The Intake — Tuesday, September 22, 2026

Atria Dawn Preview is a 744B open-weight model built for agentic tasks. Shanghai AI Laboratory released it to Hugging Face on September 11, 2026. No press release accompanied the upload. A technical report with 140 listed authors appeared on arXiv three days later. The architecture is Mixture of Experts, built on GLM-5.2. It is MIT-licensed with a 256K-token context window. Training used a Verifiable Experience Pipeline. The method uses executable environment feedback rather than human annotation as the primary training signal.

Vendor-reported benchmark scores: DeepSearchQA 96.0%, BrowseComp 92.5%, AutomationBench 53.8%, CyberGym 86.5%, SWE-bench Pro 59.6%. No independent third-party evaluation has been published as of September 22. If you're evaluating open-weight agentic models at this scale, these scores are self-reported. The absence of independent verification is the relevant constraint before acting on them.

If you're running the Pi coding agent and watching token costs accumulate across long task loops, NVIDIA, NTU, and MIT researchers published an MIT-licensed extension this week that changes those numbers.

SoL-Pi adds four mechanisms to Pi 0.85.1: Action Fusion, ObservationPack, Evidence-Preserving Reducer, and Online Context Compact. The researchers evaluated it on EdgeBench's 51-task suite. Comparison models were GPT-5.6 Sol and Opus 5. Token traffic fell 44.7–49%. API costs dropped by approximately one-third. Task completion retained roughly 94% of baseline scores. The npm package is MIT-licensed. An accompanying arXiv paper documents the methodology. The discovery process spanned 535 executable environments and more than 3,000 runs.

If you're already on Pi 0.85.1 and your cost model is dominated by token volume across agent loops, SoL-Pi is the extension to add.

If you've been building on the Messages API and working around automatic compaction — or writing detection logic to catch when compaction happened — Anthropic moved that control to the caller side on September 14.

The new interface uses the compact-2026-09-04 beta header. Callers send a compaction parameter and receive back a signed compaction block. The block summarizes prior messages and replaces them on subsequent requests. Recent turns are preserved word-for-word. Existing cache validity and thinking blocks carry through the transition.

If your session logic depends on knowing exactly when compaction occurs, on-demand compaction is the path where you decide the moment rather than the platform.

Grok 4.7 is xAI's latest model, released September 21, 2026. It carries a 500K-token context window. The model accepts text and image input. xAI did not disclose parameter counts. Access is via the xAI API, Cursor, Grok Build, and third-party platforms. The pricing structure doubles at the 200K-token-per-call threshold. If you're building agents that process large documents or hold long conversation histories, the 500K-token window is the relevant capability to test.

The release also includes Grok Voice Transcribe 2.0. It adds speaker diarization and word-level timestamps. Multichannel input support and improved turn detection are also new. If you're building applications that need voice transcription alongside long-context text processing, both are live on the xAI platform as of this week.

---