Tracker Shows GPT-6 Astra Launched Already Four Months Stale

A model's release date and its training cutoff are not the same thing, and a live tracker published this week on Hacker News puts a number on how far apart the two usually are.

The project, How Stale Is Your AI?, was submitted to Hacker News as a Show HN post by a user named joozio and reached 74 points in its first 19 hours. It was built by Pawel Jozefiak, who lists his role as AI marketing lead on his own site, and it runs live clocks counting up both dates for 20 models across eight labs.

What a training cutoff actually measures

The release date is when a lab shipped a model. The training cutoff is when the model stopped reading, so anything that happened afterward is simply absent from what it knows unless it goes and looks it up. The tracker's own framing: the gap between the two dates is how far behind a model already was on the day it launched, before a single user typed anything into it.

GPT-6 Astra is the tracker's starkest example. OpenAI released the model Sept. 3, and its own developer documentation states a knowledge cutoff of Apr. 30, a gap of roughly four months between the day training data stopped and the day the model reached customers. The same OpenAI page lists a 1,050,000-token context window for the model.

Which labs actually publish a cutoff

Of the 20 models the tracker follows, only 10 have a training cutoff their lab has published at all; the other 10 ship with a release date and nothing else on record.

Model	Lab	Released	Cutoff
Claude Haiku 4.5	Anthropic	Oct. 15, 2025	Jul. 2025
Claude Sonnet 5	Anthropic	Jun. 30, 2026	Jan. 2026
Claude Opus 5	Anthropic	Jul. 24, 2026	May 2026
Claude Fable 5.1	Anthropic	Sep. 1, 2026	Jun. 2026
Grok 4.6	xAI	Aug. 12, 2026	Feb. 1, 2026
GPT-6 Astra	OpenAI	Sep. 3, 2026	Apr. 30, 2026

Anthropic is the only lab in the tracker with a published cutoff on every model it ships; four of the site's 20 entries are Claude models for exactly that reason. Mistral AI, DeepSeek and Meta each have at least one model on the tracker with no published cutoff at all, meaning five of the eight labs the site follows withhold the figure for at least one current model.

The gap keeps mattering even with search

The Show HN thread that introduced the tracker split on whether any of this still matters once a model can search the web. One commenter argued that tool use and reasoning have made stale training data "much less of a problem than it used to be," while acknowledging it "still matters." Another countered bluntly that all of a model's reasoning still comes from its pretraining data, whatever it retrieves at query time. A third described a model that refused to accept a Venezuelan political development as real until pointed to a Reuters article, and said the same failure "hasn't been an issue in months" on the tools they now use. A fourth commenter described the opposite experience: ChatGPT insisting Joe Biden remained president well after the current administration had changed, despite the model being told the correct date, and concluded that a user still has to remember to request a live search rather than trust the model's default answer.

The tracker does not resolve that argument. What it adds is a number: on the models where labs publish both dates, the gap between them runs from about two months, in Claude Opus 5's case, to more than six months for Grok 4.6. Ten of the tracker's 20 entries carry no published figure to check at all, which is the tracker's real subject as much as any single model's staleness.

Both GPT-6 Astra's cybersecurity threshold and DeepSeek V4.1-Flash's own benchmark results were reported without a training cutoff attached to either model's announcement, which is exactly the gap the tracker was built to expose.