Paras

Perpetual Builder. Prev: Amazon building Ads. Qualcomm building CPUs.

[Paras @mumtor99]

The only difference between Max and Ultra Mode is that Ultra Mode will launch a lot of sub-agents because that's what the Ultra Mode system prompt tells it to. If you want to launch fewer sub-agents, don't use Ultra Mode. Use either Extra High or Max and tell it to use sub-agents sparingly.

[Paras @mumtor99]

@jturntdev You can turn on auto review or run it in a sandbox with full permission. That way, you will not run into this issue. Auto review should work fine on its own. Maybe your harness config is set to user-based review, and that's what's causing this.

[Paras @mumtor99]

@reach_vb Ultra mode sub agents are also using ultra reasoning. Very wasteful TBH. No way to configure which models to use as sub agents. See

https://t.co/5OrBvSnj9R

[Paras @mumtor99]

Something odd is going on with Codex CLI and the GPT-5.6 Sol model when it comes to using other GPT models as sub-agents.

It appears GPT-5.6 Sol defaults to multi_agent_version=v2, and in v2 the sub-agent API no longer exposes the model and reasoning parameters. As a result, if I'm running GPT-5.6 Sol Ultra as the primary model, every sub-agent also appears to run as GPT-5.6 Sol Ultra.

In multi_agent_version=v1, I had a great workflow where I could explicitly launch sub-agents with different models and reasoning levels. That seems to be broken out of the box with GPT-5.6 Sol in both Codex CLI and the ChatGPT app.

This is easy to verify. Ask GPT-5.6 Sol to launch GPT-5.6 Terra or GPT-5.5 as a sub-agent, and it will report that it can't say for sure which model was actually launched. If you then inspect the Codex CLI codebase, you'll find that the sub-agent inherits the same model as the primary agent rather than the model that was requested.

Unless I'm missing something, this effectively removes the ability to build heterogeneous multi-agent workflows—for example, using GPT-5.6 Sol Ultra as the orchestrator while delegating repository exploration or routine tasks to cheaper or faster models.

[Paras @mumtor99]

@pauloportella_ @pvncher this is a real issue. can't use subagents of different models and reasoning at all. i am using this weird hack to start a new thread in gpt 5.5 to lock in v1 multi_agent_version and then switch to 5.6 to launch subagents with specific model and reasoning effort

[Paras @mumtor99]

5.6 defaults to v2 sub agents and v2 doesn’t honor the model param or named-agent profiles. I am using a workaround that seems to work but i hate it. start with 5.5, send one turn to lock v1, then switch to 5.6.

@pvncher this sucks as i can't use terra sub agents with sol without starting each cli thread with this workaround. same issue in chatgpt codex app as well.

[Paras @mumtor99]

@thsottiaux help!

[Paras @mumtor99]

Something odd is going on with Codex CLI and the GPT-5.6 Sol model when it comes to using other GPT models as sub-agents.

It appears GPT-5.6 Sol defaults to multi_agent_version=v2, and in v2 the sub-agent API no longer exposes the model and reasoning parameters. As a result, if I'm running GPT-5.6 Sol Ultra as the primary model, every sub-agent also appears to run as GPT-5.6 Sol Ultra.

In multi_agent_version=v1, I had a great workflow where I could explicitly launch sub-agents with different models and reasoning levels. That seems to be broken out of the box with GPT-5.6 Sol in both Codex CLI and the ChatGPT app.

This is easy to verify. Ask GPT-5.6 Sol to launch GPT-5.6 Terra or GPT-5.5 as a sub-agent, and it will report that it can't say for sure which model was actually launched. If you then inspect the Codex CLI codebase, you'll find that the sub-agent inherits the same model as the primary agent rather than the model that was requested.

Unless I'm missing something, this effectively removes the ability to build heterogeneous multi-agent workflows—for example, using GPT-5.6 Sol Ultra as the orchestrator while delegating repository exploration or routine tasks to cheaper or faster models.

[Paras @mumtor99]

@thsottiaux you are my SOLshine my only SOLshine

[Paras @mumtor99]

Anthropic's model pricing does not make any more sense after Grok 4.5 release. Anyone who is paying for Sonnet or Opus on API is a fucking idiot and should realize it sooner rather than later.

[Paras @mumtor99]

@pvncher "hey, check all the work fable did and find obvious issues that it missed"

[Paras @mumtor99]

@sqs Factory ai subscription is a decent model to replicate

[Paras @mumtor99]

@ns123abc It's a strategic thing to say that GPT-5.6 is not as smart as Fable, but it's really, really good. You should think hard about why everyone at OpenAI is going to do this.

[Paras @mumtor99]

We will all hear a lot today that @OpenAI 5.6 is great, but it's not as smart as Fable. I believe this is to make sure that our access to 5.6 does not get further gated by the govt. I think it's a very smart strategy, and I appreciate the prudence. Very very excited about the launch!

[Paras @mumtor99]

@pvncher one of my daily threads is 80 days old since 5.4 launch. to be fair, i do have to remind it to refer to repo docs once in a while mid turn when i realize it forgot a small detail.

[Paras @mumtor99]

@yacineMTB ah fuck! i thought they fixed it. so it's same old Opus just a BIGGER lier.

[Paras @mumtor99]

@intellectronica Soon Sol Shepard

[Paras @mumtor99]

@badlogicgames When will you learn from your mistakes? Same applies to me.

[Paras @mumtor99]

@alexandr_wang It would be great if this model is accurate in long horizon tasks like GPT 5.5. It hard to spend time on anything else unless it can compete with GPT 5.5 on price and overall performance.

[Paras @mumtor99]

@kunchenguid @trq212 Great question. But I am sure you will never receive a reply to this from them.

[Paras @mumtor99]

@theo I took a call to stay away from Fable from the very first release and make do with GPT 5.5 XHigh. I am happy.