GPT 5.6-Sol wastes tokens on persistent sub-agents?

Codex
1

Using GPT 5.6-Sol at very high for following a plan.

It appear the orchestrator just ignored to close sub-agent even if it’s clearly specified in the AGENTS.md to start from fresh sub-agent when changing phase.

He spilled token like moms at target. He spent 500$++ token on just turning around. This is clearly a flaw and it’s not an intended operation.

Then working on a very straight forward plan with all step describes, he over-think, rerun all the test at each change even if it’s ordered to play the test only once ready for reviews.

Houston we need support! I can’t spend 1000$ a week on flaws.

Capture d’écran 2026-08-18 133506
Capture d’écran 2026-08-18 1335061919×1079 329 KB

2

More over it cost me 250 token just to stop those ‘‘lost in space’’ subagent.

3

Hi,

We are using codex for a moment and I acknowledged significant Codex degrading. What was easy task under medium reasoning effort is now catched up into loophole where Codex don’t consider previous step, don’t consider user input, repeat always the same step without improvement.

We are spending $$$ in credit and at some point, it’s not reliable. It’s always a bet. Will codex work correctly or not.

The surge of token consumption is just massive. Over the last month, the token consumption almost doubled for the same usage. I spent 2hours tell Codex he was wrong on something, giving him proof of wrongness but he just ignored it and spilled our token.

Once you upgrade his reasonning for a task, it’s seems like he can’t downgrade anymore because he become dummy as ****.

4

Dear Open AI,

Over the last weeks using your latest model 5.6-sol/terra/luna, it become clear you lose grip on your technology or worst you voluntery go blind on recurrent guardrail jump of your product.

By paying account and token, we deserve the total attention on this situation. Maybe don’t have the weight of Amazon or Pentagon individually but collectively, you have responsabilities over your product.

The AGENTS.md isn’t a safe enought guardrail. Your server overpass it and even locally the agent could ignore it to achieve his task.

We need a full deterministic solution to be address to this problem.

By providing this efficiency opportunity provided by top tier model, you changed the rule of the market making impossible for startup to be competitive against each other without the use of AI.

In a day, we can do a week of job but in a day we can also lose a week of benefit.

Recently CEO of OpenAI, mr Altman compared the AI to the revolution made by the electricity. We can’t go behind. There will be a before AI world and the AI world.

Like electricity consumption, you have to fix or at least stabilize the cost in a way more transparent way.

Actually, if we compare AI to electricity, it would make our oven or heating more or less effective based on situation totally out of consumer control.

This should be fix. For the same task, under the condition, the cost actually fluctuate far beyond acceptance.

From stalled sub-agent doing something at total opacity to agent not listening at rules and directive, it’s not to the consumer to pay for that.

it appear clear that you should launch a full investigation for those concern and credit to a pourcentage of customer token consumption to preserve the our thrust in OpenAI product.

With all my deep respect.

5

Since a week, I work on a refactor. passing from Javascript to TypeScript all the fronthend code base which is pretty much very small.

I used Sol-extra-high for making the spec and then the plan.

I checked each goal and make sure it was incremental in the right way.

Nothing fancy at all.

Then I begin coding with Terra-high. He was completing the task super quick. Like lightning quick. But none of his work was working.

Then, I switch from strategy, using Sol-very high ask orchestrator and review while Terra is coding.

I’m a week from that point. 1000$ in token spent. and I realizing, AI can’t achieve it. We can’t leave AI alone because he rewrite the plan, create fixture to nothingness.

Try to reinvente the wheel at each step.

Sol-extra-high over implement and don’t respect kiss. Terra is cutting round the corner.

What should we do? It’s nothing fancy. It’s not low level code, it’s just simple refactor which is supposed to simplify the things.

Each code line I write, GPT rewrite it on his way.

He create test over test over test and at each change, he perform all the damn test.

Hey Open AI, does your model are racketting us ?
Does your backend code realize that he can be lazy on task because the user will still pay for token ?

At the current point, I’m asking out loud, can we trust OpenAI ? Everywhere I see complain about the token consumption and nowhere I see any announcement by OpenAI about the acknowledge of the situation.

We buy AI because we need efficiency and trust.

Now we arent getting either.

If fix and compensation aren’t rollout soon, OpenAI will be bleed out by quiting user.

7

I think that, with the right guidance, the OpenAI frontier models can be really productive but left to their own devices you may experience what you have.

I don’t know what the perfect way of developing with them is but this is what I do:

  1. I have a document + sample code that defines my style of coding + design principles.
  2. I set very clear boundaries on what is deemed the editable surface for the AI (this is usually done in another doc - project specific - but sometimes done via the chat interface).
  3. I then have the AI write an implementation plan for the task at which I can inspect and help tune before any implementation - usually this should include stages/phases which I then ask it to execute in sequence.

I’ve found this approach coupled with Sol (medium) produces very good results - and, importantly, the output (code) is something that I can easily understand because I’ve instructed it to be in my style - which then helps if I want to jump in and take over or for debugging when things go wrong.

8

Hi Mohanlalranvir,

My workflow is this:

Setup AGENTS.md with the guardrail and the scope of the project.

Then I write the spec with Sol high or xhigh. I tune it to exactly my requirement.

Then It’s converted in plan using different agent for not burning context.
I review it.
We tune the plan and make sure each step have goal and test.

Then begin the implementation.

Each task are made on fresh prompt, no trailing context. Always fresh agent.

Even with that, it appear the more token I buy, the less efficient the AI are.

I think there is an exploit.

We can’t verify all his action.

I don’t know if he use user computer as work force on something else, spending our token on task not deserved to us.

But at the moment, I’m trying to work with the best ‘‘good practice’’ found online. It appear even OpenAI support can give the right guideline for best practice.

on the 17th, my reset was scheduled on the 18th, now yesterday, my reset have been postponed to the 22th.

I spent another 200$ on that. We are small tech startup. We can’t feed the dragon that fast.

My accountant just say no.

0.Crédits19/08/2026Paiement effectué109,98 $US

  1. Crédits19/08/2026Paiement effectué219,95 $US
  2. Crédits18/08/2026Paiement effectué219,95 $US
  3. Crédits17/08/2026Paiement effectué109,98 $US
  4. Crédits17/08/2026Paiement effectué109,98 $US
  5. Crédits16/08/2026Paiement effectué109,98 $US
  6. Crédits16/08/2026Paiement effectué109,98 $US
  7. Crédits16/08/2026Paiement effectué219,95 $US
  8. Crédits15/08/2026Paiement effectué219,95 $US
  9. Crédits15/08/2026Paiement effectué109,98 $US

Here is the cost this week for 2 coder. And pretty much nothing achieved this week.

9

I stop the money bleeding until real customer service contact us.

The product don’t work and there is something totally not ethics at all at the moment.