AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, advancing AI code generation and developer tools. Fable 5.1 achieves significantly improved performance on multiple coding benchmarks: 73.4% on CursorBench 3.2.0 (vs 70.5% for Fable 5), 55.8% on Terminal-Bench 4.0 agentic coding, and 60.9% on Humanity's Last Exam with tools. The model demonstrates 25% lower costs for typical workloads and up to 45% savings for agentic work due to 75% reduced cache read pricing ($0.25 per million tokens). Key improvements include enhanced capabilities for long-running problem-solving, better root-cause analysis in debugging, and reduced false positives in safeguards—cybersecurity safeguards now fire 60% fewer times on benign content while supporting vulnerability discovery work. Mythos 5.1 (same underlying model with permissive safeguards) showed exceptional performance on scientific tasks: designing protein binders with hit rates near 50% (vs typical 10-15%), optimizing GPU kernels for genomics models up to 2.5x faster, and creating high-resolution Venus elevation maps. The models are available on all major platforms with API access via claude-fable-5-1, while Mythos 5.1 requires trusted access through cybersecurity and life sciences verification programs.

MILO: A block-wise low-rank KV cache compression framework for efficient many-shot in-context learning. The paper addresses the memory bottleneck in many-shot ICL by exploiting low-rank redundancy in key-value caches. MILO achieves up to 50% KV cache memory reduction and 1.8x throughput improvement on Qwen2.5 models through block-wise compression with dynamic rank allocation based on information entropy. Experiments on classification and reasoning benchmarks show the method outperforms prior baselines like ASVD and Palu by up to 4.5% accuracy while maintaining negligible performance degradation. System optimizations include fused Triton kernels, CUDA streams, and CUDA graphs for efficient reconstruction during inference.

Jan released Jan-Code-4B, a lightweight 4B parameter code generation model fine-tuned for fast local inference on consumer hardware. The model outperforms Jan-v3-4B-base-instruct and Qwen3-4B-Instruct across three coding benchmarks: Aider (19.0 vs 18.0), Livecode Bench v6 (51.0 vs 45.8), and AIME25 math reasoning (53.0 vs 47.0). Jan-Code-4B supports code generation, editing, refactoring, debugging, and agentic workflows with OpenAI-compatible API access, running on 8GB minimum RAM with quantized variants and deployment options via vLLM or llama.cpp.

Aikido Security released Altar-1, an open-weight security model compressed from Z.AI's GLM-5.3 (753B parameters) to 328 GB through quantization (W4A16) and expert pruning using Cerebras REAP, reducing model size by 78.2% while retaining 92% of CVE coverage. The model runs on a single node of 4x NVIDIA H200 GPUs with vLLM, supports 131k context length, and achieved 60.4% average recall on Aikido's internal 32-CVE benchmark versus 65.6% for the full BF16 parent model, with negligible fidelity divergence (0.506 nats KL divergence). Weights are publicly available on Hugging Face under the GLM-5.3 license, enabling on-premise and air-gapped deployments for organizations with data-residency requirements.