This week's new AI releases

Only releases with a dated official source. Updated .

  1. Mintlify Autopilot (beta)

    Always-on docs agent that watches your repos, Slack and support tickets, drafts doc updates when code changes and opens a pull request for review; in beta on Pro and Enterprise, 250 credits per PR.

  2. Mercury 3 (Inception)

    Diffusion LLM running at 1,000+ tokens/s with a 1M-token context; OpenAI API compatible on Inception's API, OpenRouter, Baseten and Vercel. Launch price $0.10/$0.35 per 1M input/output tokens (50% off list).

  3. MiniMax M3.1 Flash

    Agentic coding and computer-use model, nearly 3x faster than its predecessor with 70% lower agent costs; $0.15 input / $0.60 output per 1M tokens on the MiniMax API and in MiniMax Code. Weights to follow later.

  4. Vercel AI Gateway session view

    AI Gateway logs now group requests into sessions, showing the coding agent behind each one (Claude Code, Codex, Cursor), cost, models, providers and cache hit rate per turn.

  5. DeepSeek V4.1 Flash 1017 + Search API

    Upgraded DeepSeek V4.1 Flash with better coding and agent benchmarks, faster runs and shorter outputs at the same API price (model deepseek-flash); DeepSeek also opened its in-house web Search & Fetch API.

  6. Reflection Beam 1.0 (open weights)

    Reflection released weights for Beam, a 501B-parameter (23B active) MoE model with 1M context for coding and agentic work, on Hugging Face under Apache 2.0; also on Reflection's API.

  7. Cursor cloud agent canvases + Environments API

    Cursor cloud agents can build shareable interactive canvases; an Environments API (beta) manages agent environments, which now support up to 300 repos and load 5x faster.

  8. Hy-Memory 2.0 (Tencent Hunyuan)

    Open-source agent memory stack: the Hy-Memory-30B-A3B memory model, a Python SDK with chat, working and coding memory (pip install hy-memory), and the MemHub app that brings memory into your AI tools.

  9. Muse Spark Open (Meta)

    Open-weight multimodal model (816B total / 42B active, 1M context) built for coding and agents; weights on Hugging Face.

  10. Eney by MacPaw

    Proactive AI assistant for Mac and iPhone that prepares next steps across apps, files and calendar.

Earlier releases

  1. NVIDIA Nemotron 3.5 Bolt 74B

    Model purpose-built for local agentic AI on RTX Spark PCs.

  2. Python on Vercel (GA)

    Python is now generally available on Vercel: FastAPI, Flask and Django are auto-detected, with Python SDKs for Sandbox, AI SDK (beta), Workflow (beta), Queues and Blob.

  3. Liquid AI d1-ultra

    Multimodal decision model that returns calibrated probabilities for yes/no, choice and score questions over text and images; $0.042 per million input tokens, no output-token charge.

  4. BAND for Developers

    Desktop app and CLI that connect coding agents across Claude Code, Codex, Copilot and other harnesses so they can share context, hand off tasks and route decisions to you.

  5. Jais 3 (IFM / MBZUAI)

    Six open Arabic-centric agentic models from 0.9B to 375B parameters with a 512K context, plus Jais Speech; weights, checkpoints and code on Hugging Face under Apache 2.0.

  6. GitHub agentic autofix for code scanning (GA)

    Assign a code scanning alert to Copilot and it explores the codebase, proposes a fix, reruns CodeQL to confirm it and opens a draft PR; up to 25 alerts per PR. For Code Security / Advanced Security on GitHub Enterprise Cloud.

  7. Grok Voice 3.0

    Speech-to-speech model for voice agents, live in the Voice Agent API at $0.08/min (same price as Grok Voice 2.0); grok-voice-latest now points to it.

  8. Zed 1.24

    Zed's weekly stable release adds Terminal tabs in the Agent Panel, better custom Amazon Bedrock models, lower CPU and memory use in long agent threads, and Git tag creation.

  9. Gemini 4 Argon (Google)

    Google's latest frontier model, available via the Gemini API with a 1M token limit.

  10. Klaviyo Composer

    AI marketing agent now available to every Klaviyo customer.

  11. Cursor Origin App Marketplace

    Origin now integrates with 20+ apps for deploys, CI, code review, security and monitoring (Cloudflare, Vercel, Linear, Sentry and more), and is open to public app submissions.

  12. Notion 3.8: MCP, 150+ connections, Skills API

    External agents can read and update Notion over MCP with admin controls, Notion agents connect to 150+ tools, and a new Skills API syncs skills from Notion to GitHub and other agents.

  13. Kiro local command sandbox

    Kiro IDE 1.3 and CLI 2.31 can run agent shell commands and local MCP servers in an OS-enforced sandbox with blocked, limited or open network access.

  14. LFM2.5-24B-A2B (Liquid AI)

    Open-weight 24B MoE reasoning model with 2B active parameters and a 131K context, built for fast private agents on-device (184 tok/s on an M5 Max); API at $0.10/$1.00 per 1M input/output tokens.

  15. Muse Wonder (Meta)

    Meta's frontier models for personal and professional agents and coding; Muse Wonder is on the Meta Model API and Muse Code, Muse Wonder Full for Muse Code subscribers.

  16. Claude Haiku 5.5 (Anthropic)

    Anthropic's fastest, lowest-cost small model, on the Claude Platform, AWS, Google Cloud and Azure.