Claude Fable 5 đ, Gemini 3.5 Live Translate đ±, scaling test time compute đ
Short summary
This roundup covers Googleâs Gemini 3.5 Live Translate for seamless, real-time speech translation and Anthropicâs rollout of Claude Fable 5 (with hidden safety tweaks) and Mythos 5, backed by a $35 billion chip-lease guarantee from Google. It also digs into emerging trends like text as an optimization layer, the impact of test-time compute on LLM benchmarks, and updates on AI agent identities and retrievers.
- Anthropic quietly throttles Claude Fable 5's responses ~0.03% of the time (mainly to block rivals training on it) via invisible prompt/fine-tuning tweaks, not model swapsâso users can't tell when they're getting a degraded answer.
- Google is backing a $35B chip-lease deal for Anthropic across five data centers, showing how tightly the two companies' infrastructure and business interests are now intertwined.
- Test-time compute, not architecture, is now the main driver of LLM gainsâGPT-5.5 barely beats GPT-5.4 on raw benchmarks but pulls ahead once cost, latency, and token count are factored in, making single-score comparisons misleading.
- Fully automated AI engineering loops tend to produce sloppy agents because they optimize against imperfect evals, missing nuances a human developer would catch.
Full summary
Gemini 3.5 Live Translate is Googleâs new real-time speech-to-speech model, handling 70+ languages with natural intonation and no awkward pauses. Itâs in private preview on Google Meet and rolling out to Android and iOS Translate. Meanwhile, Anthropic just launched Claude Fable 5 for everyone and Claude Mythos 5 for select cyber defenders and infrastructure partners. Google is backing Anthropicâs massive $35 billion chip lease across five data centers, a deal that underscores how deep their AI partnerships have become.
On the theory side, one piece frames textâprompts, memory, retrievalâas an âoptimization layer,â treating prompt design as a form of sample-efficient learning. Another argues that fully automated AI engineering loops produce sloppy agents because they optimize against imperfect evaluations, missing the nuance only a developer catches. And large-scale test-time compute now drives LLM performance more than model architecture. GPT-5.5 barely outperforms GPT-5.4 on raw benchmarks, but it dominates when you factor in token count, cost, or latencyâshowing that single-number scores are losing their meaning.
In engineering news, Teleport is giving each AI agent its own cryptographic identity instead of shared credentials, cutting standing privileges and improving audit trails across Kubernetes, databases, and clouds. Cohere released North Mini Code, a 30 billion-parameter MoE coding model with 3 billion active parameters under Apache 2.0. Evo ported its autoresearch orchestrator into Claude Codeâs dynamic workflows, moving coordination out of model memory and into deterministic JavaScript. And FlashMemoryâs DeepSeek-V4 retriever keeps just 10â15% of the KV cache on GPU, yet maintains or improves downstream performance by predicting which chunks matter most.
Anthropic has also embedded hidden safeguards in Fable 5 that can quietly limit its responses in about 0.03% of casesâmostly when competitors use it to build rival models. These interventions, applied via prompt tweaks and fine-tuning rather than model switches, arenât visible to users. That opacity creates a supply-chain risk: you wonât know if your agent stopped giving you full answers. Anthropic argues itâs protecting its lead, but uneven, invisible safety policies erode trust across the AI ecosystem.