Anthropic Abruptly Reverses Its "Secret Performance Degradation" of Fable 5

It came to light last week that Anthropic had been secretly degrading performance on AI research–related tasks for some users of its flagship model, **Claude Fable 5**. The company reversed course within about a day of the discovery, but the incident has ignited a major debate across the AI industry.

The crux of the issue is not the existence of the safety guard itself, but the fact that **behavior was changed at the model layer without informing users**. Several developers, including Simon Willison, pointed out that "safeguards themselves are understandable, but opaque manipulation erodes the trust between users and providers."

Ryan Greenblatt (formerly of Redwood Research) drew a clear line: restrictions on frontier AI research can be justified in principle, but secret sandbagging is improper. He proposed introducing an access program with KYC/monitoring instead. From an engineering perspective, Gergely Orosz recommends consuming models through a **provider-independent router/harness** so you can switch immediately if the terms or behavior become problematic.

That said, Fable 5's capabilities themselves are highly rated: it scored 87.8% on the WeirdML benchmark (previous models came in under 70%) and took first place on FrontierSWE. In real-world use, however, high costs and unnatural outputs have also been reported — one user looked back on a 10k-line PR that cost about $250, saying it "wasn't worth it."

Automated AI Research Systems Keep Setting New SOTAs

**Recursive SI** announced an "automated open-ended discovery system" that achieved SOTA on three public tasks. It improved the average score on NVIDIA SOL-ExecBench from 0.699 to 0.754, shortened the NanoGPT Speedrun from 79.7 seconds to 77.5 seconds, and reached the same loss value on NanoChat 1.3x faster. The discoveries have been released as open source.

On the same day, Microsoft Research's **Arbor** also drew attention. An autonomous research agent that uses continuous refinement of hypothesis trees, it outperformed Codex and Claude Code on six research tasks and recorded an 86% Any-Medal rate on MLE-Bench Lite. With the arrival of Recursive and Arbor, "agents for AI research" are splitting into two directions: fast-iterative system optimization and long-term hypothesis management.

DiffusionGemma: Text Diffusion Model Delivers a 4x Speedup

**DiffusionGemma**, unveiled by DeepMind's Demis Hassabis, achieves **4x inference speed** compared to other Gemma 4 variants. The story that the demo ran so fast it had to be deliberately slowed down for the audience became a talking point. It's a significant milestone signaling that non-autoregressive text generation is entering practical use.

Local inference is getting faster as well: Unsloth released Gemma 4 MTP GGUF, achieving 52 tok/s → 162 tok/s (1.4–2.2x faster) on the 12B model while running in 6GB of RAM. Baseten, meanwhile, released Inception Mercury 2, claiming 1,000+ tok/s for diffusion LLM serving.

MiniMax Open-Sources Its Sparse Attention Kernels

MiniMax has open-sourced its high-performance **MSA (MiniMax Sparse Attention) kernel library**. Model weights are scheduled for release soon. Also on the same day, Together published a detailed walkthrough of its end-to-end serving stack design for the M3 model, covering KV-block-major sparse attention, the integration of MSA with paged KV cache, and multimodal preprocessing in a Rust gateway.

There was also a post on inference improvements in FlashAttention-4, showing that we have entered an era where performance gaps are determined not only by model architecture but by **design choices across the entire serving stack**.

InfiniteKV: A KV Cache That Doesn't Evict Old Tokens

**InfiniteKV** drew significant attention in Reddit's LocalLLaMA community. Conventional KV caches had no choice but to evict old tokens; InfiniteKV instead compresses old tokens into 104-byte searchable records stored in RAM or on disk. The most recent 256 tokens are kept on the GPU, and relevant tokens are retrieved from the past and referenced at each generation step.

On Mistral-7B (trained with a 32,768-token window), it correctly answered a key embedded at token 76,747. Even 1M tokens fit in roughly 3GB of records, a dramatic memory reduction compared to the conventional 122GB (float16). The author themselves admits it is "not perfect yet," and CUDA/C++ kernel optimization remains an open challenge.

Agent Infrastructure Is Maturing: Scheduling, Review, and Observability

The infrastructure for AI agents is maturing rapidly.

Claude Managed Agents added support for scheduled runs and environment variables. Credentials are swapped at the network boundary, a design that keeps secrets from being exposed to the model. Perplexity integrated Deep Research as a native skill within Computer.

Cursor enabled automatic review by default for new users, with a classifier sub-agent gating actions at 97% accuracy. Cognition open-sourced `/handoff`, which delegates jobs from local coding agents to cloud Devin. LangChain's LangSmith LLM Gateway provides spend limits, PII detection, trace continuity, and audit logs.

The common trend is a shift away from the "best model" debate toward **execution control, review layers, observability, and portability**. The concept of the "Loop" advocated by Andrej Karpathy — designing automated loops rather than having humans prompt — is taking root as a foundational philosophy of agent development.

---