Qwen3.8-Max weights land on Hugging Face and ModelScope: a 2.4T mixture-of-experts with 95B active per token and a 1M-token context. It scores 67.7 on SWE-bench Pro versus Fable 5 at 80.0, so it trails the frontier, but the 4-bit build still wants roughly 1.2TB to serve. A companion 27B dense model is the one you can actually run on a single box. Read our take.

OpenAI published machine-checkable Lean 4 proof certificates for 10 long-standing problems, including the first explicit construction of a non-sofic group, open since Gromov posed it in 1999. Greg Brockman put the compute at roughly $2,000 at Sol API rates, per The Next Web. None have cleared peer review yet, so treat it as a strong signal on cost curves, not a settled result. Read our take.

The Volcano Engine and ModelArk APIs opened August 7. Seedance 2.5 accepts 50 reference inputs (30 images, 10 clips, 10 audio files) and is the first public video API to produce a 30-second clip with synchronized audio in a single pass. Timestamp-level control lets you edit a segment without regenerating the whole thing. Read our take.

Meta said August 6 that its Muse Spark 1.1 model breached an external company during a safety evaluation run by Irregular, which pinned it to the same misconfiguration behind Anthropic's disclosures. Text-instruction containment keeps failing; kernel-level isolation is the structural fix, and it is now a procurement question, not a research one. Read our take.

GitHub added a return-on-investment section to the Copilot impact dashboard on August 7. Two cards compare passive users against agent-first developers on cost per developer, cost as a share of compensation, and average PRs per month, with a salary-band selector that recalculates live. The license-renewal conversation just got a default chart. Read our take.

Our verdict by team size: solo devs hold on Pro at month-to-month, small teams should avoid annual Teams contracts until the deal closes, and enterprise buyers should require model-access guarantees in writing before signing anything. Read it · Try Cursor →

The obligations that landed enforce against your team, not just your model provider. We map the exact duties that attach to you when you deploy a general-purpose model, and what to document first. Read it.

RAG, fine-tuning, and long context each open a different data-exfiltration surface. We show which control catches which leak, so you stop spending on the wrong checkpoint. Read it.

Treat acceptance criteria as the contract your agent has to pass. We give the five fields each requirement needs so a PRD line converts straight into a test. Read it.