Loading video...
Video Failed to Load
教你一个免费白嫖使用 Codex + Computer Use + Blender 的机会。 GPT-6 Astra 的额度即使是 20X 的 Pro 用起来也消耗的巨快。 看到说 Nex-AGI 能媲美 GPT-6 Astra，不管了我直接上手先来试试看✌🏻 我的想法是先让它来处理一个长程的识图任务： 丢了一张图给它，然后让它参考图例，开发一个网页小游戏：网页版麻将 等了一个半小时终于跑完，我测试完效果，我不得不说，我要继续白嫖这个模型了。 现在免费不说，能力还这么强，大家有空都用免费的模型来试试 Codex + Computer Use + Blender 吧。 我来告诉大家白嫖流程 1、当然要先安装了Codex 2、安装Codex++, 配置OpenRouter 3、选择免费模型 nex-agi/nex-n2.5-pro:free 4、Codex 中安装插件 Computer Use，并启用 5、让 Codex 安装 Blender客户端，再加上安装 Blender MCP 6、让Codex 使用... show more
63,661 views • 12 days ago •via X (Twitter)
35 Comments
Using free Nex-N2.5-Pro on OpenRouter to run long-sequence tasks like web Mahjong from an image reference is such a solid cost hack! Great breakdown of setting up Codex++, Computer Use, and Blender MCP to avoid burning through Astra quotas
The trick that makes this 90-minute run actually work without blowing up cost or state drift isn't just snagging the free credits—it's routing the tasks so the heavy visual compute doesn't burn token context. You point Codex CLI to the open-tier Nex-N2.5 Pro endpoint, map the local headless Blender session via a simple Python MCP server, and constrain the visual feedback loop so the model only pulls full-screen screenshots when checking WebGL rendering failures in the browser. By delegating the raw 3D mesh math to Blender script commands and using the model's visual groundings strictly for game-state alignment and UI verification, you preserve your context window across hundreds of autonomous tool calls without hitting rate limits or watching the agent lose the thread mid-build.
Getting a working web game out of a single image is a huge flex for an open weight model.
Free access to a model that can combine long-sequence reasoning, Computer Use, and coding is a great opportunity to experiment. The web Mahjong test is a pretty interesting use case. 🔥
That Codex + Computer Use + Blender MCP pipeline bypasses expensive rendering tokens entirely. Letting the agent self-manage local asset creation and layout generation over a 90-minute autonomous loop without draining your wallet is the ultimate workflow hack. Did you run into any session timeouts or did the loop execute clean all the way to the final build?
The combination of Codex, Computer Use, and Blender opens up a lot of creative workflows. Being able to experiment without constantly worrying about token costs makes it even more accessible. 🚀
Running an hour-and-a-half long-horizon agent task to build a full web Mahjong game from a legend image entirely on a free OpenRouter tier completely changes the economic model of heavy prototyping. Bypassing token burnout on GPT-6 Astra while keeping that level of logic execution is a massive win. How smoothly did Nex-N2.5 Pro handle the complex tile-matching and scoring rule logic during that extended run?
nex agi，next agi，口气很大，拭目以待
他们这名起的 原来有深意🤦♂️
试完了，这免费模型真能扛长任务。丢一张图进去，一个半小时后网页麻将能正常摸打吃碰，规则也没崩，比想象中稳。额度焦虑少了一大半，先收藏这套流程。
干货满分！搞开发原型的核心就是算账。‘用免费模型跑 Blender 脚本，行了再上贵价渲染’这个两阶段路由策略直接戳中了工业级自动化生产的真痛点。麻将这种多逻辑交叉嵌套的网页游戏都能挂机跑出来，说明开源底座加网关生态在长程自主除错（Self-debugging）上的工程上限已经被完全打开了
Combining OpenRouter’s free endpoints with Codex and Blender MCP is a smart setup for heavy visual testing. Using an agentic feedback loop to handle early-stage Blender drafting before running expensive rendering workflows solves a major cost bottleneck for creators.
Best free alternative right now - Nex-AGI rivaling Astra but free on OpenRouter. Codex + Computer Use + Blender MCP workflow is smart: Astra burns quota in 20X Pro, this is unlimited. Mahjong via image recognition in an hour proves it works.
You can currently call the Nex‑N2.5‑Pro API for free on OpenRouter:
Tutorial everyone needed: Install Codex, set OpenRouter to nex-n2.5-pro:free, enable Computer Use, add Blender MCP. Then let it white model → Seedance 2.5 free rerun. Zero token worry vs GPT-6 Astra's quota hell.
这个速度快吗？速度和效果好的话，用 blender 出白膜，seedance 2.5 出视频确实是非常好的工作流。
我操，是真的吗，还能免费用，终于不能看账单爽用了
This is honestly the kind of free AI workflow worth experimenting with. 😮💨 Codex with Computer Use and Blender, and a model that can handle long-context visual tasks? If the free access holds up, I’m definitely trying this.
这套 Nex-AGI + Codex + Computer Use + Blender 的长程识图流程太香了，1.5 小时从一张图到能完整打麻将的网页小游戏，确实把免费模型的上限拉得很高
整体感觉它更像能自己解决问题的助手，而不是只给建议的聊天框，适合挂在日常工作流里多测。
真的可以么？我赶紧去试试
这下不懂 Blender 的人也能先让模型代劳，网页和白模都能摸一遍了
免费白嫖这条路我记下了。麻将网页这种看得见又点得动的活，最能看出 Computer Use 是不是真干活。你流程里最容易翻车的是 Blender 还是网页那段。
仿佛第二个牛来模型
天菩萨，还有这种操作，啥也别说了，马上去用！
that's a smart move, always worth exploring alternatives. curious to hear how nex-agi holds up in practice
The vision of AI directly operating professional software is fascinating. Imagine describing a 3D concept and having the agent build, inspect, revise, and export the model autonomously.
好家伙，这游戏看起来挺好玩！😍
AI agents are getting more powerful every day. The combination of Codex, Computer Use and Blender opens a lot of possibilities.
不错哦 竟然还能免费的用
免费模型能力还强，配合Codex和Computer Use能指挥Blender干活，太爽了。
没有听说过，这么强嘛，我马上去试试
Nex-N2.5-Pro 免费使用 生成LOGO 视频很快
网页版麻将这种长任务都能落地，说明 Nex 在工具调用和持续执行上确实有两把刷子
免费底座 + 自动化工具链才是王道！让 Nex-N2.5 先把白模和长逻辑啃完，再去调高价模型，这套两阶段路由策略直接把工业研发成本打下来了。
