Artificial Intelligence

OpenAI has released Astra for Law, pairing GPT-6 Astra with a 230-million-URL Legal Search Index and CourtListener integration, initially available only to selected U.S. law firms under a Trusted Access program. The company reports a 54.0% correctness rate on a private 200-question legal research benchmark, though the result is self-reported and has not been independently reproduced. Key details including the full list of participating firms, complete index licensing terms, and Trusted Access program rules have not been publicly disclosed.

Anthropic has published a self-reported, sample-based index showing its Claude AI reached an automation level classified as 'leads' on 26% of the company's model research and development work in August 2026, up from under 1% in February 2026. The measurement covers a frozen basket of 378 discrete tasks drawn from a July 2026 employee sample, weighted by person-time, and has not been independently audited. Anthropic explicitly states Claude is not operating autonomously and that the figure does not indicate human replacement, while acknowledging that judge-model dependence and a static task basket limit what the number can show over time.

PrismML's Ternary Bonsai 2 27B compresses a 27.36-billion-parameter model to 5.93 GB at 1.76 bits per weight, with the company reporting 98.2% benchmark retention against its Qwen3.8-27B FP16 baseline across 20 tasks — figures that have not been independently reproduced. The model also requires PrismML's custom llama.cpp fork rather than standard tooling, a deployment constraint that stock runtimes cannot satisfy. Early community testing confirmed the model runs on consumer hardware but surfaced memory overruns, hardware compatibility gaps, and task-quality failures not captured in the vendor's aggregate score.

Emerald AI, Google, and NVIDIA founded the AI Energy Management Alliance on September 16, 2026, alongside 18 partners including Anthropic, National Grid, and Constellation, with the stated goal of making AI data centers flexible grid resources. The alliance's headline figure — that power-flexible AI factories could unlock 100 GW of U.S. grid capacity — is an Emerald AI company projection, not an independently validated engineering finding, and public AEMA materials establish no binding utility demand-response commitments. Demonstrated results are narrower: a Phoenix pilot reduced GPU-cluster power by 25% for three hours, and Emerald AI reports additional multi-site demonstrations, but none of these validate the full 100 GW estimate.