Astra changed the bar. This closes the gap — on whatever model you already use.
GPT-6 Astra (OpenAI, Sept 3 2026) pushed computer use, scope discipline, and ambiguity handling further than any model before it — enough that OpenAI's own president called it the start of an "AGI era," a claim independent evaluators immediately split on. What actually moved wasn't magic — it's a specific, repeatable set of behavioral rules. Parallax turns those rules into a harness, and pairs it with 36 real prompts, for the 10 leading closed- and open-weight chat models.
One harness. Ten models. Same discipline every time.
This is the protocol itself — the exact text prepended to every task below, live-typed for whichever model is currently selected in the library.
Eight verified capabilities, with the honest caveat attached
Every card below is grounded in OpenAI's own announcement and system card, cross-checked against independent evaluators. The "reality check" line is there on purpose — the AGI framing is contested, and pretending otherwise would make this toolkit less useful, not more.
Prompt library: harness + 36 tasks, rendered per model
Pick a model. Every prompt below is the full, ready-to-paste text — the Astra-derived harness (tuned for that model's specific quirks) plus the task, already combined.
Run a harnessed prompt against a free-tier model
Bring your own free API key from a provider below. The request goes straight from your browser to that provider — this site never sees or stores your key unless you tick "remember," which just puts it in your own browser's local storage.
Model line-ups and free-tier availability change constantly. The model slug field is editable — if a suggested slug 404s, check the provider's current model list and swap it in.