About MetaPrompt

MetaPrompt covers the technology of AI for business and consumers, every layer at equal weight: AI models and their APIs for language, voice, image, video and avatars; agents and the protocols they use; benchmarks and measured progress toward AGI; the chips, memory, data centers and power beneath them; the token demand and inference costs that tie it together; the technology of robots and humanoids; and the export controls, regulations and standards that shape it all.

The site is built as a stack of ten layers. Every post sits on one of them, and that layer's color marks it wherever it appears: its chip, its table, its charts.

Every figure is labeled

Every figure on MetaPrompt carries one of three labels, shown by shape as well as by name. The three are never mixed in one ranked column.

Measured

Tested independently, by a benchmark suite, a test lab or researchers other than the maker.

Vendor-stated

The maker's or vendor's own figure, from a datasheet, a price page or an announcement.

Estimate

Modeled or inferred by a named source, such as data-center capacity estimated from satellite images and permits, or a model's undisclosed training compute.

Chip throughput always names its number format and sparsity, such as "FP8 dense". FP4 figures are never ranked against FP8 ones, nor sparse against dense. Data-center and power capacity is split by status, and only like is ranked with like:

Operational

Built and running.

Under construction

Being built, with construction confirmed by the operator, a regulator or an independent source.

Announced

Planned and announced, with no construction confirmed yet.

Token volumes name the period they cover and the company's own definition of a token.

Benchmarks, handled with care

• Version and harness, always. A benchmark score names the benchmark's version and the harness that ran it, because the same model can score very differently on either.
• Contamination and saturation are flagged, with the source that reported them, when a benchmark's tasks have leaked into training data or top models have run out of headroom.
• Vendor-run results are never ranked against independent ones.
• Arena ratings are preference votes. Elo-style arena scores measure which answers people preferred, not accuracy, and are labeled that way.
• AGI is measured, not predicted. MetaPrompt reports measured progress on tests such as ARC-AGI and Humanity's Last Exam. Forecasts of AGI or ASI always name who made them, their definition of AGI and their date, and are never stated as fact.

Three formats

• Rankings sort one measurable thing, such as price per million tokens, memory capacity, benchmark score or power, and are named for what they measure ("cheapest", "largest", "fastest"), never "best".
• Features look at one question with sourced analysis: what changed, why, and what the numbers show.
• Explainers show how one thing works, from a tokenizer to high-bandwidth memory, each built around a table.

How we source

• Primary sources first: makers' datasheets, documentation, price pages and release notes, company filings, regulators and standards bodies, peer-reviewed papers and reproducible benchmarks. Each page lists what it used, and every claim is footnoted.
• Dated and versioned: every table says when its figures were checked, and statements about a model, chip or standard name its version.
• As the source states it: prices stay in the vendor's own unit (per million tokens, per minute, per image) and currency. Conversions only ever appear in a labeled extra column.
• Trimmed, never padded: if reliable data runs out at 12 rows, the table stops at 12.

What we don't do

• No investment advice. No stock prices, price targets, buy or sell calls, or stock rankings. Company capex, revenue and market share appear only as data about the technology, exactly as filed or reported. AI-themed crypto tokens are never priced or ranked; on MetaPrompt, "tokens" means the units language models read and write.
• We explain, we don't hand out exploits. Posts may explain how jailbreaks, prompt injection and voice or face cloning work, with sources. They never publish working jailbreak or safety-bypass prompts, or step-by-step ways to clone a real person's voice or face without their consent.
• No advice for your own situation: nothing here is medical, legal or financial advice.
• No news articles, reviews, opinion pieces or interviews, and no coverage of adult or NSFW AI or of deepfake tools aimed at real people.

Not affiliated

MetaPrompt is an independent site. It isn't affiliated with any AI lab, chipmaker, cloud provider, data-center operator, regulator or researcher it covers. Names, logos and trademarks belong to their owners.