26.8 and Kimi K3; Kimi's bill $2.37, ours $0.27); DeepSeek V4F-0731 publishes no MMMU-Pro result, so it has no seven-row mean; the full six-model matrix is at https://unbiased.ai/model-card/. Measured drawback: on very difficult reasoning tasks Pareto can run up to 3x slower (measured July 2026 against the prior baseline; a comparison against this slate publishes when measured — see the model card's Latency section).

Public evaluation harness: https://github.com/circuitandchisel/pareto-evals

## Machine-readable data

- [The Verbosity Index: measured output tokens per identical task](https://unbiased.ai/data/verbosity-index.json)
- [LLM list prices, verified and dated](https://unbiased.ai/data/prices.json)
- [Measured per-task bills and latency](https://unbiased.ai/data/measured-bills.json)
- [Full site content in one file](https://unbiased.ai/llms-full.txt)

## Original measurement and corrections (2026-08-25)
- https://unbiased.ai/blog/verbosity-index-edition-2/ : Verbosity Index Edition 2: the tokenizer confound
- https://unbiased.ai/blog/we-graded-the-openrouter-listicles/ : We graded the OpenRouter listicles
- https://unbiased.ai/blog/the-price-increase-that-did-not-happen/ : The price increase that did not happen
- https://unbiased.ai/blog/why-your-bill-does-not-match-your-math/ : Why your bill does not match your math
- https://unbiased.ai/data/verbosity-index.json : Verbosity Index Edition 2, both indices, machine-readable
- https://unbiased.ai/data/verbosity-index-e2-raw.json : all 72 per-run token and character counts