# Anthropic Restores Fable 5 as Science and Media AI Turn More Operational *By AI News Digest β€’ July 1, 2026* Anthropic is bringing Fable 5 back with tighter safeguards, while NVIDIA, OpenAI, and Google pushed AI deeper into scientific workflows, harder biological evaluation, faster media generation, and cheaper inference. ## Practical deployment was the clearest theme Today's most important updates were less about raw model novelty and more about *where* AI can be used, *how* it is constrained, and *what kinds of work* it can now handle. Anthropic's policy reset, new science-focused systems, faster media models, and sharper inference economics all point in that direction [^1][^2][^3][^4]. ### Anthropic restores Fable 5, but with a tighter safety perimeter Commerce lifted export controls on Claude Fable 5 and Mythos 5, and Anthropic said it will begin restoring access tomorrow [^5]. Anthropic separately said Fable 5 returns globally with new classifiers aimed at blocking more cybersecurity tasks; some routine coding and debugging requests will fall back to Opus 4.8 while the company refines false positives, and it is drafting a common jailbreak-severity framework with Amazon, Microsoft, Google, and other partners while expanding pre-release testing work with the U.S. government [^1]. **Why it matters:** This is a concrete example of frontier access reopening only alongside tighter safeguards and more formal coordination around misuse response [^1]. ### AI for science is getting both better tooling and harder evaluation Anthropic's Claude Science workbench lets scientists use natural language to run end-to-end research workflows, and it integrates NVIDIA's BioNeMo Agent Toolkit so agents can call accelerated genomics, single-cell, cheminformatics, and biomolecular tools inside the same environment [^2]. NVIDIA said 18 of the top 20 pharmaceutical companies already use BioNeMo, and Anthropic is taking Claude Science into public beta [^2]. OpenAI, meanwhile, introduced GeneBench-Pro, a benchmark for whether agents can navigate messy biological data, choose analysis paths, and make judgment calls that computational biology research depends on; the tasks are framed as 20-40 hour problems for human experts, and Greg Brockman said GPT-5.6 Sol is a big step forward on the benchmark [^6][^7]. **Why it matters:** The science story is shifting from general assistant claims toward domain workflows and tougher task definitions that look more like real research [^2][^6][^7]. ### Google pushed generative media further into low-latency product use Google DeepMind shipped Nano Banana 2 Lite as its fastest and cheapest Gemini image model, with text-to-image generation in about four seconds for quick ideation; Logan Kilpatrick described it as under four seconds per image at $0.034 per 1,000 images [^3][^8][^9]. At the same time, Gemini Omni Flash became available in Google AI Studio, the Gemini API, and Gemini Enterprise Agent Platform for conversational video editing, multimodal input combination, real-world knowledge use, and linking text or graphics to video actions; Kilpatrick said it is state of the art for video editing at $0.10 per second [^10][^11][^9]. **Why it matters:** Faster image generation, editable video, and image-to-video chaining in the same stack make media models more usable for iterative developer workflows, not just one-off demos [^11][^8]. ### Inference efficiency kept moving fast NVIDIA said its Blackwell inference software stack cut token costs on DeepSeek V4 by up to 5x in about one month, while stacked optimizations such as disaggregated serving, large expert parallelism, NVFP4, and multi-token prediction raised token throughput per GPU by up to 20x [^4]. NVIDIA also cited deployments by Baseten, Cognition, Deep Infra, and Together AI, and said open-source frameworks including vLLM, SGLang, and PyTorch reached the same 5x performance gains on Blackwell over roughly the same period [^4]. **Why it matters:** For teams tracking serving economics, deployable cost can now move materially within weeks, not just across hardware cycles [^4]. ### Two research results showed agents moving deeper into hard reasoning and reusable skills A simple LLM pipeline using GPT 5.5 Pro and Claude Opus 4.8 was reported to resolve nine open problems spanning theoretical computer science and commutative algebra; a separate description characterized the setup as a prover-verifier loop and said one of the solved problems had remained open for two years [^12][^13]. In robotics, NVIDIA GEAR Lab and collaborators introduced ASPIRE as an automated skill-discovery system that continuously accumulates reusable robot skills for multitask, sim-to-real, and cross-embodiment transfer, with up to a roughly 10x reduction in transfer-learning tokens and a gallery covering more than 150 tasks and more than 90 learned skills [^14][^15][^14]. **Why it matters:** The common thread is reuse: formal reasoning that can tackle open problems, and physical know-how that can transfer across tasks and hardware [^12][^15][^14]. --- ### Sources [^1]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2072163884430229756) [^2]: [NVIDIA BioNeMo Agent Toolkit Brings Accelerated AI to Life Sciences Researchers in Claude Science](https://blogs.nvidia.com/blog/claude-science-bionemo-agent-toolkit) [^3]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2071988044878516466) [^4]: [How NVIDIA’s Inference Software Stack Powers the Lowest Token Cost](https://blogs.nvidia.com/blog/inference-software-lowest-token-cost) [^5]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2072106151890809341) [^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2072004836674167294) [^7]: [𝕏 post by @gdb](https://x.com/gdb/status/2072191801122038207) [^8]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2071988047445438475) [^9]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2071988351083921690) [^10]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2071988050012303710) [^11]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2071988052478615902) [^12]: [𝕏 post by @binghuip](https://x.com/binghuip/status/2070756087998152855) [^13]: [𝕏 post by @WeinsteinOmri](https://x.com/WeinsteinOmri/status/2071940117875282030) [^14]: [𝕏 post by @DrJimFan](https://x.com/DrJimFan/status/2072004190856212902) [^15]: [𝕏 post by @guanzhi_wang](https://x.com/guanzhi_wang/status/2072001345960222743)