What AI research tools actually cost, and what they do with your data
What's the short version?
Across 52 published tool dossiers and 222 models, three findings hold. Paid entry plans cluster tightly around $10 a month while frontier model input prices spread 2500×× from cheapest to dearest. Only 0 of 52 dossiers record a default-off stance on training with your content. And of 34 tracked benchmarks, 7 no longer separate models well enough to be read as live evidence.
Free to cite with a link. The desk would rather be quoted accurately than quoted exclusively.
What does an AI research tool cost?
Across 52 published dossiers, 44 carry a usable free tier (85 percent). 32 publish a structured monthly price; the median paid entry plan is $10 a month, running from $1 at the low end to $49 for Elicit.
| Field | Value |
|---|---|
| Published tool dossiers | 52 |
| Carry a free tier | 44 / 52 (85%) |
| Publish a structured monthly price | 32 / 52 |
| Median paid entry plan | $10 / mo |
| Cheapest paid entry plan | $1 / mo |
| Dearest paid entry plan | $49 / mo — Elicit |
How many AI tools say what they do with your data?
0 of 52 dossiers have an explicit, scoped no-training fact recorded. 0 have a recorded opt-out. The remaining 52 are not classified in the structured facts yet. That is a coverage gap, not proof that their terms are silent or that they train on your data.
| Field | Value |
|---|---|
| Training off by default, documented | 0 / 52 |
| Opt-out available, reader must find it | 0 / 52 |
| Not yet classified in structured facts | 52 / 52 |
| EU data residency documented for everyone | 0 / 52 |
| EU residency on enterprise plans only | 0 / 52 |
| SOC 2 documented | 0 / 52 |
| Local-first by default | 2 / 52 |
These rows describe what each dossier records after the desk read the vendor's published terms. They are not a legal audit, and a blank is a blank — not an accusation.
How wide is the price gap between frontier models?
Across 222 models, 175 publish a per-token price. The median is $1.25 per million input tokens and $4.40 per million output tokens. The gap between the cheapest and dearest published input price is 2500× — $0.03 against $75 for Claude Mythos Preview.
| Field | Value |
|---|---|
| Published models | 222 |
| Publish a per-token input price | 175 / 222 |
| Median price per 1M input tokens | $1.25 |
| Median price per 1M output tokens | $4.40 |
| Cheapest published input price | $0.03 |
| Dearest published input price | $75 — Claude Mythos Preview |
| Input price spread, cheapest to dearest | 2500× |
| Largest published context window | 10M — Llama 4 Scout |
How many AI benchmarks still separate models?
The atlas tracks 34 published benchmarks against 2055 sourced score rows covering 222 models. 27 are active. 4 are saturated and 3 are marked defunct, because a wall of near-perfect scores tells a reader nothing and leaving it up as live evidence is worse than removing it.
| Field | Value |
|---|---|
| Published benchmark pages | 34 |
| Sourced score rows in the ledger | 2055 |
| Models with at least one sourced row | 222 |
Saturation
| Field | Value |
|---|---|
| Active | 27 |
| Saturated | 4 |
| Defunct | 3 |
Contamination risk
| Field | Value |
|---|---|
| Low contamination | 16 |
| Medium contamination | 12 |
| Unknown contamination | 3 |
| High contamination | 3 |
Score rows come from Artificial Analysis and the Vals Index, and every row on a benchmark page shows its source and the date it was checked. The desk publishes no first-party benchmark scores until grading and two-reviewer sign-off are complete.
How fresh are these numbers?
30 of 52 dossiers had pricing re-checked within the last 90 days. Checks currently run from 2026-06-09 to 2026-09-18. Freshness is the whole product here: a pricing page with no timestamp is a guess wearing a suit.
| Field | Value |
|---|---|
| Pricing re-checked within 90 days | 30 / 52 |
| Pricing re-checked within 180 days | 52 / 52 |
| Oldest pricing check on any tool | 2026-06-09 |
| Newest pricing check on any tool | 2026-09-18 |
Evidence tier
| Field | Value |
|---|---|
| Editorial · hands-on | 41 |
| Synthesized | 11 |
Dossier status
| Field | Value |
|---|---|
| Solid | 43 |
| Flagship | 9 |
How are these numbers produced?
Every figure is computed at render time from the published tools in this atlas. Draft tools are excluded, because a draft is a stub and counting stubs would inflate every number on the page.
- Drafts excluded. Counts cover tools the desk has actually finished.
- Privacy rows report what a dossier records after the desk read the vendor's published terms — not a legal audit.
- Benchmark rows are third-party, each carrying a source label and a check date on its benchmark page.
- The edition label tracks the newest pricing check in the atlas, not the day you loaded the page. A page that renumbers itself daily without new evidence is false freshness.
- The same figures are available as JSON for anyone who wants to check the arithmetic.
How do you cite this report?
Quote any figure with a link back to this page. If a number here contradicts something you have measured, the desk wants to hear it — corrections get published.
APA citation
VerdictPal editorial desk. The Atlas Report, edition 2026-09. Retrieved from https://verdictpal.com/report
BibTeX
@misc{verdictpal-atlas-report-2026-09,
author = {{VerdictPal editorial desk}},
title = {{The Atlas Report}},
year = {2026},
edition = {2026-09},
url = {https://verdictpal.com/report},
note = {CC-BY 4.0},
}