Ramp AI data shows that Anthropic's most advanced model, Fable 5, accounts for only 8% of the company's model spending, while the cheaper Opus 4.8 takes up 28%. Priced at $20 per million tokens—over ten times more than its predecessor—users are shifting to older 「good enough」 models due to cost. Observations indicate teams no longer default to using the strongest model for all coding tasks, but instead allocate Fable 5, Opus, and other models based on task value. Especially in large-scale daily usage, inference cost per token has become a key constraint.

Harvey and Fireworks have released a research preview of the legal model Tenet, based on Kimi K3 and post-trained via asynchronous reinforcement learning for long-horizon legal agent tasks. The team claims it nearly doubles task completion on their internal Legal Agent Benchmark, improves by 20% on the contract subset, and increases full pass rates by 9 and 2 percentage points respectively. Training used rank-64 LoRA and approximately 150 NVIDIA B300 GPUs over two months; currently, weights, model cards, or APIs are not open, and results have not entered public leaderboards.

FreeToken has launched an edge-native MoE inference engine claiming to run the 753B-parameter GLM-5.2 on a single workstation GPU. The system dynamically assigns off-cache experts to either GPU fill or CPU in-place computation based on measured PCIe and host memory bandwidth, then precisely merges results from both paths—without modifying routers or using approximations. Its global LRU expert cache, semantic boundary state checkpointing, and elastic memory management allow adjustment of GPU cache and reconstruction of host pools without service interruption, specifically addressing VRAM limitations of ultra-large MoE models.

DeepMind has introduced the Pandora Router, modeling model routing as a 「Pandora's box」 decision: first use a low-cost but noisy score to evaluate candidate models, and only invoke more expensive quality estimation if the expected benefit of additional information exceeds its acquisition cost. In EmbedLLM tests, average check cost dropped from 1.986 to 0.075; across MATH, RAG, and EmbedLLM settings, this method achieves the lowest or jointly lowest sum of routing regret and check cost, aiming for dynamic trade-offs between routing accuracy and overhead.

Φ-Bench has released a systematic evaluation focused on large model infrastructure engineering, covering cross-module understanding, inference serving, system optimization, system reliability, and end-to-end project iteration—rather than just single-function code generation. Results show Claude Opus 5 leads overall, Kimi K3 excels in inference serving and system optimization, and GLM 5.2 performs well in system reliability, though no model dominates across all areas. The evaluation concludes that current models still lack capabilities in long-term project management, hardware understanding, continuous validation, experimental design, and reliable attribution, with complex optimizations typically requiring multiple rounds of trial and error.

An Anthropic-related study examines the 「hidden file problem」 in multi-agent decision-making: when critical evidence is分散 across different agents, most model families achieve correct rates of only 17% to 36% after group discussion; a single agent given full evidence almost always makes the correct judgment, with Mythos 5 being the exception, achieving about 85% group accuracy. Failures are attributed to insufficient diversity among models and discussion mechanisms that fail to support dissenting information, suggesting that increasing agent count does not necessarily improve decision quality—systems must actively protect minority views and ensure private evidence is integrated.

The Online-SDFT prototype enables continuous learning of a 230M-parameter language model directly on Android devices, using delayed and ambiguous user interactions—such as opening or ignoring notifications—as soft supervision targets. The system uses a teacher-student setup with the same model: a teacher without LoRA adapters generates action probability distributions based on hindsight outcomes, which then updates the student model. Combined with controlled exploration and a small balanced replay buffer, it avoids local optima. In synthetic testing, the method matched hidden preferences in 506 out of 720 decisions, achieving 70.3% accuracy and 44.8% cumulative regret, compared to 28.2% accuracy for the static base model.

A local AI agent quantization benchmark reveals clear behavioral differences among BF16, FP8, INT8, and INT4 in long-context tasks: INT8 and FP8 maintain good model fidelity, whereas INT4 and NVFP4 exhibit consistent degradation in strict-format tasks like tool calling, with token flip rates reaching as high as 50% in some tests. The results also show that even when performing identical mathematical operations, different attention backends may alter the next token; KV Cache quantization is a major risk in long conversations, with INT8 allowing recovery but INT4 errors accumulating over context length, leading to irreversible drift.