Recent AI dev tool releases, with sources.
Reflection's first model, Beam, is now on Hugging Face under Apache 2.0: a 501B-parameter MoE (23B active) with a 1M-token context, built for coding and agentic work. It ships in BF16, FP8 and NVFP4 with day-0 support in transformers, vLLM and SGLang, plus a Rust-based harness.
Source →Cursor cloud agents can now build interactive canvases as part of their work, and you can share a finished one with your team by sending its link. The same update put an Environments API in beta and raised cloud environments to 300 repos, with large ones loading 5x faster.
Source →Cline Cloud sessions, rolling out gradually, run a task in a hosted sandbox on a GitHub repo and branch you pick, and keep going after you close your editor. Cloud tasks show up in Recent and History with their status and notify you when they finish.
Source →Autonomous launched Harness, a free open-source Mac/Linux app that runs Claude Code, Codex, Cursor, OpenCode and others side by side across machines, with a terminal client and iOS/Android apps for voice replies. Agents keep working after you close the app; traffic is end-to-end encrypted (MIT, github.com/autonomous-ai/openharness).
Source →Anthropic took Opus 5.5 Fast Mode out of research preview: no waitlist and no beta header. Same model, up to 2.5x higher output tokens/sec at $8/$40 per MTok, with default rate limits on Start/Build/Scale tiers; also live on Claude in Microsoft Foundry (Azure-hosted).
Source →OpenCode made DeepSeek V4.1 Flash Free (model id deepseek-v4.1-flash-free) $0 for input, output and cached reads for a limited time on Console/Zen. Pick the Free variant in /models; the paid deepseek-v4.1-flash tier stays listed separately.
Source →Meta released an open-weight multimodal model (816B total / 42B active params, 1M-token context) aimed at coding and agentic work. It's available via Groq, OpenRouter, Together, Fireworks and others, and on Hugging Face.
Source →OpenCode 2.0 runs as one shared background service across TUI, web and desktop, with a rebuilt API/SDK and a JavaScript Code Mode. 1.0 plugins and the old server API break; upgrade with `opencode upgrade --major`.
Source →Code scanning's AI Scan for pull requests is now generally available, complementing CodeQL for languages it doesn't cover. Agentic autofix for code scanning alerts also reached GA.
Source →HydraFusion got an updated model pool and better orchestration, plus a new Max option. Both are available in VS Code, the Copilot app and more.
Source →Google's newest frontier model is now GA in Copilot for Pro+, Max, Business and Enterprise users. It's billed at provider list pricing under usage-based billing.
Source →JetBrains shipped a stable LSP-based extension bringing IntelliJ's Java/Kotlin intelligence to VS Code and Cursor, with dedicated Cursor Agents Window support. It now requires an IntelliJ IDEA Ultimate subscription after a 30-day trial.
Source →Anthropic's automatic PR reviewer left research preview for Team and Enterprise plans. Reviews are about 3x faster than in preview, and admins pick Quick, Standard or Deep per repo.
Source →Moyai is a self-hosted cloud coding agent: assign a task in Slack or the browser and get back a PR with test results. It runs Codex, Claude Agent SDK and others through your LiteLLM gateway.
Source →