GPT-6 Astra lands, and Harness rebuilds its foundations

This week the news was loud on both fronts. OpenAI shipped GPT-6 Astra and its president declared the AGI era open — while DeepSeek Harness quietly did its biggest internal rebuild since launch, breaking its session format and storage on purpose. One story is about capability; the other is about foundations. Both matter more than the headlines suggest.

The big story: GPT-6 Astra, and the AGI declarations

OpenAI released GPT-6 Astra — billed as "the model with the highest intelligence level yet," and the claims escalated fast: the OpenAI president said the AGI era has arrived, Jensen Huang declared "AGI has come," and a perfect Korean SAT score made the rounds as the proof point. Goldman Sachs called it the moment the AI bull market had been waiting for; others reached for "the iPhone moment."

The counterweight landed in the same week, from inside OpenAI itself: the chief scientist's warning that "we are building an alien brain that nobody understands — and nobody is ready." That's the split to watch: the marketing says arrival, the research leadership says we're still holding something we can't fully explain. Both can be true at once, and the tension between them is the actual story of this model cycle.

DSH this week: 0.1.2 goes RC, 0.1.3 begins

The part that actually matters: breaking changes

The features are nice. The breaking changes are the week's real DSH story — three of them:

• Session format v2. Old v0/v1 logs migrate through immutable adjacent generations — the right pattern for a format bump, and a reminder to back up before upgrading.
• A new SessionHandle API. agentLoop.create() is now async, and a session lock guarantees one process per session. Plugin and SDK authors will touch this.
• The SQLite backend is gone. The optional SQLite persistence backend was removed entirely — existing data isn't deleted, but you'll need an older version to export it. If you read our SQLite heads-up a few weeks ago, this is the sequel.

The team also flagged a known performance regression in 0.1.3-alpha.1 affecting some session loads, promised fixed next release. Candid, again — which is becoming a pattern worth appreciating.

Elsewhere in AI

• NVIDIA is reportedly buying Hugging Face for $12.9B — the "GitHub of AI" landing in the GPU giant's orbit, and open-source AI's biggest consolidation yet.
• Anthropic's IPO slides again, now expected before the US midterm elections, with the prospectus due late September.
• China approved its first AI-assisted innovative drug — a quiet milestone for AI in drug discovery, not just chat.
• Chinese LLMs keep leading the global usage charts — nineteen straight weeks, with Tencent's Hy4preview taking the top spot this round.
• "Model fatigue" is now a named thing. The majors' rapid-fire updates have analysts writing about users tiring of constant model churn.

Our take

GPT-6 Astra is a capability headline; Harness's week is a foundations headline, and foundations are where the durable lessons live. Breaking a session format and deleting a storage backend in an alpha is exactly how young infrastructure should move — early, loudly, with a migration path — and Harness keeps showing it can do that without pretending otherwise. The adjacent-generation migration and the "export with an older version" notice are the right kinds of honesty.

The OpenAI split is the week's real tension: "the AGI era has arrived" and "we're building a brain nobody understands" from the same company in the same week. Believe both, and calibrate accordingly — more capable models don't make the safety question smaller, they make it louder.

Practical line for the week: back up before you upgrade DSH (format v2 is a one-way door), and don't let GPT-6's fanfare distract you from the fundamentals — models churn, runtimes compound.

Sources

• github.com/deepseek-ai/deepseek-harness/releases
• 新华网 — OpenAI 发布 GPT-6 Astra · 澎湃 — GPT-6 来了
• NVIDIA 拟收购 Hugging Face
• AI 周报 0831-0906 — Anthropic IPO 推迟、AI 辅助创新药获批
• OpenAI 首席科学家预警