# Fable 5’s Real-Work Gain, Google’s Gemini Push, and New Infrastructure Bets *By AI High Signal Digest • July 2, 2026* Anthropic’s Fable 5 returned with stronger real-work benchmark results and tighter safeguards, Google rolled out a broad Gemini/Gemma product wave, and the compute race intensified through major funding and cloud strategy moves. The brief also covers faster generation research, calibration advances, new agent tooling, and key governance updates. ## Top Stories *Why it matters: the biggest updates combined stronger real-world capability, broader multimodal rollout, and a sharper focus on compute economics.* - **Anthropic’s Fable 5 returned with both stronger real-work results and tighter controls.** On the Remote Labor Index, Fable 5 completed 16.1% of 240 real remote-work projects at a professional standard, up from Opus 4.6’s 4.2% and roughly double the next model. Anthropic also redeployed it globally with new cyber classifiers that it says block the reported technique in over 99% of cases, with blocked requests routed to Opus 4.8. [^1][^2][^3][^4] - **Google shipped a broad Gemini/Gemma release wave.** The company introduced Nano Banana 2 Lite for image generation, Gemma 4 12B for on-device use, Gemini Omni Flash APIs for custom video workflows, Gemini 3.5 Live Translate across 70+ languages, and NotebookLM upgrades for reasoning, code execution, and document generation. Gemini Spark also entered beta for U.S. Google AI Ultra subscribers with MCP support and app integrations. [^5][^6][^7] - **The infrastructure race kept accelerating.** Together Compute raised an $800 million Series C at an $8.3 billion valuation, while a separate cited update said it is serving 400T tokens per month as demand for open models rises. Bloomberg also reports Meta is planning to sell access to excess AI compute and hosted models from its infrastructure. [^8][^9][^10] ## Research & Innovation *Why it matters: researchers are pushing on the three bottlenecks that now matter most—speed, reliability, and world modeling.* - **NVIDIA’s TwoTower points to a cheaper speed path than full retraining.** The method repurposes a pretrained 30B model into a two-part diffusion language model where one copy holds context and the other writes token chunks in parallel, preserving 98.7% of original quality at 2.42× faster generation with only ~8% of the original training data. [^11][^12] - **RLMF targets a persistent LLM weakness: confidence calibration.** The approach uses a model’s own self-judgments as a training signal, first calibrating faithful confidence estimates and then editing outputs into natural uncertainty language; the reported result is state-of-the-art faithful calibration while surpassing standard RL by up to 63%. [^13] - **Neural Theorizer (NEO) pushes world models toward explicit reasoning.** The system learns compositional theories from raw observation without language or LLM supervision, aiming to discover reusable primitives rather than only predict pixels; it was selected for an ICML 2026 oral presentation. [^14][^15][^14] ## Products & Launches *Why it matters: the newest tools are packaging strong models into workflows teams can act on immediately.* - **Devin Security Swarm turns agentic coding into security ops.** Cognition says the new Agentic MapReduce system scans whole codebases, validates exploitability in sandboxes, and can ship remediation PRs; on a 50-vulnerability GHSA set across 14 languages, it found 36 issues at 30% lower cost per finding than the next most accurate alternative. [^16][^17][^18][^19] - **Notion added an HTML block for AI-generated interactive outputs.** Teams can ask AI to turn content into interactive explainers, prototypes, or diagrams directly inside a page, reducing the gap between draft output and something collaborators can test. [^20] - **VS Code expanded its agent workflow surface.** The July release adds chat banners for failing CI checks and review feedback, better multi-session management in the Agents window, and sandboxed terminal commands on macOS and Linux. [^21][^22][^23] ## Industry Moves *Why it matters: companies are competing on infrastructure, operational efficiency, and distribution as much as on raw model quality.* - **Odyssey raised $310 million at a $1.45 billion valuation.** The Palo Alto lab, backed by Amazon and AMD Ventures, is building world models for interactive real-time simulations rather than fixed video generation. [^24] - **Shopify showed how much margin can come from model operations, not just model choice.** Its Model Optimization Flywheel converts product expertise into evals and repaired training data; in one GraphQL agent example, annualized serving cost fell from $27 million to $1 million after 4× prompt compression while still beating frontier models on quality. [^25] - **Runway expanded enterprise distribution through Bertelsmann.** Its tools will be integrated across Bertelsmann businesses including RTL Group, BMG, and Bertelsmann Marketing Services. [^26] ## Policy & Regulation *Why it matters: frontier deployment is increasingly being shaped by formal testing and public-governance frameworks, not just model releases.* - **Anthropic is building a more formal safety coordination layer.** Alongside Fable 5’s redeployment, it says it is drafting a jailbreak-severity framework with Amazon, Microsoft, Google, and other partners, and expanding U.S. government collaboration on pre-release testing and safeguards. [^3] - **The UN’s independent science panel released a preliminary AI report.** The report is positioned as an evidence-based assessment of AI’s current state and argues that benefits and harms will depend on government choices. [^27][^28] ## Quick Takes *Why it matters: these smaller updates still show where evaluation, media generation, and local deployment are heading.* - Claude Sonnet 5 scored 1391 Elo on AA-Briefcase, second behind Fable 5, but max effort averaged 183 turns per task. [^29] - Reve 2.0 debuted at #2 on Artificial Analysis’s text-to-image leaderboard and uses structured layout prompts for easier editing. [^30] - Fish Audio S2.1 Pro launched with 83-language TTS, voice cloning, and 56.3 characters/second generation; API access is free through July 24. [^31] - Qwen3.6-27B-NVFP4 arrived on Hugging Face, optimized for Blackwell GPUs and cutting local-memory requirements by about 2.5×. [^32][^33] --- ### Sources [^1]: [𝕏 post by @CAIS](https://x.com/CAIS/status/2072360965522489789) [^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2072376968729817531) [^3]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2072163884430229756) [^4]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2072199399833240054) [^5]: [𝕏 post by @Google](https://x.com/Google/status/2072385546551247013) [^6]: [𝕏 post by @Google](https://x.com/Google/status/2072409252417708218) [^7]: [𝕏 post by @Google](https://x.com/Google/status/2072409250312425824) [^8]: [𝕏 post by @vipulved](https://x.com/vipulved/status/2072321276094673083) [^9]: [𝕏 post by @tri_dao](https://x.com/tri_dao/status/2072429334758121556) [^10]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2072309840823689247) [^11]: [𝕏 post by @NVIDIAAI](https://x.com/NVIDIAAI/status/2072394812301480067) [^12]: [𝕏 post by @LiorOnAI](https://x.com/LiorOnAI/status/2072402904867365167) [^13]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2072470330535473485) [^14]: [𝕏 post by @SungjinAhn_](https://x.com/SungjinAhn_/status/2072230094614237320) [^15]: [𝕏 post by @LiorOnAI](https://x.com/LiorOnAI/status/2072380547603829224) [^16]: [𝕏 post by @cognition](https://x.com/cognition/status/2072368177388937239) [^17]: [𝕏 post by @cognition](https://x.com/cognition/status/2072368172271890629) [^18]: [𝕏 post by @cognition](https://x.com/cognition/status/2072368170157904069) [^19]: [𝕏 post by @cognition](https://x.com/cognition/status/2072368174968828330) [^20]: [𝕏 post by @NotionHQ](https://x.com/NotionHQ/status/2072354853821755822) [^21]: [𝕏 post by @code](https://x.com/code/status/2072341823255715986) [^22]: [𝕏 post by @code](https://x.com/code/status/2072341822068727899) [^23]: [𝕏 post by @code](https://x.com/code/status/2072341824308461993) [^24]: [𝕏 post by @TheRundownAI](https://x.com/TheRundownAI/status/2072334737557893563) [^25]: [𝕏 post by @ShopifyEng](https://x.com/ShopifyEng/status/2072405411756724677) [^26]: [𝕏 post by @c_valenzuelab](https://x.com/c_valenzuelab/status/2072328321363313040) [^27]: [𝕏 post by @ODET_UN](https://x.com/ODET_UN/status/2072336761682813433) [^28]: [𝕏 post by @Yoshua_Bengio](https://x.com/Yoshua_Bengio/status/2072350544056688935) [^29]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2072427328689619241) [^30]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2072443180004544909) [^31]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2072454019902267415) [^32]: [𝕏 post by @NVIDIARTXSpark](https://x.com/NVIDIARTXSpark/status/2072408414932894206) [^33]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2072413378669134306)