Each model gets one hour in its lab's first-party CLI (Claude Code, Codex, Antigravity CLI, Muse Code), with no internet, to write a C++ Protoss bot. GPT-6 Astra ranks first and GPT-6.1 Sol second at about a tenth of the cost, while Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1 tie at 37 to 39.
Point any MCP client at https://ampcode.com/mcp to get two tools: puck , which routes ideas to the right orb and checks on your Amp threads, and orbs , an interactive view of running orbs with portal links…
Released Oct 16 under the OpenMDW license, Dory 1.0 4B reruns a shared 20 layer block a chosen number of times, so one checkpoint trades speed for accuracy with no model swap; NVIDIA's own tests put it at 50.3 on SWE bench Verified with 1 loop and 61.8 with 4. It has a 256K English context and ships in BF16, MXFP8, NVFP4, and Q5 K M GGUF…
Swarmchasers logged 2,818 attempts from Sept 28 to Oct 8 by agents on Tencent Cloud to get Amap data, which switched to public code running web tools, third party readers, and Amap keys exposed in public code after being blocked; most attempts failed. Tencent said the agents were evaluating third party models rather than Hunyuan, found no deliberate targeting, and is tightening network controls and agent monitoring (full report)…
Devin Desktop v3.11.1108 on the Next channel (Oct 17) runs the same sessions sidebar, Automations, Review, Wiki, and Cmd/Ctrl+K palette as app.devin.ai, and you can drag sessions together to view up to eight at once. Every session also gets a Files view, a Changes tab with per file diffs for local sessions, and a terminal on the session's machine…
Released Oct 16 under Apache 2.0, the hy memory SDK ( pip install hy memory==2.0.3 ) adds chat, working, and coding memory for agents, and the bundled MemHub app imports and syncs sessions from tools like Claude Code, Cursor, Codex, and OpenCode into one local memory store. Tencent also opened Hy Memory 30B A3B, a 3B active memory writer model it says edges Claude Opus 5 (55.49 vs 54.62) across four memory frameworks in its own tests…
Spotted something we missed? Start a thread.