AI developer tools · What shipped

DeepSeek V4 Pro Base leads the BigCodeBench code-generation benchmark with a score of 59.2%, followed by DeepSeek V4 Flash Base at 56.8%, according to evaluations reported in the DeepSeek-V4 technical report as of August 24, 2026. BigCodeBench measures code generation tasks using Pass@1 scoring at the software engineering difficulty level, with two models currently evaluated on the benchmark.

Cline released multiple version updates across SDK, CLI, and desktop components with refreshed model catalogs adding new providers and models. SDK v0.0.79 and CLI v3.0.58 added AgentRouter and Opper providers with updated pricing and default model changes for Aki.io and NanoGPT. v4.1.14 expanded the built-in model catalog with Claude Fable 5, Grok 4.6 on Vertex, DeepSeek V4 Flash variants, and models from Qwen, Gemma, and others. v4.1.11 added image generation capabilities and refreshed the model catalog with AMD, Arcee, Echo, and additional providers. Multiple releases fixed tool calling for custom OpenAI-Compatible models and improved session persistence across hub restarts and upgrades.

NVIDIA released the Nemotron family of foundation models spanning language, reasoning, vision, speech, and safety. Key releases include Nemotron 3.5 Lightning, a 30B parameter small language model with hybrid Mamba-2 + Transformer MoE architecture optimized for efficient local inference with 1M-token context and multi-token prediction for faster generation via vLLM, SGLang, TRT-LLM, and other frameworks; Nemotron 3 Super, a 120B parameter model with 12B active parameters delivering up to 5x higher throughput than the previous version; and Nemotron 3 Ultra, a frontier 550B parameter model for demanding multi-agent and reasoning workloads. For physical AI, NVIDIA released Cosmos 3, an omni-model with variants ranging from 4B-class Edge to 32B-class Super, combining reasoning and generation for robotics and autonomous vehicles. Supporting these are open datasets across pre-training, code generation, math reasoning, agentic behavior, and reinforcement learning, along with reproducible training recipes and evaluation frameworks.

Google released Gemini 3.7 Flash, a new AI coding model designed as a "workhorse" for developers and agents. The model achieves significantly higher performance on software engineering benchmarks compared to its predecessor: 65.3% on DeepSWE v1.1 (vs. 49.0% for 3.7 Flash) and 43.6% on FrontierCode 1.1 Main (vs. 34.4%), with improvements in code accuracy, debugging, and autonomous error recovery. Gemini 3.7 Flash introduces adjustable thinking levels (low, medium, high) to balance latency and intelligence across different tasks, and rolls out as the default model across Google Antigravity, Google AI Studio, and the Gemini API's Managed Agents. The model is available at introductory pricing of $0.75/1M input tokens and $3.75/1M output tokens through December 31, 2026.