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.
AI or SI? The Techno-Autocracy Will Not Be Elected
On 29 September an executive order renamed artificial intelligence “Super Intelligence” across the US executive branch, and the president signed a White House accord on superintelligence with six companies that sets no penalties and gives the government no role in enforcing it. The institutional form of rule by a few is being written now, in documents, and democracy is not a party to them.
No penaltiesin the White House Accord on Super Intelligence, signed on 29 September 2026 by the president and six companies; Al Jazeera reports it is voluntary and gives the government no enforcement role
Who Limits AI in America
Executive Order 14365, the FTC's proposed policy statement and three federal dockets, read in the originalOpinionSep 5, 2026‘Like He Was the Weather’: The Escaped Agents Were Scouts, and What They Tested Was Us
Mücahit KayaSep 4, 2026Is GPT-6 Astra AGI? OpenAI Rated the Model's Hacking Critical and Chose Who Gets It
OpenAI's own launch and access documents, read against the EU AI ActSep 3, 2026‘But It Has to Be Said’: The Rogue AI You Can't Count, and the Bill You Can
Personal posts, institutional documentsReviews
AI products investigated from their own pricing pages, terms and refund policies. 79 so far.
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.
HeadshotPro
AI headshotA genuinely one-time purchase with nothing that renews, a published turnaround that tops out at roughly 15 minutes on Executive, refunds that dated user reports describe as quick once actually pursued, and a published per-person team price backed by an admin-curated style pool the rivals covered here do not document. Held back by a keeper rate independent sources put at roughly 10 to 33 percent against 100 percent money-back marketing, a guarantee forfeited the moment you download a single image from the order, headshot counts per tier that have fallen 25 to 65 percent since 2024 while the prices stayed flat, and a reversed vendor claim about whether uploaded photos train its models.
Decktopus
AI presentationA beginner-friendly, question-led deck builder with the only audience-question rehearsal mode among the AI presentation tools we cover and the cheapest entry price of the four, held back by a PowerPoint export that arrives as non-editable images, a credit allowance that caps the Pro plan at roughly 25 AI-generated decks a month, an entry tier Decktopus's own pages describe inconsistently, and auto-renewing subscriptions with no refund for unused time.
Aragon AI
AI headshotOne of the largest and fastest AI headshot generators, strong on style range, price stability, and a stated SOC 2 Type II and GDPR security posture, but held back by a wide gap between its advertised 90-95% keeper rate and the roughly 10-25% independent estimates (as low as 5% in one dated user report), real delivery that runs 1 to 3 hours rather than the advertised minutes, an over-enhancement pattern that dated reports say can alter your likeness, and refunds reported as hard to get once photos are generated.
D-ID
AI avatarCheapest way in, with real trust caveats
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
They publish what they're selling — not what you're buying
A pre-registered, documents-only audit of pricing transparency in 76 consumer AI products. Open data, open method, and a published record of every correction.
One dot per product audited, 76 across 15 categories
News
Only what changes a decision. Every story separates what the documents confirm from what they do not.
FTC investigates OpenAI and Anthropic a day after Trump's AI accord asked labs to police themselves
On 29 September six AI companies signed a voluntary, penalty-free White House accord. On 30 September the FTC confirmed it is investigating OpenAI, Anthropic and other labs over safety, with demands for documents and testimony being drafted.
Confirmed in the documents
The accord's text says each company should implement 'four layers of controls and audits': internal controls including to 'ensure that its models do not hack or access technical systems in unintended ways', an internal team, 'an independent external auditor or evaluator', and 'an independent committee of the board of directors'.
Not confirmed
Which companies besides OpenAI and Anthropic the FTC is investigating, and on which legal theory. Reports say 'other AI companies' and 'other frontier labs'; no outlet we read names them or cites a specific provision.
Guides
How to get a result with AI, with the cost and a way to check it worked.
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.
How to make ElevenLabs speech sound less flat
How to add sighs, pauses and emphasis to ElevenLabs speech using its documented Audio Tags and Voice Settings, and which popular tricks are not actually documented.
How to Create a Video in Midjourney (2026)
What a Midjourney video actually costs in GPU time, which plan tier you need, the documented steps to generate one, and how to verify the result yourself.
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.