@iruletheworldmo It's not about memory and session. effort Extra High = xhigh, juice = 128 is the initial and *correct* value.
English
NeterOster
77 posts
@blingdivinity In fact, I earlier tested this on GPT-5.4: injecting Juice via the developer prompt still produced statistically significantly fewer reasoning tokens than the effort parameter—though that only speaks for 5.4.
English
@NeterOster you are dunning-krugering so hard right now.
1. i know this
2. its irrelevant; the instruction injection works anyway.
with the juice instruction the model consistently reasons for longer.
English
@blingdivinity Yes, it may well think longer; in fact, simply asking the model to "think carefully" also yields a statistically significant increase in thinking time. But that doesn't prove it matches the Juice in the real system prompt, unless you can back it up with data.
English
@StevanBoljevic @Lentils80 It's just a unitless value in the system prompt. OpenAI likely used a juice-aware approach in training to make Juice softly correlated with the actual reasoning effort.
English
@Lentils80 But is the unit here tokens or arbitrary units?
English
🚨 GPT-5.6 Sol's juice values (thinking budgets) have been severely degraded compared to release day
If Sol now feels faster and more "efficient", this is probably why
Terra and Luna juice values aren't affected, so their thinking budgets are now higher than Sol's
English
佬友neteroster发现GPT-5.6-sol的Juice值除了Max整体下调了一档
大概Tibo移除5小时使用限制，只是为了拉用户量，甚至官方教学CLIProxyAPI配置OAI接口配置ClaudeCode（claudex）
然而现在都直接明着降智，搞得大家不得不开Max了
Tibo@thsottiaux
Morning. 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
中文
补充修正：OpenAI Codex DX（@pvncher）澄清，5.6 默认扩大后的上下文不会因为超过 272K 就按更高费率计费；只有手动把窗口再往上扩才另说。
不过目前 API 文档里仍写着 >272K 整单加价，和 Codex 产品侧说法尚未完全对齐。
出处：
x.com/pvncher/status…
eric provencher@pvncher
This is misinformation. Your tokens are not being billed at a higher rate over 272k tokens with the new expanded context window with the 5.6 family of models. If you manually expand the context window beyond that however, that’s a different story.
中文
GPT-5.6-Sol 消耗倍增的真相找到了。
建议在 ~/.codex/config.toml 加上两行：
model_context_window = 272000
model_auto_compact_token_limit = 240000
能让你的 Codex 额度耐用 1.5-2 倍。
Vincent@Vincent_AINotes
终于知道 Codex 5.6 跑长任务时，额度为什么可能掉得特别快了。 GPT-5.6 Sol 基础价没涨。 但输入一旦超过 272K tokens， 整次请求都会按更高价格计算。 同样一次长任务，成本可能达到 GPT-5.5 的 2.34 倍。👇
中文
@soniox_ai Great! I like soniox STT models and it's really great that you frequently updating it to make it better (v5). In fact I do believe you are making the best STT models in the world (keep it up!) but not getting enough attention.
English
Voice AI cannot be English-first.
ATLO, a startup in South Korea, is building AI companion apps, robots, meeting assistants, and smart home devices.
They use Soniox extensively for real-world voice AI.
Korean is not an edge case. Every language matters.
The future of AI is global, multilingual, real-time, and voice-first.
Thank you, Sunghyun Park and ATLO.
English
Oh my god it scored worse than Composer 2! Not even 2.5! And it cost 4x more to run!!!
This might be the worst major lab model drop of all time. Llama 4 tier. Insane.
Michael Truell@mntruell
Gemini Flash 3.5 is now on CursorBench, our main coding agent eval. We’ll keep updating the leaderboard as new models come out. cursor.com/evals
English
@yvbbrjdr 有没有一种可能openai api的tpm是算输入的... 每个请求200k的话tps不需要非常夸张就能打出limit
中文
让我们做一个数学题：
目前openrouter上的snapshot显示gpt-5.5是39tps，4千万token每分钟即666666tps相当于同时开着17094个连接。
我都不知道怎么评论这个事情🫠
Y11@seclink
周日晚间，OpenAI 开发者 Peter Steinberger（即广为流传的 OpenClaw 框架背后的“ClawFather”）分享了一张截图。 截图中显示，他在测试自己的 ClawSweeper 项目时，其 GPT-5.5 数据流的用量已触及第五层级（Tier 5）的上限——即每分钟 4000 万个 Token。 OpenAI 首席执行官 Altman 随即介入并承诺会予以修复，团队也立即重置了他的用量限制。 此次互动不仅凸显了代理式 AI 工作流对计算资源所产生的巨大需求，引发了社区的惊叹，同时也促使计算机科学先驱 Grady Booch 发出警示，提醒人们警惕资源浪费问题。
中文
@Conbenksl @fkysly 而且这玩意反代起来根本没难度，简单用甚至不用什么sub2api，请求地址就是 chatgpt.com/backend-api/co… apikey就是账号的access key，从codex auth.json复制出来，直接调用就行了
中文
@Conbenksl @fkysly 并不会，这甚至是被鼓励的，以前2api还验系统提示词，后来openclaw那事过后直接不验了，现在逆出来就是和纯血API完全一样的，各种特性都有，当时oai的意思是不像A\一样会那么小气不让在其他地方用
中文
@tison1096 其实缺乏明确的量化指标也未必就不能做，dynamic rubrics RL 就是一条前景方向，目前发展也比较快（当然了其实目前SOTA LLM肯定用了类似技术，不然不可能到现在的水平），未来也未必不会有更好的方法. 我之前也对LLM能不能真正创新持悲观态度，但深入了解了一些工作+回顾历史之后，我想说还是保持敬畏
中文
能理解到目前这种 LLM AI 根子上是概率模型，其实很多 AI 能力的边界是可以推理出来的。
语料大力出奇迹可以做很多突破，甚至纯靠拟合做到中等偏上的水平。但是一旦离开标准路线就会出问题，类似之前 AlphaGo 面对神之一手崩盘。
我之前提到过，不同于围棋有明确的优化目标，软工缺乏明确的量化指标，AI 并不能彻底解决这里的问题。
Rex Pano. -My Kemono My Stars-@RexPanotras
@im_magneto 这种神化AI的言论真的有点烦人了。这么说吧，AI在软件可维护性、软件安全、架构设计、工程纪律这些问题上还是极不擅长的，而且AI当前的基于概率统计的范式就决定了AI很难克服这些问题
中文
@plantegg @fengsir16 系统提示词污染+输出巨短就截断，能一样就怪了。下面是做中转的评价：综合能力: Max~= bedrock>AG>cursor>>kiro
说真的，用 kiro 的简直不如用 gpt xhigh, 另外发现了一个或许比较有前途的逆向，估计实际效果可以和反代后的 AG 相比
中文
很多人说 Claude code 封号太残暴了，我一直的建议是不要把时间花在这种价值不高的对抗上，但是 anthropic 的模型确实是当前最强的，所以该怎么办呢？
1）购买 aws 的 kiro，里面包含所有 anthropic 的模型，价格大概是 cc 的 10%，直接用 kiro (IDE or cli)
2) 将 kiro 里面的模型代理出来接到 Claude code 里面：github.com/justlovemaki/A…
AIClient-2-API：简单说就是"白嫖"——把原本只能在特定客户端（如 Kiro IDE、Gemini CLI）内使用的免费大模型额度，转换成标准OpenAI API 格式，让 Cherry-Studio、NextChat、Cline 等任意工具都能调用 kiro Opus 4.5、Gemini 3.0 Pro 等高级模型
玩得开心，大好的 AI 时代，不要去浪费精力折腾反封号
价钱只有 10% 只是我的各用了几个月的体感
中文
@zhaoxiongding @recatm spark has ~200Tops bf16/fp16： pastebin.com/CdSAiGzx
for fine-tuning / training workload this is more important than memory bandwidth (also with better software support) and it is more scalable
English
