06-reference

every vibe check gpt 6 2 astra sol

2026-11-17·reference·source: Every·by Katie Parrott
openaigpt-6-2astrasolmodel-comparisonevalsvibe-check

Why this is in the vault

Every's two-week, seven-tester verdict on OpenAI's GPT-6.2 Astra and Sol: good at bounded, checkable work, still behind Claude and Gemini for open-ended writing and building - a fresh external read on where OpenAI's tier sits against the Fable 5.2 generation.

The core argument

Only the free preview rendered; the full Vibe Check is subscriber-only, so this note covers the teaser and not the benchmarks or screenshots. OpenAI launched Sol alongside Astra, pitching an everyday workhorse plus a model for the hardest projects. Every's verdict: Astra is useful for clearly defined coding and operations work, but its longer writing loses the thread and ambitious builds need prodding. Sol gave most testers even fewer reasons to switch.

Named previews of individual tests: Mike Taylor (head of evals) saw Sol plan a trip well enough to give it his first perfect score on that task, yet he still will not adopt it in daily work. Kieran Klaassen's coding builds showed Astra's speed paired with a problem in the tests it left behind. Dan Shipper's manuscript review had Astra miss a factual contradiction that Fable caught. Tyler Nishida's interface designs were inventive but had unfinished interactions and stubborn visual defaults. Closing heuristic: try Astra when you can define the job and check the result; for open-ended writing and building the team still reaches for Claude and Gemini.

Bias note: Every is a Claude-heavy shop and its testers are daily Claude users, so "still reaches for Claude" is a stated house preference as much as a measurement.

Mapping against Ray Data Co

Medium mapping. The "define the job and check the result" heuristic matches RDCO's dispatch design, where bounded single-item transforms go to cheap fan-out agents with an explicit output shape and open-ended judgment work stays on the stronger tier (feedback_delegation_model_effort_pairing). Astra's tests-it-left-behind problem is a concrete case for keeping an independent fresh-eyes critic (feedback_fresh_eyes_subagent_for_own_artifacts) rather than trusting a fast model's self-written verification, and the manuscript miss that Fable caught is the same argument for critics being a different model or zero-context. No action needed on model routing today: nothing here argues for moving any RDCO workload to Astra or Sol. Re-check if the full piece's benchmarks become available.

⚠️ Sponsorship

No third-party sponsor. The email is Every house promo: it drives to the paywalled Vibe Check, cross-links Every's own writers and the earlier Fable 5.2 Vibe Check, and carries the standard subscribe upsell. No bearing on the findings beyond the Claude-leaning house angle noted above.

Related