Daily digest
13 items · ~13 min · Week 2026-W37
Must-read (1)
ARC Prize shows GPT-6 Astra's 99.9% ARC-AGI-3 score came from OpenAI's custom harness, not the model
OpenAIARC Prize's own writeup shows Astra scored 62.7% (for $26,098) on ARC-AGI-3 under its standard harness but 99.9% (for $18,817) under OpenAI's Provider Adapter, which preserves opaque reasoning state between requests; even with reasoning effort set to zero the adapter scores 96.7%. Separately, Fortune found five metrics in OpenAI's launch post were revised after publication (Astra's hallucination rate moved 4.2% -> 2% -> 4.2%; Fable 5.1's FrontierMath 87.8% -> 78% -> 83%), and ARC Prize disavows the AGI framing: 'we lack evidence to call this AGI yet.'
Worth knowing (5)
Zhipu's first interim report as a listed company: revenue up 400%, API now 86.5% of the business
Zhipu AI (Z.ai)Zhipu's (02513.HK) first half-year report since its IPO, widely covered on Sept 6-7, shows H1 2026 revenue of RMB 954 million, up 399.7% year over year and already above full-year 2025. Open-platform/API revenue grew 2735.7% to RMB 825 million and now makes up 86.5% of revenue (15.2% a year ago), while net loss narrowed 12.1% to RMB 2.07 billion; ARR hit $1.6 billion by end of August.
OpenAI says its 'automated research intern' goal is met, targets fully automated researcher by March 2028
OpenAIIn a Sep 6, 2026 blog post ('Research acceleration: The view inside OpenAI') OpenAI says it reached the goal set in fall 2025 of an 'automated research intern' — a system handling well-defined tasks that would take a skilled researcher days — and is on track for an automated AI researcher by March 2028. Internal figures: 3.1 agent-workdays per human workday as of mid-August 2026, median researcher spending over $600/day on inference and the 90th percentile over $7,000/day.
OpenAI chief scientist warns in 'An Alien Mind' essay that no one is prepared for recursive self-improvement
OpenAIOpenAI's chief scientist published a Sep 6, 2026 essay arguing internal results make him 'strongly' expect current progress speed to continue into recursive self-improvement, writing 'I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.' He says OpenAI will keep building defenses and may 'unilaterally withhold further scaling as needed,' but argues broader interventions beyond lab self-restraint are required.
RoboTok: internet-scale retrieval of human demonstrations for dexterous robot policies
RoboTok is a data engine that, given a query manipulation video, retrieves relevant human demonstrations from web video for training dexterous robot policies. It learns a latent motion space from 3D hand trajectories in actor-centered frames, making retrieval robust to camera viewpoint, scene appearance and occlusion while remaining searchable at internet scale.
LatentPress: compressing context into continuous memory tokens a frozen LLM reads directly
A two-author paper introduces a third context representation beyond text and screenshots: continuous memory tokens written into a frozen decoder through its input-embedding interface, with no text reconstruction at inference. A small reader-matched writer (4.2M-26.2M adapter parameters, about 0.1% of the decoder) compresses 4-16x; on LongMemEval it scores 0.504 at 7.7x compression versus 0.490 for uncompressed evidence, far above text summaries (0.184) and OCR-based compression.
For reference (7)
Morgan Stanley briefing: MiniMax ARR at $800M and Zhipu at $1.6B as model competition turns into a tiered elimination round
MiniMax / Zhipu AIA Morgan Stanley private briefing reported on Sept 6-7 puts MiniMax's end-August ARR at $800M (weekly basis) with a raised analyst view of $1.3B by year-end, and Zhipu's ARR at $1.6B in August ($2B+ on a weekly basis) with year-end guidance lifted to $2.4B. The same briefing dates GLM-6 for October on GLM-5.3-Flash's architecture and lists MiniMax M3.1, H3.1 and M3 Pro for H2 plus a 10T-parameter model planned next year; MiniMax shares rose nearly 8% on the news.
Alibaba's Qwen Office passes 30 million users within about a month, more than half from enterprises
Alibaba (Qwen)Coverage on Sept 4-7 reports Alibaba's Qwen Office (千问办公) agent reached 30 million users within roughly a month of launch, with over 52% of accounts tied to real enterprise organizations, driven by DingTalk distribution rather than ad spend. The product introduced paid subscription tiers on August 10 — the second major Chinese AI platform to do so after ByteDance's Doubao — and positions Qwen3.8 as the model layer across DingTalk, Alibaba Cloud and CRM/ERP/OA systems.
GPT-6 Astra reaches the $20 ChatGPT Plus tier after a rollout Altman had to apologize for
OpenAIFollowing up on the Sep 3 flagship launch: on Sep 6, 2026 OpenAI began rolling GPT-6 Astra out to $20 ChatGPT Plus subscribers; an OpenAI X post confirmed availability 'to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex. It's also live in the API.' The rollout to paying users was delayed days after launch, prompting a public apology from Sam Altman for the 'messy' release; free users still only have GPT-5.6 Sol and GPT-5.
Editable Visual Design: coding agents generate layered HTML/CSS visuals instead of flattened images
A team around Junyan Ye proposes a new visual-generation paradigm where a coding agent uses a VLM as the creative brain and an image model as an on-demand visual-world simulator: the agent generates isolated assets, writes native HTML/CSS, and refines the design against rendered feedback. The result is editable, layer-decoupled artifacts, in contrast to end-to-end diffusion output (GPT-Image-2, Nano Banana) which yields flattened bitmaps with error-prone text.
Last Translation Benchmark: human-authored tasks that still break frontier translation models
A large consortium (Vilem Zouhar, Niyati Bafna, Stella Biderman and others) releases LTBv1, a live collection of human-authored, peer-reviewed texts, images, audio and video that break leading machine translation models as standard benchmarks saturate. Each example ships with handcrafted verification rules describing concrete failure modes, enabling reliable, reward-hack-resistant, actionable evaluation.
RealSWE: coding-agent benchmark rebuilt around real, casual user requests
The authors quantify the gap between SWE-bench-style tasks and real usage: 88% of real prompts carry only a bare problem statement (versus 7% of benchmark problems), and 87% are casually written versus 94% formal in benchmarks. RealSWE derives 381 multi-variant task families from SWE-bench Verified and Pro to test compositional behavior under realistic request styles.
Krea connects to ChatGPT for image and video generation inside conversations
KreaOn September 5, 2026, Krea announced a ChatGPT integration that lets users generate images and video with Krea's model lineup directly inside a ChatGPT conversation. Existing Krea moodboards and styles carry over into the chat workflow, and outputs can be saved back to the user's Krea library.