JustForFUNFUNFUN
639 posts
如果把模型当成人来看的话，5.6 sol ultra的表现很可能第一次够到了及格线。5.5时不时就会让我拍大腿，更早以前的都是烂泥扶不上墙。所以早期vibe coding就很烦躁，仿佛跟一个弱智在说话。
sol ultra 看起来差不多稳定有我大一的水平了，就是不断在学习best practice 但是还不太会主动用的时候🤪
中文
Our Single-rollout Asynchronous Optimization (SAO), is able to train stably for one thousand steps and consistently outperform GRPO and its variants on agentic coding and reasoning benchmarks,
such as SWE-Bench Verified, BeyondAIME, and IMOAnswerBench. arxiv.org/pdf/2607.07508
English
旱的旱死，涝的涝死
有 V2EX 网友跑完 GPT-5.6 与 Claude Fable5 的 20X 周限生成了一个网站
简直辣眼睛🤣
AI 要是有意识，怕是要拒绝这个任务
地址我放评论区，慎点！
中文
Tips when moving to 5.6
-Clean up old instructions. 5.6 follows them carefully, including undesirable ones buried in old skills. We had to clean many of those
-Avoid Ultra if cost-constrained. Multi-agent improves evals+latency but costs much more. Or ask @thsottiaux for resets
English
@thsottiaux @theo Why did this thing need to be clarified? this is simple math and 1+1 question in LLM
English
(1) is not correct, it is not due to 2X charging after 272k context, we don't charge for longer context on the subscription for GPT 5.6 Sol as we control all the settings. It is due to something else, which I will attempt to explain below.
The overall trajectory length, which is the total some of all context windows across compactions, changes little based on the reasoning effort. Similarly the quality of the overall output is similar across context lengths above 272k. The benefits of higher context lengths are mostly overall speed (as you don't wait for compaction), ability to deal with humongously large input and potentially cost if the system is well tuned and you hit your cache perfectly.
The actual reason is the what you can see depicted in the chart below, which is the difference in the orange line and the blue line. It is caused by overall cost of cache reads going up with the size of the context being shuffled back and forth between toolcalls. The sweet spot in terms of cost is therefore not necessarily to use the maximum possible context length.
What we're working on is tuning the system differently so that we can go back higher without it resulting in higher usage being charged. Hope this clarifies a bit.
English
There's a few issues with Codex that stacked really badly, causing people's usage to get nuked. Crazy how well they all lined up to be as rough as possible
Issue 1 (shown here): gpt-5.6 costs 2x more after 272k tokens. Codex upped their limit to 372k tokens, meaning long threads were billed 2x higher
Issue 2: "Ultra" subagents were also spawned with Ultra, causing insane nested subagent usage that burned a TON of tokens. Apparently intended, but (hopefully) being toned down in system prompt
Issue 3: Sol and Terra use the "v2" subagent layer in Codex that is still early/unfinished/disabled by default. v1 spawned subagents with a fresh history, v2 copies the entire long context.
When combined, you end up with:
1. A ton of reasoning tokens filling context windows, triggering 2x billing
2. Subagents spawning with that filled context window, instantly billed at 2x
3. Ultra spawning too many of these high reasoning full-window subagents
If you had fast mode on when this happened, you got hit with an ADDITIONAL 2.5x
No wonder we were burning so hard...
Tibo@thsottiaux
Updates for Codex and ChatGPT Work users. No nerfing, only good stuff! - We have landed inference optimizations and are passing down savings to all the subscriptions for GPT-5.6 Sol. That should result in around 10% more usage on its own. - We noticed that by changing the context size limit in the product to 372k for GPT-5.6 Sol, up from 272k for GPT-5.5, it resulted in more usage being charged than intended. We have reverted to 272k and will work to roll back out to 372k in the days to come. You should notice that usage drains significantly less after this change. - To understand where the extra usage was coming from, we ran some experiments where reasoning efforts were changed (referred to as juice values under the hood) and have reverted this. - There is slightly more usage of multi-agent than intended in high and xhigh reasoning effort, we are fixing this going forward. Also fixing a small other thing we noticed with auto-review where we can be more efficient. And we continue to have the 5h limit temporarily not apply. Enjoy the rest of the weekend!
English
GPT-5.6 Sol is already very fast
but don't forget, "Sol" is expected to migrate to Cerebras later this month, which could bring all subscription tiers up to 100 tokens/s and allow it to dominate on both speed and price
cerebras makes it cheaper, not more expensive
English
@mattpocockuk could you pls remake some of your skills like code review, tdd, to specs in this way? it is super helpful
English
Reworded /grill-me so it can now truly be used anywhere - not just engineering.
"design tree" -> "decision tree"
"every aspect of this plan" -> "every aspect of this"
"codebase" -> "environment"
"Do not enact the plan" -> "Do not act on it"
github.com/mattpocock/ski…
English
这就是我一直以来的看法，所以你看我根本不依赖什么复杂的 skill，就简单的代码风格…说白了还是 AI 不够强大…
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。。。
中文
Someone DM'd me saying they're using /grill-me in technical interviews
Watch the candidate work. See how they answer questions and push back. See whether the AI drives them, or they drive the AI.
Honestly genius
English
A correction after robust testing.
Replace GPT 5.6 Terra (High) with GPT 5.6 Luna (Extra High) for execution.
Same performance but:
> 1.3x faster than Terra
> 2.5x cheaper than Terra
In Codex, you burn 2.5x less limits if you use GPT 5.6 Luna (Extra High) as subagents.
CJ Zafir@cjzafir
English
@thsottiaux @mylifcc 但是我调整了以后，额度消耗速度确实明显降下来了。在相同的任务负载水平之下，5小时额度窗口明显不会不够用了，此前我的5小时额度在3个小时内就会用光。
中文
