Murphy Zhai

Having fun building with AI

[Murphy Zhai @murphy_zhai]

2022 startups:
Idea is cheap, show me the code.

2026 startups:
Code is cheap, show me the idea.

[Murphy Zhai @murphy_zhai]

@jarvis11x HJ5OML

[Murphy Zhai @murphy_zhai]

@OpenAI just dropped the GPT-6 lineup—and the missing of Terra was predictable.

After GPT-5.6 slashed Luna prices and Luna Max outperformed almost every Terra tier (exc Max), Terra got squeezed out. For daily workflows, Sol and Luna cover everything. https://t.co/ZVn70ttOTw

[OpenAI @OpenAI]

Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.

GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.

We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.

[OpenAI @OpenAI]

Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.

GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.

We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.

[Murphy Zhai @murphy_zhai]

互联网时代的四方模型：用户付出注意力，平台将用户注意力的一部分分发给商家或内容创造者，一部分卖给广告主变现。

现在是五方模型：个人Agent把注意力直接砍掉了大部分，并获得了注意力入口，未来广告主会转向agent。这些utility平台确实需要注意力经济之外的盈利模式了，可能是加价然后分一些给Muse。

[Murphy Zhai @murphy_zhai]

It beats Opus4.6(Max) in agentic coding, and can run locally in a 5090 GPU. Amazing! Code red for Anthropic and OpenAI.

[LM Studio @lmstudio]

Qwen3.8-27B is here! 🚀

It's a leap in capabilities for a laptop size model.

Requires ~17GB to run locally.

Model page: https://t.co/GOxQEOLrdK

[Max For AI @MaxForAI]

刚刚，OpenAI收入曝光，企业业务首次超过消费者，年化收入已达400亿美元

据CNBC等媒体披露， @OpenAI 首席财务官 Sarah Friar 在最新投资者会议上表示，OpenAI企业业务收入目前已经超过消费者业务，成为公司最大的收入来源。

今年年初，OpenAI的收入结构还是大约：

消费者业务60%
企业业务40%

当时ChatGPT Plus、Pro等个人订阅仍然占据主导。

但仅仅半年多后，两条曲线已经完成反转，而且比OpenAI此前预计的时间更早。

与此同时，OpenAI最新年化收入run rate已经达到400亿美元。

相比2025年底超过200亿美元的水平接近翻倍。

仅7月份，公司整体收入run rate就环比增长约20%，其中企业业务增长尤其明显。

更关键的是，这并不意味着ChatGPT消费者业务不增长。

OpenAI目前已经拥有超过10亿周活跃用户，同时服务超过200万家企业。

真正变化的是：企业端的增长速度已经开始超过消费者端。

这也解释了OpenAI过去半年越来越明显的产品方向。

Codex、Coding Agent、企业知识库、数据连接器、工作流自动化，以及各种面向公司的Agent能力，都在快速成为OpenAI新的商业化重点。

OpenAI自己的企业数据也显示，截至今年6月，Codex已经占企业客户Codex与ChatGPT合计输出Token的64%。

换句话说，企业客户正在从「让员工用ChatGPT聊天」，逐渐转向「让AI直接参与写代码、处理数据和执行工作」。

所以400亿美元这个数字虽然夸张，但真正值得关注的是背后的收入结构变化：

ChatGPT给OpenAI带来了全球最大的消费者AI入口之一，但真正开始承担下一阶段收入增长的，正在变成企业AI。

AI行业上一阶段争的是模型能力和用户规模。

下一阶段，可能要开始争谁能真正进入企业预算表了。

[Murphy Zhai @murphy_zhai]

14x ultrafast vs. 1.5x fast mode? My 20x plan will drain out in an hour. Not too bad if you would do a daily reset 😏

[Tibo @thsottiaux]

Sometimes you have have to go /ultrafast.

[Naval @naval]

People who are serious about software train their own models.

[Jimmy Jia @JimmyJia_]

@silsrc 强模型时代 harness 的正确形态是只需要告诉模型不要干什么,而不是教它干什么。前者是模型的权限系统, 后者不过是给旧模型打的续命补丁,每次发新模型就贬值一轮。现在凉掉的产品基本都是把补丁当成了护城河

[Murphy Zhai @murphy_zhai]

The resets helped binge coders but apparently hurt those who had a plan

[Degenerate Tweeter ⚡️ @cryptoScrutini]

@thsottiaux Please, no more resets. If I have 20% left Friday and reset Monday, I can use it over the weekend and be back at 100% Monday. Resetting Friday wastes that 20% in the new period and delays my next reset by a full cycle

[Murphy Zhai @murphy_zhai]

🤔

[Dan @DanDr1s]

OpenAI is reportedly running its biggest pre-training run ever.

And no, it’s not GPT-6.

The model is internally codenamed “Doug” and is expected around the end of the year.

GPT-6 should already be a major step forward.

But Doug is supposedly on another level entirely, to the point that Fable could look “primitive” by comparison.

[Murphy Zhai @murphy_zhai]

If you keep hitting ChatGPT / Claude limits, those 7 tips may help you:
1. One-off question → Chat; ongoing work with files/artifacts → coding agents or Cowork
2. New task, or coming back later → new session
3. Don’t keep switching models & effort levels mid-session
4. After a long idle, assume the cache is cold and your context may need to be loaded again
5. Use frontier model to plan; use cheaper models to execute as sub-agents
6. Remember, ChatGPT Pro users now have "unlimited" GPT-5.6 Sol
7. Use /status, /usage, or /context to check token usage pattern regularly

[Tibo @thsottiaux]

That's right, GPT-5.6 Sol is awesome and can be used pretty much anywhere, including in the CC harness.

To celebrate this, together with the fact that I'm not going anywhere... I have reset usage limits for all paid users of ChatGPT Work and Codex.

Have fun out there!

[Murphy Zhai @murphy_zhai]

@thsottiaux Is Luna excluded from subagents? My Sol keeps telling me Luna is not registered as a model for subagents. Is it only me?

[Murphy Zhai @murphy_zhai]

This aged pretty well.

Not exactly “fire the management team,” but Google just moved Hassabis out of day-to-day management, put Koray in charge, and lost Jeff Dean along the way.

Feels like Google finally decided the AI org needs to ship differently.

[Naval @naval]

We’ll know Google is serious about AI when they fire the current management team.

[Naval @naval]

We’ll know Google is serious about AI when they fire the current management team.