How to read an AI claim

Three questions before you trust a claim

1. Ask for the measurementWhich named test or private set, on which date, scored by whom.
2. Ask for the billList input and output per million tokens on the vendor page. A loop pays both again.
3. Ask what you cannot call yetLive API, lab claim, replay, or a file you download — keep those doors separate.

Worked checks from this week’s briefing

1. Model claimWorked check: Carbon’s introductory list is $2/$10 per million tokens — not on the public API yet; DeepMind’s Index v5 claim of 59 is DeepMind’s report. Test named. Price named. Door separate. Briefing
2. Harness claimWorked check: NVIDIA’s NemoClaw post says LangChain’s Deep Agents tune for Nemotron 3 Ultra cut cost about 10× per run versus leading closed models — that delta is the harness, not a new weight file. Briefing
3. Weight-file claimName who measured the weights — an RNG recipe that regenerates 730,368 values is not a trained checkpoint claim. Briefing

Four list prices to keep in your pocket

Toy bill: one million input tokens plus one million output at the vendor’s published list rate. Real jobs are usually heavier on input.

• GPT-6.1 Terra (OpenAI): $1 input / $5 output per million tokens · toy 1M+1M bill $6.
• Claude Opus 5.5 (Anthropic): $4 input / $20 output per million tokens · toy 1M+1M bill $24.
• Claude Haiku 5.5 (Anthropic): $0.10 input / $0.50 output per million tokens · toy 1M+1M bill $0.60.
• Gemini 4 Argon (Google): $2 input / $10 output per million tokens · toy 1M+1M bill $12.

Read on the vendor pages dated with this models chapter. Full table and how a loop multiplies the bill.