AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, advancing AI capabilities for coding and knowledge work. Fable 5.1 costs approximately 25% less than its predecessor for typical workloads (up to 45% less for agentic tasks) due to 75% reduced pricing on cache reads. The model demonstrates improved performance across multiple benchmarks: Terminal-Bench 4.0 (55.8%), Humanity's Last Exam with tools (65.0%), CursorBench 3.2.0 (73.4%), and OSWorld 2.0 (41.7% strict). Fable 5.1 shows stronger reasoning for debugging and root-cause analysis, with real-world validation from users including Jane Street Capital, Cognition, and others reporting superior performance over Fable 5 and Opus 5 across coding and long-running problem-solving tasks. Fable 5.1 introduces more precise safeguards reducing false positives by 60% in cybersecurity contexts while now permitting vulnerability discovery work. Claude Mythos 5.1, an unrestricted variant, is available through trusted access programs for vetted cyberdefenders and life sciences professionals. Scientific capabilities expanded to include protein binder design achieving 50% hit rates (versus typical 10-15%) and computational biology optimizations delivering 1.4x-2.5x speedups on deep learning models, reducing GPU costs by 30-60% on genome-wide analyses.

Ollama released v0.34.0-rc1, adding ChatGPT Desktop integration to run Ollama models directly in the app, improving structured output performance on Apple Silicon, and adding support for OpenAI-compatible client tool search and response compaction. v0.33.3 brought Gemma4 image and audio support on MLX engine with MLX and llama.cpp updates. v0.32.15 improved time-to-first-token performance by caching resolved model metadata, reducing TTFT from approximately 995ms to 524ms in benchmarks. Recent releases also added support for Qwen3.8 27B with optimizations for Apple Silicon, DeepSeek Harness integration for agentic coding tasks, and structured output support across various model runners.

Meta shipped Muse Spark 1.3 on September 2, 2026, an agentic coding model that achieves approximately 20% fewer tool calls and 25% fewer tokens than its predecessor Muse Spark 1.2 on equivalent tasks, according to Meta's internal engineering comparisons. The model supports a 1,048,576-token context window and maintains standard pricing at $1.25 per million input tokens and $4.25 per million output tokens, delivering an effective cost reduction despite the efficiency gains. On benchmarks, Muse Spark 1.3 scored 98.5 on MRCR long-context retrieval, 75.4 on DeepSWE v1.1, 88.8 on Terminal-Bench 2.1, and 1754 on GDPVal-AA v2 agent performance (compared to Opus's 1824), indicating optimization for coding and long-context workloads over general agent capability. A discounted contributor tier is available at $0.10 per million input tokens and $0.20 per million output tokens, requiring Meta training rights on prompts and completions.