AI Receipts

Issue #001 · · By The AI Receipts team

Gemini 4 Argon is cheap. Until Nov 30

Gemini 4 Argon's 50% coupon itemized, OpenAI's 722 AI-written math papers audited, the White House's 'not optional' incident reporting, and what 2.2% of PNC households paying for AI really means.

Hello, auditors. An Anthropic testing model submitted 20 US visa applications through the State Department’s public form: 19 in August, one in May. None were processed (Axios).

Somewhere, a consular officer asked for a photo and got a server rack. Welcome to AI Receipts: the week’s AI claims, held up to the numbers.

This week in 60 seconds

  • 🧾 Gemini 4 Argon ties GPT-6 Astra on an independent index. It’s cheap until Nov 30.
  • 🧮 OpenAI dropped 722 AI-written math papers. We counted what’s machine-checked.
  • ⚖️ The White House says AI incident reporting is “not optional.” No rule, no penalties yet.
  • 💳 2.2% of PNC households pay for AI. Not 2.2% of America.

The Big One: Gemini 4 Argon and the 50% coupon

What happened. Google made Argon globally available on Oct 13 via the API and for AI Pro and Ultra subscribers. Artificial Analysis scores it 53 on its Intelligence Index, tying GPT-6 Astra and edging GPT-6.1 Sol (52) (Artificial Analysis). Claude Opus 5.5 still leads at 58 (TNW).

Why it matters. Argon’s real trick is honesty. Its hallucination rate (how often it guesses wrong instead of saying “I don’t know”) is 15%, versus 51% for Astra. But it gets fewer answers right, 50% vs 63%, so the combined scores land side by side: 42 vs 43 (Artificial Analysis). Argon isn’t smarter. It just knows when to shut up. Rare in models, rarer in people.

Our take. Google’s model card has Argon beating Opus 5.5 on the Vals Index, 68.9% to 67.0%. A 1.9-point win on a vendor-picked table is a coin flip with good marketing. The independent result is real, though: Google is back in the top tier.

What to watch. The price. Argon launched at an introductory $2 / $10 per million input/output tokens (a token is roughly three-quarters of a word). On Dec 1 it doubles to $4 / $20 (Google).

🧾 Mini-receipt: “Argon is cheaper than GPT-6 Astra” At the discount, Argon costs $1.99 per index task, 60% of Astra’s $3.26. But it burns ~62,000 output tokens per task vs ~27,000. The saving is the coupon, not efficiency. At full price, Artificial Analysis puts it at $3.98, about 1.2x Astra (Artificial Analysis). Verdict: ⚠️ Partly. True until Nov 30.

Cheap the way a gym’s first month is cheap.

🧾 Receipts: “OpenAI’s AI solved hundreds of open math problems”

Also published as a standalone receipt: OpenAI’s 722 math papers, audited.

The claim. On Oct 6, OpenAI released 722 manuscripts in 372 “result families” from an internal model, each billed as resolving or advancing an open problem (It Does What Now?).

Line item Number
Problems posed ~4,000
Manuscripts published 722
Failures disclosed none
Manuscripts in Lean catalogue 162 (~22%)
Families with some Lean check 235 of 372
Catalogue review status “unchecked”
Peer-reviewed 0
Model name not disclosed

Sources: The Frontier, It Does What Now?

Lean is a proof language where a computer checks every step. The catch: humans still have to confirm the Lean statement says what the paper claims.

Plasma AI counts 40 Erdős problems touched, with claimed complete solutions to 17. It rebuilt the 12 that came with full Lean proofs and now lists those as solved (Plasma AI). OpenAI’s own README warns that “some of the unformalized results could have issues.”

Verdict: ⚠️ Partly. “Hundreds of results” is a fair paper count. “Solved” is premature: under a quarter of manuscripts have catalogued Lean proofs, the catalogue is unreviewed, and nothing is peer-reviewed. And the most telling number, how many of ~4,000 attempts failed, is missing.

Show your work. It’s literally math.

Rapid fire

  • ⚖️ White House: AI firms must “immediately disclose incidents.” No executive order, rule, incident definition, or penalty yet (Axios, AGI Watch).
  • 🧠 Meta Muse Wonder costs $4 / $20 per million tokens, same as Opus 5.5; “Wonder Full” is subscription-only (Meta).
  • 💳 “2.2% pay for AI” is PNC card data on its own customers in May. Payers spend ~$31/month (PNC, PDF).

Try this: the 30-second claim audit

Paste any AI announcement into your chatbot with this prompt:

You are a skeptical analyst. For the announcement below, list:
1. Every numeric claim, and who measured it (vendor or independent).
2. What the number is compared against, and whether that comparison is fair.
3. Any price that is temporary, introductory, or plan-gated.
4. What is NOT disclosed (failures, sample size, model name, dates).
End with a verdict: Checks out / Partly / Doesn't add up, in one sentence.
Do not add facts that aren't in the text.

[PASTE ANNOUNCEMENT]

Try it on Google’s Argon post and see if it spots the Nov 30 date. Then check its work, obviously.

Tools (checked Oct 14)

  • Claude for Google Workspace: Claude in Docs, Sheets, and Slides. Public beta on all paid plans (Anthropic); Pro is $17/mo billed annually or $20 monthly (pricing).
  • Gemini 4 Argon API: $2 / $10 per million tokens until Nov 30, then $4 / $20 (Google).
  • Muse Code: Meta’s coding agent. Plans $5 / $15 / $50 / $200 a month; Wonder Full needs $50 or up (The Threshold).
  • erdosproblems.ai (Plasma AI): marks each Erdős problem open, claimed, or solved from the evidence (Plasma AI).

Spot a wrong number? Reply to the email or see our corrections policy. Prices change: see the AI model price tracker.