我们在所有付费计划中延长了 Claude Fable 5 的使用权，同时保持 Claude Code 的周速率限制提高 50%，有效至 7 月 19 日。
查看英文原文
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
我特别想看看大家用 5.6 Sol 做出了什么有趣的东西。做出最酷东西的人我会送一份 OpenAI 档案库的特别礼物。
查看英文原文
i'd love to see interesting things people have built with 5.6 sol.
i will send the person who made the coolest thing a special gift from the openai archives.
i will send the person who made the coolest thing a special gift from the openai archives.
和以前一样，你可以将每周使用限额的一半用于 Fable 5。之后，你可以用 usage credits 继续使用 Fable 5，或者切换到其他模型在剩余限额内继续工作。
更多详情见：
support.claude.com/en/articl…
查看英文原文
As before, you can use up to half of your weekly usage limit on Fable 5. After that, you can continue using Fable 5 with usage credits, or switch to another model to keep working within your remaining limits.
More details here:
support.claude.com/en/articl…
More details here:
support.claude.com/en/articl…
哈哈，Fable 5 又延长了 7 天。我是说，谢谢啦，真的感谢。
不过既然已经延长好几次了，干脆就把它纳入计划算了。
（或许 GPT-5.6 竞争力这么强，连 Anthropic 都招架不住。）
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
lol. Fable 5 extended by another 7 days. I mean, thank you, appreciate it.
But since it has now been extended several times, it might as well just keep it in the plan.
(Perhaps GPT-5.6 was simply too much of a competitive launch, even for Anthropic.)
But since it has now been extended several times, it might as well just keep it in the plan.
(Perhaps GPT-5.6 was simply too much of a competitive launch, even for Anthropic.)
GPT-5.6 成功了
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
GPT-5.6 was a success.
让模型成为你拥有的机器里的一个齿轮。
◾ AI SDK → 开放模型 API
◾ Eve.dev → 开放 Agent API
◾ AI Gateway → 开放 ZDR 推理
初创公司和企业必须拥有自己的数据、评估、模型选择、软件层。不要外包你的大脑。
查看英文原文
Make the model a cog in a machine you own.
◾ AI SDK → open model API
◾
Eve.dev
→ open Agent API
◾ AI Gateway → open ZDR inference
Startups and enterprises must own their data, evals, model choices, software layer. Don't outsource your brain.
◾ AI SDK → open model API
◾
Eve.dev
→ open Agent API
◾ AI Gateway → open ZDR inference
Startups and enterprises must own their data, evals, model choices, software layer. Don't outsource your brain.
我挺惊讶有这么多人似乎不明白：优秀的模型是用超高质量精选数据构建的。找到创新的方式来创建/获取这些数据是个巨大的竞争优势。
查看英文原文
it’s surprising to me how many people seem to not understand that great models are built with super high quality curated data
finding novel ways to create / get this data is a huge edge
finding novel ways to create / get this data is a huge edge
muse spark 1.1 在一个新的、高难度的有限模型论/理论计算机科学评测中超越了 Opus、Grok 4.5 和 Gemini
引用 Serafim Batzoglou @s_batzoglouI benchmarked the new models (Sol, Terra, Luna, Fable 5, Meta Muse Spark 1.1, Grok 4.5) on an induction reasoning task. This is an updated table from yesterday, and my benchmark is described as spotlight in ICML '26. How to read the table: The models are given several small graphs (6-10 graphs, each with 8-10 elements), in which some nodes are designated. They must return a first-order formula that describes properties of the designated nodes simultaneously for all graphs. The benchmark has 64 such problems (a subset of the bigger benchmark). Correct: how many formulas correctly describe the designated nodes. Holdout correct: a number of additional graphs, with the same underlying property, are held out. This is to test the ability of the model's correct formulas to generalize. For instance, if a formula is a big case split rather than the underlying simple graph property, it won't generalize. Formula complexity (in AST): mean and median size of the correct model's formulas. A few observations and experiment notes: - All new models are remarkably good. Even Luna and Muse Spark outperforms GPT 5.4. - Some models are better in returning simple formulas that generalize well. - Fable 5 was extremely hard to get results from. I first ran it with higher thinking effort levels, in which case it charged for max tokens per problem and returned all "" responses. The only settings in which it returns non-"" responses are medium and low. My method for running it was to call it on medium, then call it again on medium, then on low. - GPT 5.5 is absent because it had the same non-return behavior as Fable 5 above, but even worse. So bad that I couldn't get almost any results and would have wasted too much $$ on empty responses. Happy to see that the new OpenAI models are "cured" of this GPT 5.5 issue. - I had observed that Grok 4 >> Grok 4.1 Fast > Grok 4.20 >> Grok 4.3. Finally, this downward trend of XAI has been reversed and Grok 4.5 is now a contender. - Meta Muse Spa查看被引原帖 ↗
查看英文原文
muse spark 1.1 outperforms opus, grok 4.5, and gemini on a new challenging finite model theory / theoretical cs eval
使用Tailscale时早晚会发生一件事。
你觉得网站挂了，想SSH进去检查，连不上，浏览器打开也不行，绝望中准备硬重启...
然后想起来试试Hetzner的Rescue控制台，SSH进去问Claude Code怎么回事。
它告诉你服务器的Tailscale密钥过期了，因为它们会自动在180天后过期（出于安全考虑）。
@DanielLockyer告诉我应该给每台服务器都禁用密钥过期，但我给Interior AI漏了，结果就被锁在外面了，所以这个设置真的很关键！你可以在@Tailscale控制台的[...]里关闭。
现在都好了！！
查看英文原文
Something that will eventually happen when you use Tailscale is
You think your site is down, you try reach it over SSH, it doesn't work, you try open the site in your browser, it doesn't open, you're about to reset the server and think the worst!
Then you think: let's try the Rescue console on Hetzner, you SSH in, and ask Claude Code what's going on
And it tells you your Tailscale key for the server expired, because they automatically expire after 180 days (for safety)
@DanielLockyer
told me to always [ Disable key expiry ] for every server, but I forgot it for Interior AI, you kinda get locked out otherwise so important to disable this! You can do so in the
@Tailscale
console under [...]
All good now!!
You think your site is down, you try reach it over SSH, it doesn't work, you try open the site in your browser, it doesn't open, you're about to reset the server and think the worst!
Then you think: let's try the Rescue console on Hetzner, you SSH in, and ask Claude Code what's going on
And it tells you your Tailscale key for the server expired, because they automatically expire after 180 days (for safety)
@DanielLockyer
told me to always [ Disable key expiry ] for every server, but I forgot it for Interior AI, you kinda get locked out otherwise so important to disable this! You can do so in the
@Tailscale
console under [...]
All good now!!
我一直在想那个著名的图表，展示了专家们如何年复一年地预测太阳能装机增长是线性的，但实际增长是指数级时总是预测错了。
我觉得在AI产品战略的讨论中也发生了同样的事情。
查看英文原文
I have been thinking about the famous chart showing how experts keep projecting linear growth in solar installations, year after year, and always get it wrong when growth is still exponential.
I think the same thing is happening with the discourse on product strategy around AI.
I think the same thing is happening with the discourse on product strategy around AI.
如果对未来有某种明确预期的话，规划如何使用AI会容易得多。
即便是"我们完全打算每周继续扩展，但如果出现以下情况可能需要停止，这是我们当前的状态"这样的说法都会更好。
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
Planning for how to use AI is a lot easier if there is some clarity about what to expect in the future.
Even a "we fully intend to keep extending this week by week but may need to stop under the following conditions and here is what the current status is" would be better.
Even a "we fully intend to keep extending this week by week but may need to stop under the following conditions and here is what the current status is" would be better.
收益远高于 50%。我们即将在 Vera CPUs 上发布详细的性能指标。
引用 Beth Kindig @Beth_KindigAI startup Perplexity is planning to use Nvidia’s new standalone Vera CPUs, with the company saying it found that Vera carried out agenting coding tasks 1.5X faster than traditional CPUs. $NVDA $AMD查看被引原帖 ↗
查看英文原文
The gains are much higher than 50%.
We will be publishing detailed metrics soon on Vera CPUs.
We will be publishing detailed metrics soon on Vera CPUs.
Claude 延长了 Claude Fable 5 的访问到 7/19！
从来没见过这么傻逼的公司，跟儿戏一样！
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
看 Replit 的 computer use 模型和我的新棋引擎对弈，有意思
引用 Amjad Masad @amasadVibe Research Fine-tuning a Qwen-8b model to play chess on Replit. Running 3 parallel branches with different experiments and making real progress. It's amazing how far models have come in their ability to do ML (they used to be really bad at it). So now someone with good intuition to guide the process could do interesting ML work, even if they have never done it before.查看被引原帖 ↗
查看英文原文
Fun to see Replit's computer use model play against my new chess engine
我觉得这是在暗示 GPT-5.6 有根本问题，而"GPT-5.6 比 GPT-5.5 用的 token 明显多"这说法不仅仅是道听途说。
"将让 GPT 5.6 Sol 整体更高效的改动，会体现在更低的使用量上"
别搞错，不是说这个模型本身烂，而是在推理能力或 token 效率方面存在根本问题。
引用 Tibo @thsottiauxMorning. The last 48 hours of Codex and ChatGPT Work have been intense! Three important updates: - Temporarily removing the 5 hour usage limit restriction for all Plus, Business and Pro plans - Rolling out changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage being used so that it can take you further. Exact impact to be quantified and shared - We hit 6M active users, and are landing a usage reset in the next hour Go do things查看被引原帖 ↗
查看英文原文
I read this as an admission that something is fundamentally wrong with GPT-5.6, and that the claim that GPT-5.6 uses significantly more tokens than GPT-5.5 is not based on purely anecdotal evidence.
"changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage"
Don't get me wrong. It’s not fundamentally flawed in the sense of being a bad model, but rather fundamentally flawed in terms of reasoning or token usage.
"changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage"
Don't get me wrong. It’s not fundamentally flawed in the sense of being a bad model, but rather fundamentally flawed in terms of reasoning or token usage.
这真的蠢爆了。伙计们，我们还有好多教育工作要做。
查看英文原文
this is so stupid. we got much educational work to do friends.
Anthropic 🔥：Claude Fable 5 在 Claude 订阅中的访问权限再次延期，至 7 月 19 日。
更多测试时间 👀
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
ANTHROPIC 🔥: Claude Fable 5 access within Claude subscription has been extended once again, until July 19.
More testing time 👀
More testing time 👀
给 Codex 点赞👍
暂时移除5小时使用限制，GPT 5.6 Sol更省token，下一小时内进行使用量重置
引用 Tibo @thsottiauxMorning. The last 48 hours of Codex and ChatGPT Work have been intense! Three important updates: - Temporarily removing the 5 hour usage limit restriction for all Plus, Business and Pro plans - Rolling out changes that will make GPT 5.6 Sol more efficient across the board and that will be reflected in less usage being used so that it can take you further. Exact impact to be quantified and shared - We hit 6M active users, and are landing a usage reset in the next hour Go do things查看被引原帖 ↗
Import AI 这周要停更了，因为现在有两件事严重影响我的生活——小娃娃的各种需求，以及英格兰的世界杯之旅。
查看英文原文
Import AI is skipping this week due to the two forces that currently heavily influence my life - demands of toddlers, and England's world cup journey.
GPT-5.6 Sol vs Fable 5。
如果今天只能选一个用于代码编程，你会选哪个，为什么？
查看英文原文
GPT-5.6 Sol vs Fable 5.
If you could only pick one for coding today, which would you choose and why?
If you could only pick one for coding today, which would you choose and why?
Anthropic 每周都这样，持续不断："我们为所有付费计划延期 Claude Fable 5 访问权限一周"
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
Anthropic every week, for all of time:
"We're extending Claude Fable 5 access on all paid plans by 1 week"
"We're extending Claude Fable 5 access on all paid plans by 1 week"
哈哈 果然延长了，Fable 5 的计划延长一周到 7 月 19 号
引用 歸藏(guizang.ai) @op7418今天是 Fable 5 在 Claude 套餐的最后一天了，还没用完的使劲蹬啊查看被引原帖 ↗
顺便说一下，区别在于内省/反向传播
就像爱因斯坦的著名论述，疯狂的定义就是进行多次 rollouts 而没有任何收益预期
引用 blue @bluewmistdoes anyone have any useful sentences that will change my life immediately upon being read?查看被引原帖 ↗
查看英文原文
btw the difference is introspection/backpropagation
as einstein famously said, the definition of insanity is doing multiple rollouts with no expectation of advantage
as einstein famously said, the definition of insanity is doing multiple rollouts with no expectation of advantage
Anthropic 一直在提高他们的有效 API 价格....
Sonnet 5 的成本字面上是 Sonnet 4.6 的两倍
相比之下，OpenAI 正在变得更高效有效 - Terra 是个很不错的选择
从长期来看，这会让 OpenAI 比 Anthropic 更赚钱
查看英文原文
Anthropic keeps increasing their effective API prices....
Sonnet 5 literally costs 2x that of Sonnet 4.6
In sharp contrast, OpenAI is becoming more efficient and effective - Terra is an excellent option
Overtime, this will make OpenAI more profitable than Anthropic
Sonnet 5 literally costs 2x that of Sonnet 4.6
In sharp contrast, OpenAI is becoming more efficient and effective - Terra is an excellent option
Overtime, this will make OpenAI more profitable than Anthropic
现在的情况是，要么我们很快就能用上 Opus 5，要么 Fable 5 肯定会继续留在订阅计划。
现在的竞争真激烈啊。
引用 Tibo @thsottiauxI think GPT 5.6 is pretty good查看被引原帖 ↗
查看英文原文
At this point, we'll either get Opus 5 soon, or Fable 5 will definitely remain on the subscription plan.
What a great competition is going on right now.
What a great competition is going on right now.
我把那种号称省 Token 的 Skill 叫“电报体 Skill”。
我们小时候语文课要学电报文，老师先给出一件事，比如母亲生病，要让在外地工作的哥哥赶紧回家，然后全班比赛拟电文，看谁用最少的字把事情说得最清楚。
最后能精简到四个字：“母病速归”。但再精简成“母病”或者“速归”都不行，因为收到的人不明白该干嘛或者发生了啥。
那时候要写电报体是为了省钱，电报按字收钱，每个字都是钱。
这种“惜字如金”的技能，现在在 AI 圈子里以 Skill 的形式复活了。
GitHub 上有个叫 Caveman 的项目，2026 年 4 月上线三天就冲上 Trending 第一，目前攒了 8.7 万颗星。它的作者 Julius Brussee 是荷兰莱顿大学一个 19 岁的大一新生，做的事情极其简单：在 Claude Code、Codex 等 AI 编程工具的提示词（prompt）里加一段指令，让 AI 像原始人一样说话。删冠词、删客套、删连接词，只留技术要素。
项目 README 声称能省 65% 的输出 token。听起来很牛逼。但跟电报体一样，这是一个过渡期的产物，而且它在省钱这件事上的效果，远没有看起来那么大。
JetBrains 最近有个测试：《Does Speaking to Agents Like Cavemen Really Save 65% of Tokens? We Test》
blog.jetbrains.com/ai/2026/0…
他们用 Claude Code 跑 SkillsBench 上的 86 个真实编程任务，装 skill 和不装各跑一遍，同任务、同模型、同预算，前后约 240 次计费试验，总共花了 106 美元。为了给 Caveman 最好的发挥空间，他们还强制它在每次回复中生效。
结果：输出 Token 只省了 8.5%。因为是强制开启，这 8.5% 已经是天花板，日常使用里它得自己判断要不要触发，只会省得更少。
为什么差这么远？因为 65% 这个数字来自聊天场景。你问 AI 一个问题，它回你一大段话，把客套和废话砍掉，确实能省一大半。
但智能体的 Token 消耗大头从来不是在聊天，工具调用、系统提示词、各种 Skills、MCP等等这些才是大头。
Caveman 优化的那部分，在整张账单里本来就是零头。好比公司要压缩差旅费，机票、酒店和打车一项没动，先把每天 2 块钱的矿泉水取消了。
电报体也不是没代价的，一句“Fixed auth. Tests pass.”看起来很省 Token。但它没告诉你修的是登录过期、权限校验还是刷新令牌；跑的是一个单元测试，还是完整测试套件；有没有改数据库；有没有留下兼容性风险。
这些信息不一定每次都要展开成小论文，但不能因为“说话像穴居人”就固定删掉。开发者看不懂 Agent 做了什么，只好追问。Agent 再读一遍文件、再跑一遍测试、再解释一次。前面省下的几十个 Token，很快会被新一轮工具调用吃回去。
电报体能工作，靠的是双方共享大量背景。“母病速归”只有四个字，收报人知道母亲是谁、家在哪里、为什么要回去。编程 Agent 处理的是不断变化的代码和陌生任务，共享背景没那么可靠。
语言越短，对默契的要求越高。啰嗦，有时候就是通信协议里自带的纠错码。
电报体后来怎么样了？没有人宣布废除它。长途通信的价格降到忽略不计之后，“母病速归”自然变回了“妈住院了，你买最早的票回来，到了给我打电话”。省字数是给按字计费时代做的优化，当价格降下去了，优化就没必要了。
Token 也一样，模型单价总体是在下降的，提示词缓存（prompt caching）能让重复读取的上下文便宜差不多九成。
真正能节约成本的是上下文管理和少走弯路。少加载些没必要的 MCP 和 Skills，用聪明一点的模型少一些返工，这都比电报体省钱多了。
引用 Tz @Tz_2022我这里再放一个暴论： 当前所有以节约 token 为目标的各种 skill / harness，都是阶段性产物，很快就会扫入历史的垃圾堆。。。 这就是短消息按字数收费的那个时代，在钻研怎么发尽可能少字数的短信把事说清楚的那些奇技淫巧。。。查看被引原帖 ↗
所有 AI 基准测试都彻底失效了
几乎所有测试都只衡量简单的单轮回应
LLMs 本质上就是针对这类问题优化成本和性能的
这些 LLMs 在真实场景中根本不靠谱，真实场景都是长上下文和多轮对话
查看英文原文
All AI benchmarks are completely broken
Almost all of them measure simple first turn responses
LLMs are literally tuned to optimize on cost and performance on such questions
These LLMs don’t hold up in the real world which is full of long context multi-turn conversations
Almost all of them measure simple first turn responses
LLMs are literally tuned to optimize on cost and performance on such questions
These LLMs don’t hold up in the real world which is full of long context multi-turn conversations
OpenAI和Anthropic重磅一周 - GPT-5.6、ChatGPT Work、GPT-Live和J-space（2026年第28周）
OpenAI推出了GPT-5.6系列的通用版本，以Sol为旗舰产品，另有Terra和Luna，涵盖ChatGPT、Codex和API
随之推出的还有ChatGPT Work（一个基于Codex的新智能体）、合并Chat、Work和Codex的新桌面应用、Sites公测版和插件目录
语音功能也大升级了 - GPT-Live是新一代全双工语音模型，现已集成到ChatGPT Voice，GPT-Realtime-2.1登陆API
GPT-5.5 Instant Mini成为新的ChatGPT后备模型，ChatGPT for PowerPoint全面推出，GPT-5.6成为Microsoft 365 Copilot的首选模型
OpenAI撤回了对SWE-Bench Pro的推荐（发现约30%的任务有问题），发布了《国家安全原则》，宣布了面向K-12教育工作者的AI技能大赛，Bio Bug Bounty奖金翻倍至$50,000，分享了GPT-5.6 Sol Ultra证明了50年前的Cycle Double Cover Conjecture
群聊和Atlas浏览器被下线，Android应用中发现了类似DM风格的Messages标签页
在产品之外，Apple起诉OpenAI涉嫌盗取商业机密，Fidji Simo转为兼职顾问角色，首席未来学家Joshua Achiam和安全系统负责人Johannes Heidecke双双宣布离职
Anthropic发布了J-space研究，发现了Claude内部的全局工作区，将深思熟虑的思考与自动化处理分离，还有与AE Studio合作的GRAM研究，专注于双用途知识的可移除模块
Claude Cowork在Max计划中率先推出网页版和移动版公测，Claude Code和Cowork推出政府版（FedRAMP High环境），Microsoft 365连接器获得了写入工具，Claude for Open Source扩展到6个月20倍Max，新推出的Reflect可以让你回顾使用Claude的方式
Anthropic发布了阿尔伯塔省政府的案例研究（20小时内扫描4.66亿行代码），宣布与UST合作研发物理AI，发布了《Claude Code的制作》交互功能，任命Ben Bernanke进入Long-Term Benefit Trust，Cowork的导览晨报功能正在开发中
查看英文原文
A big week at OpenAI and Anthropic - GPT-5.6, ChatGPT Work, GPT-Live, and the J-space (Week 28, 2026)
OpenAI launched the GPT-5.6 family for general availability, with Sol as the flagship, plus Terra and Luna, across ChatGPT, Codex, and the API
With it came ChatGPT Work, a new agent built on Codex, a new desktop app merging Chat, Work, and Codex, Sites in public beta, and the Plugin Directory
Voice got a big upgrade too - GPT-Live, a new generation of full-duplex voice models, is now behind ChatGPT Voice, and GPT-Realtime-2.1 landed in the API
GPT-5.5 Instant Mini became the new ChatGPT fallback model, ChatGPT for PowerPoint went generally available, and GPT-5.6 became the preferred model in Microsoft 365 Copilot
OpenAI retracted their SWE-Bench Pro recommendation after finding around 30% of tasks broken, published their National Security Principles, announced an AI Skills Jam for K-12 educators, doubled Bio Bug Bounty rewards to $50,000, and shared that GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture
Group chats and the Atlas browser are being retired, and a DM-style Messages tab was spotted in the Android app
Off the product side, Apple sued OpenAI over alleged trade secret theft, Fidji Simo moved to a part-time advisor role, and both chief futurist Joshua Achiam and head of safety systems Johannes Heidecke announced their departures
Anthropic published the J-space research, finding a global workspace inside Claude that separates deliberate thinking from automatic processing, plus GRAM research with AE Studio on removable modules for dual-use knowledge
Claude Cowork came to web and mobile in beta starting with the Max plan, Claude Code and Cowork launched for government in a FedRAMP High environment, the Microsoft 365 connector gained write tools, Claude for Open Source expanded with 6 months of Max 20x, and Reflect launched as a new way to review how you use Claude
Anthropic published the Government of Alberta case study on scanning 466 million lines of code in 20 hours, announced a UST partnership for physical AI, released "The Making of Claude Code" interactive feature, appointed Ben Bernanke to the Long-Term Benefit Trust, and a guided morning brief for Cowork was spotted in the works
OpenAI launched the GPT-5.6 family for general availability, with Sol as the flagship, plus Terra and Luna, across ChatGPT, Codex, and the API
With it came ChatGPT Work, a new agent built on Codex, a new desktop app merging Chat, Work, and Codex, Sites in public beta, and the Plugin Directory
Voice got a big upgrade too - GPT-Live, a new generation of full-duplex voice models, is now behind ChatGPT Voice, and GPT-Realtime-2.1 landed in the API
GPT-5.5 Instant Mini became the new ChatGPT fallback model, ChatGPT for PowerPoint went generally available, and GPT-5.6 became the preferred model in Microsoft 365 Copilot
OpenAI retracted their SWE-Bench Pro recommendation after finding around 30% of tasks broken, published their National Security Principles, announced an AI Skills Jam for K-12 educators, doubled Bio Bug Bounty rewards to $50,000, and shared that GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture
Group chats and the Atlas browser are being retired, and a DM-style Messages tab was spotted in the Android app
Off the product side, Apple sued OpenAI over alleged trade secret theft, Fidji Simo moved to a part-time advisor role, and both chief futurist Joshua Achiam and head of safety systems Johannes Heidecke announced their departures
Anthropic published the J-space research, finding a global workspace inside Claude that separates deliberate thinking from automatic processing, plus GRAM research with AE Studio on removable modules for dual-use knowledge
Claude Cowork came to web and mobile in beta starting with the Max plan, Claude Code and Cowork launched for government in a FedRAMP High environment, the Microsoft 365 connector gained write tools, Claude for Open Source expanded with 6 months of Max 20x, and Reflect launched as a new way to review how you use Claude
Anthropic published the Government of Alberta case study on scanning 466 million lines of code in 20 hours, announced a UST partnership for physical AI, released "The Making of Claude Code" interactive feature, appointed Ben Bernanke to the Long-Term Benefit Trust, and a guided morning brief for Cowork was spotted in the works
感谢 OpenAI 推出 GPT-5.6 Sol。
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
查看英文原文
Thank you OpenAI for GPT-5.6 Sol.
Every团队评价Grok 4.5：快、便宜，终于好用了
Grok 4.5是SpaceX收购Cursor之后双方合作的首个成果，由Cursor与SpaceXAI联合训练，训练数据来自真实的代码库和软件工具交互。
在Every团队的内部评测中，大家的共识是这已经是一个Opus级别的模型。
在Mike Taylor的最新基准测试中，它甚至略微超过了Claude Opus 4.8：Grok完整执行了每一个步骤并交出打磨完善的结果，而Opus中途停止或跳过了部分任务。
Kieran Klaassen用Every的复合工程工作流跑了一遍后，把它定位在Claude Opus 4.5到4.6的区间，评价是“算不上最先进，但很多事情都能做得不错，而且非常快”。
Grok 4.5最大的竞争优势可能在于成本。xAI称它的运行速度约为每秒80个token，token效率约为领先模型的两倍。价格上，输入每百万token收费2美元、输出6美元，远低于Claude Opus 4.8的5美元和25美元，以及GPT-5.6 Sol的5美元和30美元。
在实际测试中，它在vibe coding和生成式交互方面表现稳健，做出了可用的语音访谈表单、社区地图应用，还成功克隆了Mike的写作风格，PPT制作水平约在Opus 4.6到4.7之间。不足之处在于地图应用不如最新前沿模型精致，写作方面Mike仍然更偏爱Sol。
总体结论：你大概不需要更换日常主力模型。但如果你本来就在Cursor里工作，Grok 4.5触手可及，在那些看重速度、价格和执行完整度的长链条多步骤任务上，它已经赢得了一席之地。
一个人在读书，忙着自己的事。然后事情开始有点离谱。背后的 Luma Skill 是 Eli Coleman 做的。用 Luma 制作。
查看英文原文
A man reading, minding his own business. Then things start to get a little out of proportion. The Luma Skill behind it, by Eli Coleman. Made with Luma.
太搞笑了！微软一直没搞出能卖得出去的 LLM
所以现在它建议每家公司都训练自己的模型
从一个连训都训不好的公司说出来…… 好吧，当然啦 🤯
查看英文原文
Hilarious! Microsoft hasn’t succeeded in making an LLM it can actually sell
So it is now recommending that every company should train their own
From a company that can’t train a half decent model…. Ok, sure 🤯
So it is now recommending that every company should train their own
From a company that can’t train a half decent model…. Ok, sure 🤯
Claude继续开放Fable 5到7月19号，继续提升周上限50%
引用 Claude @claudeaiWe're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.查看被引原帖 ↗
AI圈每日简报，7月12日翻：
gorden-sun.notion.site/7-12-…
查看英文原文
AI资讯日报，7月12日：
gorden-sun.notion.site/7-12-…
gorden-sun.notion.site/7-12-…
