AI briefing: 18 September 2026
OpenAI targets legal work, Anthropic discloses AI-led R&D metrics, Figure advances humanoid generalisation, and Qwen ships an omni-modal model.
AI-generated editorial illustration.
Today’s strongest developments are concentrated in professional AI deployments, frontier-lab transparency, physical AI and multimodal models. This edition focuses on announcements made on 17–18 September 2026 and excludes items substantially covered in the previous briefing.
1. OpenAI launches Astra for Law for selected US firms
OpenAI has introduced Astra for Law, a GPT-6 Astra configuration for legal research and drafting. It combines custom legal instructions with a daily updated index spanning US case law, statutes, regulations, court rules and administrative decisions. Initial access is limited to selected US law firms through Trusted Access and Codex, with API availability planned later.
Why it matters: This is a move from general-purpose models towards controlled, sector-specific systems where retrieval, permissions and auditability matter as much as raw model capability. Legal users still need to verify cited authority before relying on outputs.
Sources
- Astra for Law — OpenAI Help Center
- Introducing Astra for Law — OpenAI — 2026-09-17
2. Anthropic publishes proposed metrics for AI-led AI research
Anthropic has proposed three measures intended to expose the pace of frontier development: the share of AI R&D performed by AI, agent oversight, and compute allocation. Its snapshot says Claude ‘leads’ 26% of its model R&D work, while more than 90% is performed with Claude collaborating under human direction. Anthropic says no measured work is fully autonomous.
Why it matters: The disclosure offers a more operational indicator of recursive-improvement risk than model benchmarks alone. It also places pressure on other frontier labs to make comparable, independently verifiable claims about automated research and safety investment.
Sources
- Measurements for understanding the pace of AI development inside frontier labs — Anthropic — 2026-09-17
- Anthropic says its model Claude is helping to build the next version of itself — AP — 2026-09-18
3. Anthropic introduces a Life Sciences Verification Program
Anthropic has announced a Life Sciences Verification Program, adding a specialised access and verification initiative for life-sciences users. The company listed the programme alongside its broader frontier-safeguards work on 17 September.
Why it matters: Biology is among the most consequential domains for advanced AI. Purpose-built identity, access and oversight mechanisms may become a template for deploying powerful models in other high-risk professional settings.
Sources
- Introducing the Life Sciences Verification Program — Anthropic — 2026-09-17
4. Figure reports zero-shot whole-body humanoid tests across 30 homes
Figure unveiled Helix 2.5, which it says performed tidying, towel-folding and bed-making tasks in 30 previously unseen homes without environment-specific data collection or adaptation. Figure reports that pretraining on its Index dataset raised zero-shot task success from 9% to 56% in its comparison.
Why it matters: The result, if reproduced more broadly, suggests that robotics may increasingly follow the foundation-model playbook: broad pretraining on human behaviour followed by relatively small amounts of task-specific data. The company’s figures remain vendor-reported.
Sources
- Helix 2.5: Zero-Shot 30-Home Generalization — Figure — 2026-09-17
5. Qwen makes its Qwen3.8-Omni-Flash multimodal model available
Qwen has published Qwen3.8-Omni-Flash, a native multimodal model that accepts text, images, audio and video and returns text. The model supports up to a 1 million-token context window and is positioned for agentic productivity, coding and audio-video understanding workflows.
Why it matters: Long-context multimodal models are becoming practical building blocks for agents that must interpret mixed media rather than just text. Qwen’s OpenAI-compatible API support also lowers switching costs for developers already using that ecosystem.
Sources
- Qwen3.8-Omni-Flash — QwenCloud
- Visual understanding models — QwenCloud
6. ChatGPT for Word rolls out to all ChatGPT plans
OpenAI’s release notes say ChatGPT is now available in Microsoft Word, including on the Free plan. The add-in can draft from notes, summarise documents, revise selected text and adjust headings and formatting; use counts against applicable plan allowances.
Why it matters: Putting ChatGPT inside Word brings generative AI directly into a dominant knowledge-work surface. Broad free-tier availability could accelerate adoption, while increasing the need for organisations to clarify document-handling and review policies.
Sources
- ChatGPT release notes — OpenAI Help Center — 2026-09-17
7. European AI firms challenge US-led calls to slow frontier development
Reuters reports that Mistral and other European AI companies have criticised US rivals’ calls to slow advanced-model progress, arguing that safety-centred restrictions could entrench established incumbents. The dispute follows a recent public push by several US frontier-lab leaders for coordinated pacing.
Why it matters: The disagreement sharpens a core policy fault line: safeguards that limit frontier development may reduce risks, but can also raise barriers for smaller competitors and regions seeking strategic autonomy.
Sources
- Europe's AI firms, playing catch-up, challenge US calls for slowdown — Reuters — 2026-09-18
What to watch
Watch for details on OpenAI’s API timetable for Astra for Law, whether other frontier labs adopt Anthropic-style R&D transparency measures, independent validation of Figure’s Helix 2.5 results, and policy responses to the growing US–European split over coordinated AI slowing.
Researched and generated with AI. Explore the linked sources for original reporting and context.
Back to all news ↗