# Agent and Healthcare Model Gains Meet Long-Horizon Reality

*By AI High Signal Digest • July 14, 2026*

GPT-5.6 and Muse Spark 1.1 set new markers for agent and healthcare performance, while fresh benchmarks expose persistent weaknesses in long-horizon execution and safe autonomy. The brief also covers efficient long-context research, video and agent product releases, Oak Lab’s launch, and emerging U.S. open-model policy discussions.

## Top Stories

*Why it matters: leading models are improving on agent and medical benchmarks, but new evaluations still show large gaps in sustained execution and safe autonomy.*

- **OpenAI’s GPT-5.6 is expanding across products and scoring near the top of agent evaluations.** The Sol, Terra, and Luna family is rolling out in ChatGPT, Codex, and the API, and is now generally available through Amazon Bedrock. GPT-5.6 Sol ranked **#2** on Agent Arena from 7,800 real-world agent sessions, with #1 steerability and #2 confirmed task success; the benchmark tests long-horizon workflows with web, filesystem, and terminal access. [^1][^2][^3][^4][^3]

- **Muse Spark 1.1 posted strong healthcare results, while radiology tests retain a human-performance gap.** On HealthBench Professional’s 525 clinician tasks, it had a higher overall score than GPT-5.6 Sol and was statistically on par on a length-adjusted score, at $1.25/$4.25 per million input/output tokens versus $5/$30 for Sol. On RadLE 2.0, it outperformed GPT-5.6 Sol and Gemini 3.1, and led the handover-readiness index at 48.5 against a 52.0 human-expert baseline—but the benchmark reports that no model reached average human-expert performance overall. [^5][^6][^7][^8]

- **Long-horizon agency remains unresolved.** Long-Horizon Terminal-Bench evaluated 18 frontier models on 46 reproducible terminal tasks requiring up to 90 minutes and 120–320 steps. The best mean reward was 0.505; no model solved a third of tasks, and 29 tasks remained unsolved by every model. Separately, the persistent-enterprise simulation Morpheus concluded that tested frontier LLMs are not continual learners in dynamic business settings. [^9][^10]

## Research & Innovation

*Why it matters: progress is moving beyond larger models toward systems that use context, computation, and physical coordination more efficiently.*

- **DeepSeek-V4 is presented as a full-stack redesign for native 1M-token context.** The reported architecture combines compressed sparse and heavily compressed attention, multiple residual streams, fused mixture-of-experts kernels, and on-policy distillation. At 1M tokens, the report estimates V4-Pro uses about 27% of DeepSeek-V3.2’s single-token inference FLOPs and 10% of its KV cache; V4-Flash targets roughly 10% and 7%, respectively. [^11]

- **Sakana AI’s Smart Cellular Bricks demonstrate decentralized physical intelligence.** Published in *Nature Communications*, the work uses identical neural-network-equipped cubes that communicate locally to infer a shared shape, identify missing modules, and guide repair. Tests across nearly 200 physical bricks reported 100% convergence; the system tolerated up to 15% module failure and located damage with 95% accuracy. [^12][^13]

- **Hugging Face and vLLM removed a common inference bottleneck for open models.** Transformers implementations can now run through vLLM at native speed, avoiding separate research and production implementations; reported benchmarks matched or exceeded native vLLM throughput from 4B to 235B parameters, including tensor-parallel and MoE setups. [^14]

## Products & Launches

*Why it matters: video generation and software agents are being delivered through lower-cost, more accessible workflows.*

- **Google’s Gemini Omni Flash debuted at #1 in Artificial Analysis’s text-to-video and image-to-video leaderboards.** The natively multimodal model accepts text, images, and video; produces 3–10 second 720p/24fps clips with native audio; and supports conversational editing. It costs $0.10 per generated second and is available through Gemini API, AI Studio, the Gemini app, and free in YouTube Shorts and Create. [^15]

- **Devin Fusion entered agent preview with Fable 5.** Cognition reports lower cost per task than Opus 4.8 through better delegation and reasoning chains, while noting that savings are not uniform—serial debugging still needs accumulated context. [^16][^17][^18]

- **ChatGPT Sites entered public beta.** Paid-plan users can turn prompts, files, or rough ideas into dashboards, reports, prototypes, and lightweight apps, then build in ChatGPT Work or Codex, preview privately, and publish via URL. [^19]

## Industry Moves

*Why it matters: continual learning and open-weight deployment are becoming strategic fronts alongside frontier-model development.*

- **Richard Sutton and Khurram Javed left Keen Technologies to found Oak Lab.** Their stated approach centers reinforcement learning and intelligence maintained through run-time experience; they argue current deep-learning methods require fundamental reworking. Oak Lab says it will first demonstrate limitations in simple settings before pursuing domain-independent algorithms and larger-scale systems. [^20][^21]

- **Open-weight usage is rising alongside closed models.** One gateway reported open-weight models accounting for 29% of tokens, up from 11% in April; Baseten says companies are increasingly deploying open-source AI alongside closed providers. [^22][^23]

## Policy & Regulation

*Why it matters: U.S. policy discussions may increasingly tie open-model treatment to international capability comparisons.*

- **The Trump administration and AI industry are reportedly discussing a capability framework for U.S. open-source models benchmarked against leading Chinese open-source models.** According to the report, a proposal would streamline U.S. open and licensed models to market when their capabilities match or fall below Chinese open-model capabilities; the same report raises concerns over possible malicious software or exploitable back doors in Chinese models. [^24]

## Quick Takes

*Why it matters: capability gains are arriving alongside practical improvements in collaboration, access, and deployment.*

- GPT-5.6 Sol Ultra was reported to have produced a short construction for Erdős problem #793 on 2-primitive sets. [^25]
- Claude Artifacts now supports public sharing, multiplayer editing, and creation through Claude Tag on Team and Enterprise plans. [^26][^27][^28][^27]
- Step 3.7 Flash joined Baseten’s library with 198B total parameters, 11B active parameters, native image/video input, and a 256K context window. [^29]
- OpenAI reported 7M active users across Codex and ChatGPT Work. [^30]

---

### Sources

[^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2075271421149020426)
[^2]: [𝕏 post by @AWSNewsroom](https://x.com/AWSNewsroom/status/2076776261397909647)
[^3]: [𝕏 post by @arena](https://x.com/arena/status/2076709326711037991)
[^4]: [𝕏 post by @arena](https://x.com/arena/status/2076709329173127216)
[^5]: [𝕏 post by @MedicalSphereAI](https://x.com/MedicalSphereAI/status/2076776649807573075)
[^6]: [𝕏 post by @_jasonwei](https://x.com/_jasonwei/status/2076723921441988758)
[^7]: [𝕏 post by @DrDatta_AIIMS](https://x.com/DrDatta_AIIMS/status/2076689436092326276)
[^8]: [𝕏 post by @DrDatta_AIIMS](https://x.com/DrDatta_AIIMS/status/2076689408040849507)
[^9]: [𝕏 post by @Yucheng__Shi](https://x.com/Yucheng__Shi/status/2076548493913817406)
[^10]: [𝕏 post by @skyfallai](https://x.com/skyfallai/status/2076713589788864920)
[^11]: [𝕏 post by @ZhihuFrontier](https://x.com/ZhihuFrontier/status/2076653008616825251)
[^12]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2076597965804765283)
[^13]: [𝕏 post by @hardmaru](https://x.com/hardmaru/status/2076633306008072592)
[^14]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2076763231788339669)
[^15]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2076747075036045645)
[^16]: [𝕏 post by @cognition](https://x.com/cognition/status/2076714965344342382)
[^17]: [𝕏 post by @cognition](https://x.com/cognition/status/2076714967118594247)
[^18]: [𝕏 post by @cognition](https://x.com/cognition/status/2076714968334913786)
[^19]: [𝕏 post by @jxnlco](https://x.com/jxnlco/status/2076511474600857950)
[^20]: [𝕏 post by @RichardSSutton](https://x.com/RichardSSutton/status/2076663628301058329)
[^21]: [𝕏 post by @kjaved_](https://x.com/kjaved_/status/2076663868160459214)
[^22]: [𝕏 post by @rauchg](https://x.com/rauchg/status/2076713720731042174)
[^23]: [𝕏 post by @baseten](https://x.com/baseten/status/2076831791806390455)
[^24]: [𝕏 post by @pequityresearch](https://x.com/pequityresearch/status/2076748647568449802)
[^25]: [𝕏 post by @prz_chojecki](https://x.com/prz_chojecki/status/2076749164067565872)
[^26]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2076789349145092230)
[^27]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2076789351745458624)
[^28]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2076789352869613872)
[^29]: [𝕏 post by @baseten](https://x.com/baseten/status/2076862431385850034)
[^30]: [𝕏 post by @thsottiaux](https://x.com/thsottiaux/status/2076735790567338203)