AI developer tools · What shipped

Claude Haiku 5.5 launched as the cheapest, fastest, and most capable small model for high-volume, cost-sensitive tasks. Claude Sonnet 5.5 and Claude Opus 5.5 also shipped as part of the new Claude 5.5 family, with Opus 5.5 performing at the level of Claude Fable 5.1 while costing 40% less to run than Opus 5. Claude Fable 5.1 and Claude Mythos 5.1 were previously released as the world's most advanced models for coding and knowledge work.

Recursive launched Flint, a free open-source coding model designed to run on laptops, built largely with AI assistance from the company's Eureka Machine research system. The mixture-of-experts model uses 2.5 billion of its 40 billion parameters per task, achieving approximately 131 tokens per second on a MacBook Pro M4 Pro—roughly twice the speed of comparable models like Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On SWE-Bench Verified, a benchmark for fixing real-world software bugs, Flint outperformed other models under 50 billion parameters, with only Alibaba's Qwen3.8 27B scoring higher across Recursive's six-test average. The model is available for free download on Hugging Face under the Apache 2.0 open-source license.

oMLX 0.7.2 ships faster DeepSeek V4.1 inference across all quantization levels with fused decode kernels, Lightning MTP support in distributed inference across multiple Macs, and Qwen3.8-Flash-Next quantizations on Splash. The release includes performance gains of 72-86% on decode for oQ2-oQ3 variants and 13-89% improvements in distributed tensor parallelism scenarios, alongside improvements to Thunderbolt networking warnings and multi-part GGUF handling. oMLX 0.7.1 adds GGUF downloads through Splash, much faster DeepSeek V4.1 decode with fused kernels and concurrent request optimization (+64% single request, +254% aggregate with 4 concurrent, +373% with 8), 4-bit and 8-bit KV cache formats (Affine4/Affine8) reducing decode memory by 49-74%, and fixes for GPU hangs when running multiple models simultaneously. GLM-5.3-Flash, Qwen3.8, and Qwen3.6 now run on Splash external engine with DFlash2 speculative decoding while staying in oMLX's model list and API. oMLX 0.7.0 brings completely rebuilt memory guard, faster Qwen3.8-Flash-Next and GLM-5.3-Flash inference with fused kernels and exact Lightning MTP matching serial decode output, faster MiMo V2 with blocked sliding-window attention, and fixes for regressions in ANE prefill and oQ quantization. oMLX 0.7.1rc1 introduces Muse Spark Open (Meta's 816B MoE) support, model recommendations showing only compatible oQ builds for each Mac, Splash as external engine for Qwen models with DFlash2, reworked model settings with sectioned UI, DFlash in batched engines for concurrent requests, and decision model support via POST /v1/decisions following OpenAI's API.

Mistral AI released Voxtral 3 on October 20, 2026, a speech processing platform comprising three transcription models, one text-to-speech model, and a LiveKit-based voice agent SDK. Voxtral Small Transcribe 3 achieved 2.1% average accent word error rate, placing third globally on Artificial Analysis benchmarks for speech-to-text accuracy. The integrated SDK enables developers to build real-time voice agents with combined speech recognition, synthesis, and conversational AI capabilities, reducing engineering overhead for production-grade voice applications.