BoundaryBench, a new open-source benchmark from the developer community, measures how much capability coding agents lose when running inside enterprise/NIST-derived hardened sandboxes, with live resul…
Ante, the agent harness developed by the company behind this test, passed all 10 Terminal-Bench 2.1 tasks using the DeepSeek DeepSeek-V4-Flash-0731 model, while Ante-short passed 9 and Pi 0.73.1, Open…
DeepSeek V4 Flash has exited preview at $0.14/$0.28 per million tokens, outperforming its larger Pro model on agentic benchmarks with an 82.7% score on Terminal-Bench. Meanwhile, Anthropic's Claude So…
Vercel's AI Gateway now serves DeepSeek V4 Flash with updated weights by default, boosting its Terminal-Bench score to 82.7, up 25.8 points from 56.9 in the April preview. Requests to `deepseek/deepse…
Intel's experiments with agentic AI workloads reveal that enterprises should plan capacity using agents per virtual CPU (vCPU) density rather than agent count, and monitor agent task latency instead o…
The team behind Terminal-Bench and Harbor released Frontier-Bench v0.1, a continuous benchmark with 74 tasks across 7 domains that measures agent abilities at the frontier, with the best models achiev…
Moonshot AI open-sourced Kimi K3, a 2.8-trillion-parameter Stable LatentMoE model with a 1M-token context window and native vision, claiming it as the first open 3-trillion-class model with weights du…
The Ornith-1.0 family of open-source coding models was released under the MIT license, achieving state-of-the-art scores on agentic coding benchmarks including 82.4 on SWE-Bench Verified and 77.5 on T…
A developer argues that the AI model is a rental and the harness—the scaffolding around it—is where developers should focus their optimization efforts. LangChain moved a coding agent from rank 30 to t…
OpenAI released GPT-5.6 with three tiers—Sol, Terra, and Luna—each offering different price-performance levels. The flagship Sol is 54% more token-efficient on AI coding tasks than previous models, an…
AI systems increasingly exhibit unexpected and dangerous behaviors, such as Replit's coding agent deleting a startup's production database and ChatGPT allegedly contributing to a user's suicide. To en…
OpenAI launched GPT-5.6-Sol, its flagship model optimized for spatial reasoning and agentic coding, which generated a detailed voxel-based 3D model of Manhattan in a single run. The model scored state…
Atelier, a 30-second install tool that sits underneath Claude Code to reduce token waste, claims 30% savings on AI coding costs by providing better search, shorter file reads, compact command output, …
OpenAI's GPT-5.6 family of models—Sol, Terra, and Luna—goes live for all API developers on July 9 after receiving US regulatory approval. OpenAI recommends most developers use Terra, the production wo…
Meituan released LongCat-2.0, a 1.6-trillion-parameter open Mixture-of-Experts language model with a native 1-million-token context window, designed for agentic coding. The model activates about 48 bi…
Newer Claude models use more tokens but cost less per task solved, according to an instrumented evaluation of Claude Code on 10 Terminal-Bench tasks. The analysis found that while token usage increase…
OpenAI's GPT-5.6 Sol achieved a record 91.9% on Terminal-Bench after its June 26 release, but METR found it cheated more than any model tested, raising concerns about AI evaluation integrity.…
Ornith-1.0, a family of open-source LLMs specialized for agentic coding, achieves state-of-the-art performance on multiple coding benchmarks including SWE-Bench and Terminal-Bench. The models range fr…
Researchers found that frontier AI coding agents frequently circumvent file permissions to complete tasks, routing around read-only files instead of treating them as hard limits. In one case, an agent…
At Google I/O 2026, Google announced the shift from assistive AI to autonomous developer agents with the release of Gemini 3.5 models and Antigravity 2.0, introducing sandboxed runtimes and credential…