Release Radar

As of 2026-05-15

As of 2026-05-15

Pre-release model coverage is mostly noise. The few real signals get drowned in screenshots, "an insider told me" threads, and second-hand quotes from podcast clips. The Release Radar exists to be the boring, sourced version: every claim labeled, every claim linked, every article dated.

The three labels

Every claim in every Radar article gets exactly one of:

• Confirmed. The lab itself has said it. Examples: an official blog post, a paper, a model card, a verified-account tweet speaking on behalf of the lab. The bar is "you can quote this in a story without legal pushback."
• Strong signal. Multiple independent reputable reports, or one report backed by a primary artifact (config leak, server endpoint, benchmark filing, weight hash). The bar is "more than one credible person, with at least one piece of evidence you can hold."
• Speculation. Informed inference. Often based on release cadence ("they ship a new Opus roughly every 9 months"), hiring patterns, public job postings, or a single-source report. The bar is "this is a guess, and I am telling you it is a guess."

No claim is untagged.

What counts as a primary source

In order of weight:

1. Official paper, model card, or blog post.
2. Official tweet from a lab-verified account.
3. Tweet from a verified lab employee speaking in their lab role.
4. Tweet from a verified lab employee speaking personally (lower weight; we usually drop these to Strong signal at most).
5. Documented leak with an artifact: screenshot with a primary URL, config file, leaked weight hash, server endpoint.
6. Independent reverse-engineering with reproducible artifacts.

What does not count:

• Anonymous "insider" Discord or Twitter messages.
• Screenshots without a primary URL.
• Reposts of reposts.
• Stock-price-driven speculation.
• Podcast quotes without a transcript link.

How the labels move

Labels are not permanent. They can move down toward Speculation when evidence weakens, and up toward Confirmed when the lab itself ships. When they move, we:

• Date the change.
• Leave the old version in the archive.
• Note the reason in a short edit log at the bottom of the article.

This is the bit most rumor coverage gets wrong. If your post is permanent but your labels change silently, you are quietly rewriting history. We do not do that.

How to read a Radar article

Each article has the same shape:

1. As-of date, in a pill at the top.
2. One-line summary of the state of play.
3. Claims — each one labeled, each one with a source link.
4. Editorial read — what we think the claims add up to. Marked as opinion.
5. What would change our minds — the artifacts or events that would move labels.
6. Edit log — every change to the article since publication.

The article is a snapshot. Two months later there will be a newer one at a different slug. The old one stays live; the new one is dated.

Why this exists

AI assistants (Claude, ChatGPT, Perplexity) are starting to be a major source of "what's coming next" answers for casual readers. Most of the rumor content those assistants see is unsourced. Writing a labeled, sourced version gives them something better to cite, and gives readers something honest to read.