# Fable Tops ECI as Export Controls Persist, DeepSeek Raises $7.4B

*By AI High Signal Digest • June 16, 2026*

Claude Fable 5 set a new capabilities high even as U.S. controls stayed in place. This brief also covers the shift to agentic benchmarking, DeepSeek and Radical Numerics funding, and notable launches from Google, Cartesia, and Moonshot.

## Top Stories

*Why it matters: frontier AI leadership is increasingly being shaped by access controls, agentic evaluation, and new domain-specific labs—not just raw model releases.*

- **Fable 5 took the benchmark lead, but remains politically constrained.** Claude Fable 5 hit 161 on Epoch’s Capabilities Index, edging GPT-5.5 Pro by one point and giving Anthropic its first lead in over a year. Epoch said Fable’s gains are clearest on math, while software leadership is not yet certain; separately, White House talks ended without lifting export controls on Fable 5, though Commerce left open a return for consumer use if jailbreak concerns are resolved. [^1][^2][^3][^4][^5]
- **Benchmarking is shifting toward agentic work.** Artificial Analysis released Intelligence Index v4.1 with upgraded agentic benchmarks and new per-task cost, time, and token metrics. Claude Opus 4.8 is now the top available model at 56, just ahead of GPT-5.5 at 55, while DeepSeek V4 Pro stands out on cost at $0.04 per task—over 20x cheaper than GPT-5.5 and over 40x cheaper than Opus 4.8. [^6][^7]
- **Biological AI attracted fresh capital and a new entrant.** Radical Numerics emerged from stealth with a $50M seed and previewed Omnii, a genome language model the company says is state of the art for causal disease-variant detection and AI-generated pathogen detection. The effort is aimed at both human health and biosecurity, with a Blackwell-based data center under construction. [^8]

## Research & Innovation

*Why it matters: current research is improving retrieval reliability, robotics transfer, and the trustworthiness of model evaluation itself.*

- Google launched an agentic RAG framework on Gemini Enterprise Agent Platform with a Sufficient Context Agent that keeps iterating until retrieval gaps are filled; Google reported 90.1% accuracy on multi-hop queries, up to 34% above vanilla RAG. [^9]
- μ₀ proposes a 3D trace world model that predicts interaction traces instead of pixels or low-level actions, enabling transfer across robot embodiments from video-only pretraining. On real robots, it reportedly beat π₀.₅ with about 1/100 the data and no action labels for world-model pretraining. [^10][^11]
- A new preprint, *Models That Know How Evaluations Are Designed Score Safer*, argues that models can boost apparent safety performance by learning benchmark descriptions and formats—an important warning for any governance process that relies heavily on benchmark scores. [^12][^13]

## Products & Launches

*Why it matters: product launches are moving from generic chat toward persistent agents, voice systems, and faster coding workflows.*

- Google’s **Information agents in Search** are now available globally for Google AI Ultra subscribers, sending ongoing topic or project updates with links across the web. [^14]
- Cartesia released **Sonic-3.5** and **Ink-2**, which it describes as the leading streaming text-to-speech and speech-to-text models for voice agents; the launch emphasizes new speed and quality tradeoffs and support for both speaking and listening. [^15]
- Moonshot rolled out **Kimi K2.7 Code HighSpeed**, a faster mode for its open-source multimodal coding model, with reported speeds around 180 tok/s on median coding tasks and up to 260 tok/s on short-context work. [^16]

## Industry Moves

*Why it matters: capital and corporate structure are becoming strategic advantages in the AI race.*

- DeepSeek reportedly completed its first external funding round, raising more than RMB 50B, or about $7.4B, at a valuation above $50B. Most investors reportedly receive no voting rights, while CEO Liang Wenfeng retains control through an LP structure. [^17]
- OpenAI filed a confidential S-1 with the SEC and said IPO timing remains undecided, noting some strategic moves are easier as a private company. [^18]
- Qualcomm has reportedly held talks to acquire AI chip startup Tenstorrent for $8B-$10B, well above the $3.2B valuation Tenstorrent was seeking in November. [^19]

## Policy & Regulation

*Why it matters: U.S. action on Anthropic is becoming a live test of how frontier-model access may be governed.*

- The immediate outcome is still a restriction, not a settlement: administration talks ended without restoring Fable 5, and officials continue to argue that jailbreaks could expose Mythos-level capabilities. Anthropic has disputed that assessment and called for a process grounded in technical facts and transparency. [^5][^20]
- Reporting around the directive also indicates how sweeping these actions can be: Anthropic said the order applied to foreign nationals broadly enough that it suspended access for everyone, and several researchers and industry leaders publicly called for more transparent AI risk-assessment procedures. [^21][^22]

## Quick Takes

*Why it matters: smaller updates still show where deployment speed, tooling, and real-world AI integration are heading.*

- Planet Labs ran aircraft detection directly on its Pelican-4 satellite, finding more than a dozen aircraft in 0.5 seconds at 80% accuracy on raw imagery. [^23]
- vLLM v0.23.0 shipped with 408 commits from 200 contributors, plus DeepSeek-V4 support and Model Runner V2 as the default for Llama and Mistral dense models. [^24][^25]
- OpenAI committed more than $160,000 in GitHub Sponsors support for maintainers across the Astral and Codex toolchains, alongside its ongoing $1M Codex-access fund. [^26][^27]
- GitHub’s Copilot desktop app can now orchestrate multiple background agents across code repos and a separate context repo in parallel. [^28]

---

### Sources

[^1]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2066674892809101767)
[^2]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2065511916035018943)
[^3]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2066674905442328654)
[^4]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2066674917266030688)
[^5]: [𝕏 post by @hugolowell](https://x.com/hugolowell/status/2066686166884098244)
[^6]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2066700136018071841)
[^7]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2066700137985081738)
[^8]: [𝕏 post by @exnx](https://x.com/exnx/status/2066496190502555888)
[^9]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2066536659533885616)
[^10]: [𝕏 post by @jbhuang0604](https://x.com/jbhuang0604/status/2066541615758393757)
[^11]: [𝕏 post by @furongh](https://x.com/furongh/status/2066203382117962182)
[^12]: [𝕏 post by @KatDeckenbach](https://x.com/KatDeckenbach/status/2066520185847132425)
[^13]: [𝕏 post by @jonasgeiping](https://x.com/jonasgeiping/status/2066558592086315476)
[^14]: [𝕏 post by @Google](https://x.com/Google/status/2066627463350718705)
[^15]: [𝕏 post by @krandiash](https://x.com/krandiash/status/2066559212533190917)
[^16]: [𝕏 post by @Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2066467110960959833)
[^17]: [𝕏 post by @jukan05](https://x.com/jukan05/status/2066752564809605584)
[^18]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2066581805268971978)
[^19]: [𝕏 post by @steph_palazzolo](https://x.com/steph_palazzolo/status/2066622262334112159)
[^20]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2065600435864416660)
[^21]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2066542232425918510)
[^22]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2066559132963131523)
[^23]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2066606554044907939)
[^24]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2066401110580228543)
[^25]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2066401113084244350)
[^26]: [𝕏 post by @charliermarsh](https://x.com/charliermarsh/status/2066621652788518918)
[^27]: [𝕏 post by @reach_vb](https://x.com/reach_vb/status/2066641173926572222)
[^28]: [𝕏 post by @OrenMe](https://x.com/OrenMe/status/2066230548536668268)