AI Giants Race to Scale Infrastructure and Capability, While Safety Concerns Mount

Model Releases

• Anthropic launched two new models, Claude Fable 5 and Claude Mythos 5, claiming significant advances in coding, science, and general knowledge work [46, 48]. Fable 5 is available for general use and reportedly hits 80.3% on SWE-Bench Pro, significantly outperforming Claude Opus 4.8's 69.2% [46]. While Mythos 5 shares the same base model as Fable 5, it removes safeguards in areas like cybersecurity and is currently limited to select partners via Project Glasswing [46, 48]. Both models are priced at $10 per million input tokens and $50 per million output tokens, making them more expensive than Claude Opus 4.8's rates [46, 48]. ^[46][48]
• Google launched Gemini 3.5 Live Translate, a real-time speech-to-speech model that automatically detects and translates in over 70 languages while preserving the speaker's intonation, pacing, and pitch [45, 47, 50]. The model is rolling out across multiple platforms: developers can access it via public preview through the Gemini Live API or AI Studio; businesses are getting private preview access in Google Meet, where language support increases from five to over 70 languages and enables more than 2,000 language combinations [47, 50]; and all users can use it in the Google Translate app on Android and iOS [47, 50]. Furthermore, the service is being tested by ride-hailing service Grab for driver-passenger communication, which makes over 10 million voice calls monthly [47, 50]. ^[45][47][50]
• Google has released DiffusionGemma, an experimental open weights model that generates text using a novel diffusion process for significant speed boosts in local single-user mode. This architecture starts with 256 random placeholder tokens and refines them across multiple passes, enabling it to run up to four times faster than comparable autoregressive models on dedicated GPUs [6]. The efficiency gain stems from processing tokens in parallel, which shifts the computational bottleneck from memory bandwidth—a common limitation for traditional language models—to raw compute power, allowing hardware like the H100 and GeForce RTX 5090 to remain actively utilized [6]. While DiffusionGemma trades some output quality for speed, its ability to process entire blocks of tokens at once makes it uniquely suited for tasks that don't proceed left-to-right, such as solving Sudoku or filling gaps in code [6]. ^[6]