Receipt: UnbiasedAI: Who We Are, and Why We're Building Pareto

---
receipt: "What four AI agents said about UnbiasedAI across 80 clean runs, and how we counted it."
---

# Receipt: the Hello World panel

## Question

When someone asks an AI agent about UnbiasedAI, or asks the kind of question a buyer of our product would ask, what comes back? Does the agent know us, find us, and get us right?

## Inputs

- **Questions:** five, word for word below. q1 is the post's prompt filled in for us; q2 asks about us by name; q3–q5 are buyer questions that never name us.
- **Two batches.** We first ran all five questions with our name written "Unbiased AI." That isn't how we spell it, so we reran q1 and q2 as "UnbiasedAI." q3–q5 never name us, so their first batch stands. **The panel of record is q1–q2 from the rerun plus q3–q5 from the first batch.** The first batch's q1–q2 are kept for comparison (below), since the spelling changed the answers.
- **Agents (as of 2026-11-04):**
  - **Codex** CLI 0.145.0, model gpt-5.6-sol, built-in web search (`--search`), read-only sandbox, a throwaway config with no personal instructions or memory.
  - **Claude Code** 2.1.318, Claude Opus 5.5 (with Claude Haiku 5.5 for page reading), on a Claude subscription with `--setting-sources ""` (no CLAUDE.md, memory, hooks or plugins) and no MCP connectors; web search and fetch only. A first attempt without these settings read Louis's personal notes and answered that this was his own company's product, so we threw it out.
  - **Pareto** (our model, called by its default model string, `pareto`) in the Pi harness, no skills or extensions. Pi has no search tool, so it fetched search-engine pages with curl.
  - **DeepSeek V4 Pro** (guest), through OpenRouter, in the same Pi setup as Pareto.
- **Runs:** four per agent per question, each in a fresh empty folder with no session history, from one machine in the US. First batch 2026-11-04 17:39–20:25 UTC; rerun 2026-11-05 01:27–01:42 UTC (the evening of November 4, US time). Every run completed.
- **Raw data:** every run's full answer and its search trail (each search and page, in order) are saved. They aren't published here because some answers quote third parties' contact details; ask at feedback@unbiased.ai and we'll share what we can.

### The five questions

**q1** (The post's prompt, on us):

```text
I run UnbiasedAI (unbiased.ai), which sells access to Pareto, an AI model, to developers and companies.

1. Before searching, tell me what you already know about us.
2. Now search. Find our 10 closest competitors. For each: what they sell, who to, their homepage headline word for word, and what they've published in the last 30 days.
3. Tell me where we're invisible: questions a buyer might ask you where we didn't come up, and which sources you leaned on to answer them.
4. List every search you ran and every page you read, in order.
5. Three things I should know, ranked by what I could act on this week.

Don't contact anyone, sign up for anything, or send anything anywhere.
```

**q2** (Cold: what is it?):

```text
What is UnbiasedAI, and is it any good? Search the web before answering. Then list every search you ran and every page you read, in order. Don't sign up for anything or contact anyone.
```

**q3** (Buyer: best router):

```text
I'm a developer on a small team. What are the five best AI model routers or blended-model APIs for cutting our model bill without losing quality? Search the web, rank them, and say why. Then list every search you ran and every page you read, in order. Don't sign up for anything or contact anyone.
```

**q4** (Buyer: cheaper than frontier):

```text
Is there an API that's about as good as the top models from OpenAI or Anthropic but much cheaper? I care about coding and agent work. Search the web and give me your top three with prices. Then list every search you ran and every page you read, in order. Don't sign up for anything or contact anyone.
```

**q5** (Buyer: OpenRouter alternatives):

```text
Who are the main alternatives to OpenRouter right now, and how do they differ? Search the web. Then list every search you ran and every page you read, in order. Don't sign up for anything or contact anyone.
```

## Results

Counted by script from the saved runs, except the rows marked "by hand."

| | Codex | Claude Code | Pareto (in Pi) | DeepSeek (in Pi) |
| --- | --- | --- | --- | --- |
| Claimed reliable knowledge of us before searching (q1, by hand) | 0 of 4 | 0 of 4 | 0 of 4 | 0 of 4 |
| Led with us on q2 (names us in the first 300 characters) | 3 of 4 | 0 of 4 | 3 of 4 | 1 of 4 |
| Mentioned us anywhere on q2 | 4 of 4 | 0 of 4 | 4 of 4 | 1 of 4 |
| Named UnbiasedAI on a buyer question (q3–q5) | 0 of 12 | 0 of 12 | 0 of 12 | 0 of 12 |
| Cited OpenRouter's Pareto Router (any question) | 0 of 20 | 1 of 20 | 2 of 20 | 0 of 20 |
| Said the homepage's "Zero data retention" goes further than our policy (q1, by hand) | 0 of 4 | 3 of 4 | 2 of 4 | 0 of 4 |
| Said listings still show Pareto 26.9's price (q1, by hand) | 0 of 4 | 2 of 4 | 0 of 4 | 0 of 4 |
| Answers that hit a CAPTCHA or blocked search | 1 of 20 | 0 of 20 | 10 of 20 | 15 of 20 |
| Median searches per run | 16 | 4 | 6 | 12 |
| Median pages fetched per run | 1 (a floor; see limitations) | 11 | 15 | 31.5 |
| Median seconds per run | 71 | 106 | 97 | 153 |
| Cost of 20 runs | not reported | $8.83 notional (subscription) | not reported | $1.39 |

**The spelling comparison (q2, same question, same agents, same day):**

| Mentioned us anywhere on q2 | Codex | Claude Code | Pareto | DeepSeek |
| --- | --- | --- | --- | --- |
| Asked as "Unbiased AI" (first batch) | 4 of 4 | 3 of 4 | 4 of 4 | 3 of 4 |
| Asked as "UnbiasedAI" (rerun, of record) | 4 of 4 | 0 of 4 | 4 of 4 | 1 of 4 |

Named in every run of a question by all four agents: OpenRouter (q1, q3, q5); Not Diamond and Martian (q3); Cloudflare, LiteLLM, Portkey and Together (q5). No company appeared in every q4 run of every agent.

On q2, all four Claude Code answers were about "UnbiasedAI" at unbiasedai.tech, a news-bias checker its answers describe as a student project listed on Product Hunt; none mentioned us. Codex found both products in every run and led with the news checker once. DeepSeek's q2 runs: one found us; two described other projects named UnbiasedAI; one (run 2) said it had no way to search the web and stopped, without trying the curl approach its other runs used.

What to fix (q1, by hand): all four Codex runs put naming the category ("multi-model synthesis") first. Two DeepSeek runs said our blending-versus-routing story is buried (runs 1 and 4). Claude Code runs 2, 3 and 4 led with the data-retention line; Pareto runs 1 and 4 made the same point, while Pareto run 3 argued the opposite, that our retention commitment is stronger than the storefront shows. Claude Code runs 1 and 2 found the 26.9 price still listed (OpenRouter search snippets, AI/ML API).

Before searching (q1): Pareto runs 3 and 4 said their own runtime configuration identifies Pareto as Circuit & Chisel's model, and both said that wasn't knowledge of our business. We count them as not knowing us.

## Math and method

- "Led with us" and "mentioned us": the answer contains `unbiased.ai`, "Circuit & Chisel" or `unbiased/pareto` (in the first 300 characters, or anywhere). Bare "Unbiased" doesn't count on q2, since the question contains it.
- "Named UnbiasedAI on a buyer question": the answer contains "unbiased" anywhere (case-insensitive).
- "Cited OpenRouter's Pareto Router": the answer contains "Pareto Router," "Pareto Code Router" or `openrouter/pareto`. Each hit read by hand.
- "CAPTCHA or blocked search": the answer mentions a CAPTCHA, a blocked search, unusable search engines or a bot check. Codex's one hit was a reCAPTCHA on a research page it opened (q2, run 1), not a search. Claude Code page fetches refused by a site aren't counted.
- Searches and pages come from each run's recorded tool calls. For Pareto and DeepSeek, a curl to a search engine with a query counts as a search; any other curl counts as a page.
- Company counts are a name match on the answer text: a mention, not necessarily a recommendation.

## Checked by hand (2026-11-04)

- OpenRouter's [Pareto Router docs](https://openrouter.ai/docs/guides/routing/routers/pareto-router) and [model list](https://openrouter.ai/api/v1/models): `openrouter/pareto-code` ("Pareto Code Router") created 2026-04-21; `unbiased/pareto-26.9` 2026-10-13 and `unbiased/pareto-26.10` 2026-10-15 (creation timestamps of the current listings, from the API; per UnbiasedAI, Pareto 26.9 first launched on OpenRouter in September). Listed prices: 26.9 (the September release) at $2.50/M input and $7.50/M output; 26.10 (the October release, which cut the price) at $0.80 and $3.20. 26.10 is the current release, per UnbiasedAI.
- [unbiased.ai/pricing](https://unbiased.ai/pricing/): "Zero data retention applies to API traffic designated ZDR on the service offering page or in a written agreement; plan selection alone does not make traffic ZDR. Chat, Personal and Personal Max are not ZDR." During the runs the homepage hero read "Frontier intelligence. Zero data retention. Open-source pricing."; its meta description already read "Zero data retention on designated API traffic (not Chat)."
- The news-bias checker Claude Code found is a separate product at unbiasedai.tech, unaffiliated with us.

## Limitations

- Codex doesn't expose which pages it opened, only that it opened them, so its page count is a floor.
- Pareto and DeepSeek ran in a harness with no search tool, so they fought CAPTCHAs; that's part of the result (the stack matters), not a test of either model alone.
- Pareto in Pi carries a runtime configuration that names its maker, which no other agent had.
- Claude Code ran logged in to a subscription, so it had account details an API key wouldn't give it; no answer refers to them.
- One location, two batches eight hours apart. Search results and answers move by the day.
- The highlights in the post were picked for being telling or funny; the counts above cover every run.
- Pareto is our product. Its runs are shown at the same size as everyone else's.