AI Tools Police
Signal over noise.
Independent AI reviews, research and analysis.
We investigate AI companies, products and claims from the primary documents, and publish what they say, what they contradict and what they leave out. Every claim carries its source. Nothing here is for sale.
Analysis
Arguments from primary documents, each with the case against it printed alongside.
The Skimpflation of Everything
First the price rose, then the packet shrank, and now the quality drops while nothing on the label changes. An opinion on how skimpflation escaped the supermarket into our buildings, screens and sentences, why AI is a kind of intelligence inflation, and what is still worth paying for.
50.9%of new English-language web articles were primarily AI-generated in Q4 2025, in Graphite's random sample of 55.4k Common Crawl URLs, before 49.9% in Q1 2026
Digital Socialism in the Age of AI
Mücahit KayaOpinionOct 8, 2026No, AI Doesn't Have an Off Switch
Mücahit KayaOpinionOct 2, 2026AI or SI? The Techno-Autocracy Will Not Be Elected
Mücahit KayaSep 12, 2026Who Limits AI in America
Executive Order 14365, the FTC's proposed policy statement and three federal dockets, read in the originalReviews
AI products investigated from their own pricing pages, terms and refund policies. 83 so far.
Photoroom
AI designThe specialist product-photo editor in our AI design reviews, with background removal and the other core tools free of credit cost on every paid plan and a yearly Pro price of $89.99. Held back by a free plan that Photoroom's own terms restrict to personal, non-commercial use, AI credit allowances that the vendor says vary by country, a Shopify integration that is missing from Pro and starts at Max, and a cancellation rule that keeps charging you if you cancel on the wrong platform.
Higgsfield
AI video generatorThe widest single window onto current video and image models we have reviewed, with a per-model credit table that lets you price a clip before you pay. Held back by refund terms that exclude renewals and any purchase with credits spent, 'Unlimited' plans on a queue whose speed the terms do not guarantee, and a run of dated billing and cancellation complaints that a 4.0 Trustpilot score does not show on its own.
Headshotsly
AI headshotA $4.99 one-time pack of 5 photos in 4K, the smallest paid entry pack in our headshot reviews, plus one free full-HD photo with no account or email, credits that never expire, and a refund window with no download or generation cut-off. Held back by having no independent user report at all on likeness or keeper rate, a cost of about $1.00 per photo that larger packs at all four rival tools undercut, a refund that is 'full or partial' at the vendor's discretion, no named company behind the site, no security certification, and no stated retention period for selfies on paid orders.
Dreamwave
AI headshotDreamwave is an AI headshot generator whose natural, recognisable results are praised in three dated Reddit buyer reports, and its 15 October 2026 pricing page lists one-time packages from $35. It is held back by a refund rule that ends once creation starts, a competitor's June 2024 roundup and two 2025 Reddit posts saying the price appeared only after photos were uploaded, a small and low Trustpilot record of 3.1 on 12 reviews, partial hit rates in the few batches users have described, and privacy terms that reach us mostly through a competitor's summary.
NaturalReader
AI voiceThe cheapest serious document reader we have reviewed, and a legacy accessibility tool that earns its place on reading support rather than voice quality. Held back by a licence that bars Personal audio, including the top $159 a year plan, from any public use, a no-cost product that cannot read a scanned page and contradicts the vendor's own homepage, a Commercial line sold in credits whose value is never stated, and a Trustpilot score tied with the discontinued LOVO AI for the lowest in our AI voice reviews.
BetterPic
AI headshotThe only tool in these rankings whose no-cost option is the real product rather than a cut-down side tool, with 4K on every tier including that one, the cheapest per-headshot rate of any individual pack covered here at about 65 cents on Pro, a published volume curve that falls to $13 a credit at 1,000, and the category's only priced developer SDK and API. Held back by the absence of any keeper-rate figure at all, advertised or independent, a refund that pays nothing once you download a single image, a free queue that sits behind every paid order, and a privacy policy that keeps uploaded photos for up to a year and both the trained facial model and the generated output for up to three, with deletion available on request rather than by default and retention extendable for unlimited time on some packages.
Frontier models
New models, read against their system cards, rate cards and outside evaluators. No star ratings, only what the record supports.
Gemini 4 Argon
Gemini 4 Argon costs $2/$10 per million tokens for now and $4/$20 later, and only some Fairwind security partners can use it. Which Google claims check out.
review, Sep 28, 202670.6% / 63.6%Terminal-Bench 4.0 for Claude Sonnet 5.5 at max effort: Anthropic's own run in Claude Code, and Artificial Analysis's run in its own harness, read 28 September 2026Claude Sonnet 5.5
Claude Sonnet 5.5 costs $2/$10 per million tokens, like GPT-6 Sol. Its Terminal-Bench lead over Opus 5.5 is max against xhigh. What it costs per task.
comparison, Sep 24, 2026$0.27 vs $0.71Zapier AutomationBench 1.0.6 cost per task at a level score: GPT-6 Sol at xhigh 33.2%, Claude Opus 5.5 at high 33.03% with default fallbacks, read 24 September 2026Claude Opus 5.5 vs GPT-6 Sol
GPT-6 Sol lists at half Claude Opus 5.5's price, yet cost per task turns on the effort setting. Scores and costs at each setting, plus Sol's 272K rule.
review, Sep 24, 202633.2% / 26.9%Zapier AutomationBench 1.0.6 for GPT-6 Sol at xhigh effort ($0.27 a task) and at medium, the default ($0.21 a task), read 24 September 2026GPT-6 Sol
GPT-6 Sol costs $2/$10 per million tokens. OpenAI benchmarks it against Claude Opus 5, not Opus 5.5. What outside evaluators measured, at which effort.
comparison, Sep 23, 202651.24 vs 51.15Artificial Analysis Intelligence Index v4.3.2 at each model's default effort: Claude Opus 5.5 at medium for $1.34 per task, Claude Fable 5.1 at high for $3.91, read 23 September 2026Claude Opus 5.5 vs Claude Fable 5.1
Claude Opus 5.5 lists at $4/$20 against Fable 5.1's $10/$50 and matches it at each default effort. Where Fable still leads, and which scores to trust.
comparison, Sep 23, 202659.6% = 59.6%Terminal-Bench 4.0 for Claude Opus 5.5 (max and xhigh) and GPT-6 Astra (xhigh) on Artificial Analysis's own harness, where Anthropic's table prints 66.4% against 57.9%, read 23 September 2026Claude Opus 5.5 vs GPT-6 Astra
Claude Opus 5.5 lists at 40% of GPT-6 Astra's price yet costs more per task at max effort. Same-setting benchmarks, cache and 272K pricing compared.
review, Sep 23, 202658 / 51Artificial Analysis Intelligence Index v4.3.2 for Claude Opus 5.5 at max effort and at medium, the default, read 23 September 2026Claude Opus 5.5
Claude Opus 5.5 costs $4/$20 per million tokens. Its benchmark table is max effort, its savings are quoted at medium. What the system card and evaluators add.
review, Sep 6, 20266 published scoresArtificial Analysis numbers for GPT-6 Astra on one index version, one per reasoning setting, read 6 September 2026GPT-6 Astra
GPT-6 Astra lists at $10 and $50 per million tokens, but a million output tokens runs $25 to $150 on the same model. What four outside evaluators measured, and at which setting.
Claude Opus 5.5, read against the record
We read the announcement, the 230-page system card, the rate card, seven outside evaluators and 21 dated user reports. Point at each launch claim to see what the record adds.
Claude Opus 5.5, as launched
The launch and its coverage, read against the record
58 in the headlinesThe score most coverage led with
$4 and $20 per million tokens40% of GPT-6 Astra's list price
The safety section of the launch postA few summary lines
That is max effort. At medium, the default and where the savings are quoted, Artificial Analysis scores it 51.Artificial Analysis Intelligence Index v4.3.2, read 23 September 2026
At max effort it costs $5.98 per task, against GPT-6 Astra's $3.26.Artificial Analysis cost per index task, v4.3.2, read 23 September 2026
The system card reports it acted on instructions planted in pasted text in about 7.4% of attempts at max effort, against 0% for Opus 5.Anthropic system card, pp. 125-126
EU AI Act Transparency Audit
An independent, documents-only audit of what 26 providers of general-purpose AI models publish under the EU AI Act's transparency duties: training-content summaries, telling people they face an AI, and marking AI output. Open data on Zenodo.
One dot per provider read, 26 in all
News
Only what changes a decision. Every story separates what the documents confirm from what they do not.
Claude Duet is "always on", but it checks your accounts every few hours: what Anthropic, OpenAI and Meta's background agents will do without asking
Anthropic's new Claude Duet watches your email, calendar and business apps and messages you when something needs you. Its product page says always on; its announcement says every few hours. Here is how Duet, ChatGPT Dots and Meta Muse split acting from asking, from each company's own documents.
Confirmed in the documents
Anthropic's announcement: 'Claude Duet checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster.' Its product page: 'While it's always on, it uses good judgment when it reaches out and knows what can wait.'
Not confirmed
How often Dots and Muse check your connected apps when you have not asked them anything. Neither company publishes a cadence.
Guides
How to get a result with AI, with the cost and a way to check it worked.
How to Write an AI Policy for Your Company
Write a company AI policy staff can follow: tool tiers, data rules, one owner, a review date, mapped to EU AI Act duties, NIST and ISO. Full template included.
How to Make a Song With Your Own Voice in Suno (2026)
How to record, verify and use your own voice in Suno's Voices feature after the v6 launch: plan requirements, the 15-second-to-4-minute sample rule, verification, model choice, rights, and how to check the result.
How to Control AI Crawlers With robots.txt
How to write, place and verify a robots.txt file that actually controls AI crawlers, including which bots don't obey it and which tokens aren't crawlers at all.
How to Improve Your Brand's Visibility in AI Search Engines
What Google, OpenAI, Perplexity, and Bing actually document about AI-search visibility, in their own words, plus how to check whether any of it gets your brand mentioned.
A source for every claim
Primary documents first. User reports cited with the account and the date.
Corrections in the open
Errors are fixed and dated, with what changed and why.
Never for sale
Commission never moves a score, a ranking, or what we cover.
AI use is disclosed
AI assists research and drafting. Every verdict is a named person's call.