Methodology

Sources

• News: RSS feeds from technology publishers and the AI sections of general news outlets.
• Companies: official blogs and newsrooms of AI companies such as OpenAI, Anthropic, Google DeepMind, Meta and Mistral.
• Research: arXiv, Hugging Face Papers and lab research blogs.
• Policy: AI-specific government feeds: GOV.UK artificial intelligence news, the UK AI Security Institute, US Federal Register documents about AI, NIST and the EU digital strategy.
• Funding: news reports of funding rounds.
• Models: OpenRouter’s public model list for API prices and Hugging Face’s trending list for open-weight models.

Every source, with the result of its latest check, is listed on the sources page. A source is removed when it stops publishing, blocks automated requests, or mostly publishes things that are not about AI.

Collection

• News feeds and arXiv are collected every 15 to 20 minutes.
• Company, research, funding and government sources are collected every two hours. Each run records whether the source worked and how many items it returned.
• Model prices and trending models refresh every six hours.
• The AI Wire stores the headline, link, date and the short text the publisher includes in its feed. It never stores full articles. Every story links to the original.
• Items more than 60 days old when first collected are skipped, so archives don’t reappear as new.
• Dates: the publisher’s date is used. When a source gives no date, the time The AI Wire first saw the item is used instead, and a later check never makes an item look newer.

AI briefs

Items get a short brief written by a language model (currently DeepSeek V4.1 Flash). The model sees only the headline and the text the publisher provided. It returns a one or two sentence summary, an optional “why it matters” line, a topic, the organisations named, and a score from 0 to 1 for how much the item is about AI.

Before anything is published, software checks the output:

• Any number in the brief must appear in the source text, or the brief is discarded.
• Briefs that talk about the input (“this article discusses…”) are discarded.
• A brief is shown only when the publisher gave enough text to summarise; otherwise it would just repeat the headline.
• Items scoring below 0.4 for AI relevance are left out of AI-focused lists.

Briefs are always labelled as AI-written and link to the source. They can still be wrong; the original report is the authority.

Grouping coverage into stories

When several outlets report the same event, a language model reads the last four days of headlines and groups articles about the same specific announcement, release, deal or incident. A group needs at least two different outlets. Its combined headline and summary pass the same checks as briefs.

Story score

Top stories are ranked by a score out of 100, built from three measured parts:

A story reported by four or more outlets gets the full 45 points; a single-outlet story gets none. The recency part halves every 36 hours.

Funding

• Reports of the same company raising money within 30 days are merged into one round, so a round covered by five outlets counts once.
• Roundup articles listing several companies are not treated as a round.
• Valuations are kept separate from the amount raised.
• Totals are in US dollars, converting other currencies at the European Central Bank’s daily reference rates. Until those rates have loaded, fixed approximate rates are used instead (1 EUR = $1.08, 1 GBP = $1.27). Individual rounds are shown in the currency reported.
• Signed-in readers can choose a display currency in their account preferences. Prices and totals are then converted from US dollars at the same ECB reference rates, and are approximate.
• Amounts that look implausible are shown on the round but left out of totals.
• A round is marked verified when the company is in The AI Wire’s company registry or at least two sources reported it.

Models

• Prices are per million tokens, as listed on OpenRouter. They can differ from buying directly from the provider.
• “Available since” is when a model appeared on OpenRouter, which can be later than its launch.
• Price changes are recorded from when tracking began; earlier changes are not known.
• Trending open models are Hugging Face’s trending list, saved once a day so rank changes can be shown.
• Benchmark scores are not shown. There is no reliable public source to collect them from automatically, and The AI Wire does not publish numbers it cannot check.

Company profiles

A profile brings together funding rounds, headlines naming the company, its models on OpenRouter, and its research and announcements. Everything shown is a record or a count. There is no company score or ranking.

Signals and topics

Every item is filed under one of nine topics, chosen by its AI brief or, when it has none yet, by the collector’s keyword rules. Technologies are the six technology topics. Signals sort items by what happened (a model release, a funding round, a company announcement) and count how many sources back each one. They are counts, not predictions, and carry no confidence score.

Mentions this week are compared with last week only when every source behind a term was already being collected last week. Otherwise the week before collection began would count as a week of no mentions, so no change is shown.

What The AI Wire does not do

• No human reviews stories before they appear.
• No original reporting: every item comes from a publisher, company, lab or government source.
• No confidence scores, market forecasts or company rankings.

Spotted something wrong? See the contact page.