Daily AI RoundupsFriday, 4 September 2026

GPT-6 Astra leads this edition as Simon Willison maps its pricing and long-context performance, ARC Prize finds 62.7% standard-harness and 99.9% state-preserving ARC-AGI-3 scores, and OpenAI reports a Lean-verified improvement to Rankin's prime-gap estimate. Bloomberg says DeepSeek plans a 160,000-chip Huawei cluster, while IFM releases six K2 Horizon models and Google DeepMind ships WeatherNext 3. Security and tooling signals include Anthropic's dark-web distillation claims, Coder's registry compromise, H Company's NeoMME encoders, Microsoft's MAI-Transcribe-2 and new GPT-6 Astra constants in OpenAI's Go SDK.

• frontier models
• AI benchmarks
• open-weight models
• AI compute
• AI security
• coding agents
• multimodal retrieval
• scientific reasoning

Stories in this roundup

1. 01
   Simon Willison reports GPT-6 Astra will cost $10 per million input tokens and $50 per million output tokens, with near-perfect long-context recall but a lower Intelligence Index score than Claude Fable 5.1.Simon Willison
2. 02
   ARC Prize finds GPT-6 Astra scores 62.7% on ARC-AGI-3 with its standard harness and 99.9% with OpenAI's state-preserving Provider Adapter, while using fewer actions than the human baseline on 96% of levels.ARC Prize
3. 03
   Bloomberg reports DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at an Inner Mongolia data center, potentially creating one of the largest Huawei AI clusters.Bloomberg
4. 04
   IFM releases K2 Horizon, a connected fleet of six open models from 375B-A23B to 0.9B, the three smallest setting state of the art at their scales.IFM
5. 05
   Anthropic says foreign adversaries are using the dark web to access American AI systems and train competing models, selling cheaper copycats.CNBC
6. 06
   Google DeepMind releases WeatherNext 3, a global weather model trained on real-time satellite observations that produces fresh forecasts every hour.Google DeepMind
7. 07
   Attackers compromised Coder's registry infrastructure and used it to push malicious modules, BleepingComputer reports.BleepingComputer
8. 08
   Armature measured 16,893 coding-agent sessions to map which tools Claude Code, Codex and Cursor choose to install for a given service.Armature
9. 09
   OpenAI president Greg Brockman says Astra was trained on more than 100,000 GPUs and that OpenAI reassigned 25% of its production engineers to use the model for finding and fixing vulnerabilities.Stratechery
10. 10
    Gary Marcus calls GPT-6 Astra a genuine advance but challenges Greg Brockman's AGI claim, saying the robustness of its symbolic world models remains unknown.Gary Marcus
11. 11
    Benedict Evans argues AI will not sweep away enterprise software because the hard part is identifying useful workflows and institutionalizing them across organizations, not generating code.Benedict Evans
12. 12
    OpenAI's Go SDK v3.56.0 adds model constants for the new GPT-6 Astra model.GitHub
13. 13
    Cerebras adds Qwen 3.8 27B to its public inference API, serving it at roughly 1,500 tokens per second.Cerebras
14. 14
    H Company releases NeoMME, open 260M- and 800M-parameter multimodal encoders that share one transformer across text and images; its 260M retriever processes about 51 document pages per second.H Company
15. 15
    OpenAI reports GPT-6 Astra proved a new lower bound on gaps between consecutive primes that improves Rankin's estimate by a factor of log₂ X, with a complete Lean proof released alongside.OpenAI
16. 16
    Microsoft releases MAI-Transcribe-2 in public preview, adding speaker diarization, word-level timestamps and support for 60 languages at $0.10 per audio hour through 2026.Microsoft