AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, setting new performance standards across multiple benchmarks. Fable 5.1 achieves 52.6% on Terminal-Bench-Science 0.1 (agentic scientific research), 60.9% on Terminal-Bench 4.0 (agentic coding), 60.9% on Humanity's Last Exam (multidisciplinary reasoning with tools), and 73.4% on CursorBench 3.2.0 (agentic coding)—outperforming Fable 5 and Opus 5 across these benchmarks. Pricing is reduced approximately 25% for typical workloads and up to 45% for highly agentic work due to 75% lower cache read costs ($0.25 per million tokens). The model demonstrates improved capabilities in coding, knowledge work, long-context reasoning, and scientific research, with Mythos 5.1 offering more permissive safeguards for cybersecurity and life sciences professionals. Fable 5.1 is available on all major platforms and the Claude API with production safeguards enabled.

Artificial Analysis Intelligence Index v4.3.2 incorporates 10 evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, and others, measuring AI model performance across capability-specific indexes for Finance & Accounting, Strategy & Ops, Legal, Healthcare & Medical, Engineering, and Economics. The benchmark evaluates models on cost per task, output speed in tokens per second, and overall intelligence scores, with GPT-6 Luna positioned in the most attractive cost-performance quadrant.

OpenBMB released MiniCPM5-2B in September 2026, a 2.5-billion-parameter language model that achieved an average score of 53.9 across 34 benchmarks covering coding, math, instruction following, general knowledge, long-context understanding, tool use, and agentic tasks, beating larger open-source rivals including a 4-billion-parameter model that scored 51.1. The model supports a native context window of 131,072 tokens and is designed for efficient self-hosted deployment. In the same week, Ornith AI released Ornith-1.5-9B, a 9-billion-parameter model achieving 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified for coding-agent tasks, while Liquid AI shipped LFM2.5-VL-3B-DSpark, a 279.5-million-parameter draft model for speculative decoding, and Supersonic Labs released Julia 1, a 144.3-million-parameter decision model. The releases reflect a broader shift toward squeezing capability from smaller parameter counts through distillation from frontier models, curated training data, and architecture optimizations focused on reasoning-per-parameter efficiency rather than raw scale. These developments are intensifying price competition with frontier models, as evidenced by DeepSeek's recent 70 percent output price cut and similar moves across the industry, while expanding the viability of on-device and edge deployment compared to larger hosted models.

Kimi K3 and Gemini 2.5 Pro have been benchmarked across multiple AI evaluation frameworks with version numbers and public scores. Kimi K3 scores higher overall at 71.87 versus Gemini 2.5 Pro's 50.24, with particular strength in coding tasks (62.6 vs 24.6) and knowledge benchmarks like GPQA and HLE. Gemini 2.5 Pro offers lower API costs across standard workloads—chat turns at $0.00625 versus $0.0105, and cache-heavy agent loops at $0.15 versus $0.27—while Kimi K3 provides a larger documented context window at 1.05M tokens.