Astra’s real ship is the Codex harness: notes instead of amnesia
GPT-6 Astra is already in the API. The change that matters for agents is experimental context notes — searchable across windows — going default in the coming weeks. Grok 4.7 is still cooking.
Feature2 minIn-depth
Maya Chen2 min read
- OpenAI
- Codex
- GPT-6 Astra
- Grok
- Agents
- Harness

Photo: OpenAISource page
The morning desk already has GPT-Live-1 as a mouth and Grok 4.7 as a delay. The agent story is the harness underneath Astra. When a Codex session fills the context window, models have historically compacted: summarize, discard, hope the next loop still knows why the last fix failed. Astra can keep notes across windows and search earlier turns — requirements, test output, tool traces — even when those details never made the summary. OpenAI says the feature is experimental in config.toml today and becomes the default for Astra in the coming weeks.
That is a product sentence about memory, not a benchmark. OpenAI also says the Codex harness, paired with Astra, finishes Mind2Web computer-use work 1.9 times faster than the GPT-5.6 Sol path. Cognition is wiring Astra into Devin’s harness on launch day. Those are vendor quotes. Treat them as the bid. The mechanism you can inspect is the notes file: if the model forgets a constraint, it has to know to look it up. Unknown unknowns still win.
The September calendar is models plus the things that hold them
Grok 4.7 was supposed to be a Saturday model. Musk pushed it for more reinforcement learning — length penalties, task management, a habit of giving up early. Reporting puts it near 2.1 trillion parameters against 4.6’s 1.5 trillion, slightly slower to serve. Grok 5 is still an end-of-year target with no card. Grok Bot remains the derivative that already has a conference next week. Hermes, on the Nous side, already speaks Codex OAuth and Grok’s Responses API. The stack a shop actually runs is model plus harness plus bot, not a single checkpoint.
- GPT-6 Astra: API / Azure / Bedrock; Codex notes opt-in, default later.
- Codex harness: less compaction, searchable prior windows, faster computer-use loop on OpenAI’s numbers.
- Grok 4.7: slipped past Sep 12 for RL; Grok 5 promised before year-end, unspecified.
- Derivatives already shipping: Grok Bot, Codex, Devin, Hermes — the wrappers, not the weights.
If you only track launch dates you will write that OpenAI shipped and xAI blinked. If you track harnesses you will write that the next month is about whether an agent can remember a failing test after the window turns over. That is the less glamorous race. It is also the one that decides whether Grok Bot Galaxy and Codex pull requests are demos or labor.


