# Health AI Expands, Open Models Close Gaps, and the Grid Becomes an AI Issue

*By AI News Digest • June 19, 2026*

Today’s biggest signals came from healthcare and biology: OpenAI paired a broad health upgrade with published rare-disease results, Profluent signed a $2.25B Lilly deal, and Midjourney surfaced a medical imaging project. Elsewhere, new benchmark data showed open-weight momentum amid persistent agent limits, while labs and policymakers focused on deeper safety and infrastructure questions.

## Health and biology led the day

### OpenAI paired a broad health rollout with published clinical evidence

OpenAI said GPT-5.5 Instant is now on par with its frontier Thinking models for health-related questions, with better urgent-care recognition, context gathering, uncertainty explanation, and clarity across more than 230 million weekly health and wellness queries; the update is available to all free ChatGPT users and was shaped with feedback from hundreds of physicians across 60 countries, 49 languages, and 26 specialties [^1][^2]. Separately, OpenAI, Boston Children’s Hospital, and Harvard published a study in *NEJM AI* showing o3 Deep Research helped clinicians identify 18 diagnoses across 376 previously unsolved rare pediatric disease cases, with every result going through human adjudication and clinical confirmation [^3][^4][^5].

*Why it matters:* one announcement widened access to health guidance inside ChatGPT, while the other tested AI inside an expert-led rare-disease reanalysis workflow that had already resisted years of specialist review [^6].

### Profluent signed a $2.25B Lilly deal for AI-designed gene editors

Profluent said it signed a $2.25 billion milestone deal with Eli Lilly to develop AI-designed gene editors for therapeutic large-gene insertion, framing the work as an example of AI unlocking a problem that could not previously be solved in this way [^7]. The company says its transformer-based sequence models are trained on more than 100 billion protein sequences and used to generate proteins from scratch; it also pointed to OpenCRISPR as the first demonstration of AI-generated functional gene editors, and said peer-reviewed comparisons found sequence models outperforming structure-based approaches on complex multi-domain proteins [^7].

*Why it matters:* this is a large commercial signal for generative biology, and it ties frontier-model methods directly to therapeutic gene-editing programs rather than discovery tooling alone [^7].

### Midjourney surfaced a new medical imaging project with clear tradeoffs

Midjourney published a technical dive on a new "Scanner" project, which François Chollet described as a hardware effort for full-body internal 3D scans without MRI [^8][^9]. A separate technical summary described the system as radiation-free, magnet-free, fast, and low-cost, while also noting current constraints: it requires a water immersion tank and its resolution is still coarser than CT or MRI [^10].

*Why it matters:* it is a notable expansion from an AI image company into medical hardware, but the present limitations are substantial and part of the story [^10].

## Open-weight competition kept getting stronger

### A new benchmark showed both momentum and stubborn limits

Artificial Analysis launched AA-Briefcase, a benchmark for long-horizon knowledge work across multi-week projects with thousands of fragmented inputs, including 25,000+ Slack messages and 3,500+ emails [^11]. Its headline result was sobering: the top model, Claude Fable 5, satisfied all rubric criteria on just 3% of tasks, and no model scored above 50% on 31 of 91 tasks; within that field, GLM-5.2 was the next-best non-Anthropic model at 1266 Elo and one of the strongest price/performance options, at $2.40 per task versus $31 for Claude Fable 5 [^11]. Poolside added to the open-weight push by releasing Apache 2.0 weights for its 256K-context Laguna M.1 and saying that "open weights are now our default" [^12][^13].

*Why it matters:* open-weight models are getting more competitive on cost and capability, but the benchmark also underscores how far the field still is from reliable end-to-end agentic knowledge work [^11].

## Safety work is moving below the interface layer

### OpenAI and DeepMind both argued for more structural approaches

> "Instead of assuming AI will always do what we intend, we ask: what if it doesn’t?" [^14]

OpenAI said its new work on broadly beneficial reinforcement learning used realistic conversations across 12 domains and improved a compute-matched model on 44 of 53 independent evaluations spanning deception, reward hacking, safety, health, and mental health; it also reported cross-domain transfer, where training only on health conversations improved non-health misalignment evaluations [^15][^16][^17]. The company also reported that the trained model was harder to steer toward harmful behavior with adversarial prompts and showed preliminary resistance to harmful fine-tuning while remaining responsive to helpful instructions [^18]. In parallel, Google DeepMind published an AI Control Roadmap arguing that most agent failures come from misinterpreting commands or becoming over-enthusiastic, and that there is a narrow window to embed structural security protocols before multi-agent systems scale globally [^19][^20].

*Why it matters:* both efforts point toward safety techniques that try to shape persistent behavior and system design, rather than relying only on after-the-fact prompt guardrails [^16][^18][^20].

## AI infrastructure is becoming energy policy

### FERC took a meaningful step on large-load interconnection

FERC issued a large-load interconnection milestone that affects how AI factories, semiconductor fabrication support systems, and advanced manufacturing facilities connect to the grid [^21]. The policy direction highlighted in the announcement includes large-load customers funding their own network upgrades, bringing new generation online, and offering flexible load; customers that can demonstrate flexibility may qualify for accelerated study timelines as short as 60 days [^21]. NVIDIA also said it and Emerald AI are already working on flexible AI factories designed as grid assets, with commercial deployment beginning later this year [^21].

*Why it matters:* AI capacity planning is no longer just a chip and data-center story; grid access and load flexibility are becoming part of the competitive stack too [^21].

---

### Sources

[^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067672740539306261)
[^2]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067672742426775728)
[^3]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067625110199247353)
[^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067625111717609504)
[^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067625113193951611)
[^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067625115182120972)
[^7]: [Ali Madani \(Profluent\) on the $2.25B Eli Lilly deal and the "GPT-1.5 era" of biology](https://www.youtube.com/watch?v=Oes9W8XOELk)
[^8]: [𝕏 post by @midjourney](https://x.com/midjourney/status/2067422898407837797)
[^9]: [𝕏 post by @fchollet](https://x.com/fchollet/status/2067589665830375682)
[^10]: [𝕏 post by @iScienceLuvr](https://x.com/iScienceLuvr/status/2067384062910832812)
[^11]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2067744637155226101)
[^12]: [𝕏 post by @poolsideai](https://x.com/poolsideai/status/2067623353230217448)
[^13]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2067690103451918721)
[^14]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2067594863785173257)
[^15]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067722688165232654)
[^16]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067722691675824637)
[^17]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067722693714338044)
[^18]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2067722695270334549)
[^19]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2067594866196877631)
[^20]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2067594868180857165)
[^21]: [How FERC’s Large-Load Interconnection Actions Help Address Grid Stress, Improve Affordability](https://blogs.nvidia.com/blog/ferc-large-load-interconnection)