Astra Takes WebDev. helloai's Elo Gate Still Holds

OpenAI shipped GPT-6 Astra on September 3, 2026 at the same $10 input / $50 output per million tokens as Claude Fable 5.1. The API id is gpt-6-astra. Official docs list a 1,050,000-token context window, 128K max output, and cache reads at $1 — four times Fable 5.1's $0.25 cache-read rate. Prompts over 272K input double input and cache rates and multiply output by 1.5 for the whole request. Fast mode is 2x Standard: a Fable-class sticker with a worse cache line.

The independent signal that moved this week is Code Arena, not OpenAI's scorecard. Arena.ai's September 5, 2026 WebDev board lists gpt-6-astra-max at 1797 on 1,199 votes, rank 1. Claude Fable 5.1-max is 1762 on 2,275 votes. Grok 4.6-high is 1625. Muse Spark 1.3 (xHigh) finally has a code-arena slug at 1622. The text-overall snapshot is still dated September 2 — the day before Astra launched — so gpt-6-astra is not on helloai's Elo table yet.

Vendor benches are noisier and should be read as vendor benches. OpenAI reports Terminal-Bench 4.0 at 57.9% for Astra versus 55.8% for Fable 5.1 and 37.3% for GPT-5.6 Sol. It reports ARC-AGI-3 at 99.9% using its own Responses API harness, with a footnote that two settings were changed to match real-world performance. Greg Kamradt of the ARC Prize Foundation said Astra surpassed their human action-efficiency baseline on 96% of levels. Those numbers are why the launch post talks about AGI. They are not why helloai would add a row.

helloai's tracked six is gated on LMArena text-overall Elo, sustained two weeks, with a public API. Astra clears the API and the 200K context bar but fails the Elo bar: it is four days old, the live text board has not listed it, and 1,199 WebDev votes is a debut sample. The set is already at six, so admitting OpenAI as a new provider would force a drop — Grok 4.6 at 1461 is still Preliminary and is the lowest tracked Elo. The owner already held GLM-5.3-max, 1482 on 7,668 votes, rather than drop Grok. Astra does not get a shortcut.

The routing question this week is not whether Astra is good at computers. It is whether a same-price Mythos competitor with a 4x worse cache read belongs on a directory that currently leads coding with Fable 5.1 and daily use with Muse Spark 1.3 at $1.25/$4.25. If the next text-overall snapshot puts Astra inside 25 points of the floor and keeps it there for two weeks, the six-model rule has to name who leaves. Until then, /api/recommend still has no OpenAI row, and the WebDev crown sits on a model the table does not track.