26:50Theo - t3․gg · Essay
The Unexpected Death of Codex
OpenAI folded its beloved Codex app into a rebranded ChatGPT overnight -- Theo argues they just killed the best brand in AI coding.
July 11thThe argument in one line.
GPT-5.6 Codex burns through usage limits faster than 5.5 mainly because it stopped asking for permission mid-task, and the fix isn't a hidden setting — it's writing explicit stop points into every prompt, defaulting to high (not max) reasoning, and turning off fast mode.
Read if. Skip if.
READ IF YOU ARE…
- You have a $100-200/month Codex, Claude Code, or similar AI-coding subscription and are hitting five-hour or weekly usage limits faster than expected.
- You run AI coding agents unattended for long stretches and want to control how far they go without babysitting every step.
- You maintain an AGENTS.md or CLAUDE.md file and want to know which instructions actually change agent behavior.
SKIP IF…
- You're on the free tier or use AI coding tools only occasionally — usage-limit management isn't a problem you have yet.
- You want a tutorial on writing code with AI, not on managing subscription usage and agent configuration.
The full version, fast.
GPT-5.6 Codex users are hitting five-hour and weekly usage limits far faster than they did on 5.5, and the video traces why: 5.6 was trained to keep working instead of stopping to ask for feedback, so a single message can burn up to 15% of a five-hour limit instead of 5.5's 0.1-2%. OpenAI temporarily lifted the five-hour cap, leaving only the weekly limit active. The fix isn't a hidden setting — it's five habits: avoid the unreleased 'Ultra' mode, turn off fast mode's 2.5x burn multiplier, default to high reasoning (past which cost roughly doubles for single-digit score gains), rein in eager sub-agent spawning with one AGENTS.md line, and always write an explicit stop point into prompts instead of trusting the model to know when to quit.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 02:09
01 · Cold open: rate-limit rage
Montage of user complaints (Reddit/Twitter) about GPT-5.6 Codex burning through five-hour and weekly usage limits far faster than 5.5 did.
02:09 – 03:57
02 · Sponsor: WorkOS / AuthMD
WorkOS ad for enterprise auth, pivoting into AuthMD, an open protocol for letting AI agents authenticate on a user's behalf.
03:57 – 09:36
03 · Setting the stage: Ultra, limits, and OpenAI's live updates
Explains why Ultra should be avoided for now, the five-hour vs weekly limit structure, and OpenAI staffer Tibo's same-day announcement removing the five-hour cap and rolling out efficiency changes.
09:36 – 11:29
04 · Fast mode: why to turn it off
Fast mode's 1.5x speed / 2.5x usage-burn tradeoff made sense when 5.5 stopped constantly to ask permission; it doesn't now that 5.6 runs unattended for hours.
11:29 – 17:31
05 · Reasoning levels + benchmark deep dive
Walks the five reasoning levels and DeepSWE/Cursor-bench cost-vs-score data showing 'high' as the value sweet spot; x-high and max cost much more for small gains.
17:31 – 20:30
06 · Subagents: use with caution
Codex's sub-agent implementation (v1 and v2) isn't good yet, and GPT-5.6 spins them up too eagerly; one AGENTS.md line curbs it.
20:30 – 22:48
07 · Sponsor: G2i
G2i hiring-platform ad built around a customer case study (BatAround hiring nine engineers).
22:48 – 26:26
08 · Stop points: the most important tip
The core fix — write explicit stop points into every prompt, from tight ('write a plan, then stop') to far-out ('build, test, PR, address first review round, then stop'), instead of relying on tools or reasoning level to control how far the agent goes.
26:26 – 30:09
09 · AGENTS.md walkthrough + context-window myth-busting
Shows his real AGENTS.md file, debunks a viral tip to manually cap the context window (confirmed wrong by OpenAI's Tibo), and closes on experimenting with your own config instead of copying presets.
Lines worth screenshotting.
- GPT-5.6 Codex Sol can burn up to 15% of a five-hour usage limit in a single message, versus 0.1-2% for GPT-5.5 on the same plan.
- Fast mode delivers 1.5x faster responses but drains usage limits 2.5x faster — a trade that made sense on 5.5 and doesn't on 5.6.
- OpenAI temporarily removed the five-hour usage limit for all Plus, Business, and Pro plans, leaving only the weekly cap active.
- Removing the five-hour limit also removed its role as a circuit breaker — it used to stop a single runaway session before it could exceed about 25% of the weekly budget.
- On the DeepSWE benchmark, GPT-5.6 Sol scores 69% on high reasoning for $3.47 per task, versus 73% on max reasoning for $8.39 per task — more than double the cost for a 4-point gain.
- The gap between low and high reasoning on an easy coding task measured only 50-100 tokens, because GPT-5.6 already throttles its own reasoning depth to task complexity.
- 'Ultra' is not a reasoning level, and one five-hour limit was blown through twenty minutes after the model came back from testing, caused specifically by triggering Ultra.
- Adding one line to a global AGENTS.md file — 'Only use sub agents if the user explicitly requests them' — nearly eliminates GPT-5.6's tendency to spin up unnecessary sub-agents.
- OpenAI's own staff publicly debunked a viral config tweak recommending manually capping the context window, confirming it makes the model dumber without saving money.
- A since-reverted OpenAI change that raised the context threshold from 272k to 372k tokens accidentally charged users more usage than intended, exposing a bug in their own tracking.
- The single highest-leverage prompting habit for GPT-5.6 is defining an explicit stop point — the model is now so unlikely to pause on its own that not stopping it is the default failure mode.
- GPT-5.5's lower usage per message wasn't efficiency — it came from constantly stopping mid-task to ask for permission, which GPT-5.6 was specifically trained not to do.
- OpenCode's engineering team accidentally ran a month of usage on 'medium' reasoning instead of their intended x-high setting due to a key-passing bug, and still rated it their favorite model.
- Codex's sub-agent implementation exists in two versions, v1 and v2, and neither is considered good by the creator, who recommends caution with both.
Five habits that stop GPT-5.6 Codex from burning your limits
USAGE DISCIPLINE
GPT-5.6 runs longer per message than 5.5 because it was trained to stop asking for permission, so controlling usage now comes down to prompting habits and one config line, not a hidden setting.
01Cold open: rate-limit rage
- Complaints about GPT-5.6 Codex burning through 5-hour and weekly usage limits spiked right after launch, even among users who never hit limits on the prior model.
- The usage spike is partly a training change: 5.6 was built to stop asking for permission mid-task, so it runs much longer per message than 5.5 did.
03Setting the stage: Ultra, limits, and OpenAI's live updates
- OpenAI temporarily removed the 5-hour usage limit for all Plus, Business, and Pro plans, leaving only the weekly cap in effect — which is roughly 4-5x a 5-hour limit.
- Removing the 5-hour limit is a double-edged sword: it used to act as a circuit breaker that capped any single runaway session at about 25% of your weekly budget, and that circuit breaker is gone for now.
- 'Ultra' is not a reasoning level and is not worth using yet — a dedicated breakdown is still pending, so treat it as off-limits until that lands.
- OpenAI acknowledged 6 million weekly active users and a usage-tracking bug tied to a context-window change, both surfaced live while this video was being made.
04Fast mode: why to turn it off
- Fast mode gives 1.5x faster responses but burns your limit 2.5x faster — a fine trade when GPT-5.5 stopped itself constantly and you were sitting there waiting, a bad one now that 5.6 runs unattended for hours.
- A message that would use 15% of a 5-hour limit on high reasoning becomes close to half your limit with fast mode's 2.5x multiplier — turn fast mode off unless you're actually watching the run live.
05Reasoning levels + benchmark deep dive
- On the DeepSWE benchmark, GPT-5.6 Sol jumps from 45% (low) to 61% (medium) to 69% (high) for roughly $1, $1.86, and $3.47 per task — each step is a meaningful score gain for the money.
- Past 'high,' the curve flattens: x-high adds 2 points for $1.23 more, and max adds 4 points over medium for $8.39 versus $1.86 — more than double the cost for a small gain.
- High reasoning is the efficient default because the model already scales down its own token spend on simple tasks — the gap between low and high reasoning on an easy task measured only 50-100 tokens.
06Subagents: use with caution
- GPT-5.6 was trained to spin up sub-agents more eagerly than 5.5, and Codex's own sub-agent implementation (both v1 and v2) is flagged as not good — treat it with caution.
- If sub-agent spam is burning your limit, add one line to your global agents.md: 'Only use sub agents if the user explicitly requests them' — it stops the behavior almost entirely.
08Stop points: the most important tip
- The single highest-leverage fix is writing an explicit stop point into every prompt — GPT-5.6 rarely stops on its own, unlike 5.5 which stopped too often.
- A stop point can be tight ('write a plan, then stop and ask for feedback') or far out ('build it, test it, open a PR, address the first round of review comments, then stop') — either way, define the finish line in the prompt instead of relying on tools or reasoning-level settings to rein it in.
- Defining the end condition in plain language is what lets the model run unsupervised for hours without drifting — it's a prompting habit, not a config change.
09AGENTS.md walkthrough + context-window myth-busting
- Don't manually cap the context window or auto-compact threshold in Codex's config — OpenAI's own team confirmed this makes the model dumber and doesn't save money, since compaction itself is the expensive part.
- A viral 'fix' recommending a 272k context_window and 240k auto-compact limit was publicly debunked by OpenAI staff hours after circulating — verify config advice against the vendor before copying it.
- OpenAI's own accidental telemetry bug (raising the limit to 372k) caused users to get charged for more usage than intended, and was reverted within the same 48-hour window this video covers — treat platform-level usage numbers as a moving target, not a fixed rule.
- Spend time customizing your own AGENTS.md or CLAUDE.md by hand instead of installing someone else's preset — a markdown file of your actual preferences changes agent behavior more than any dashboard setting.
Terms worth knowing.
- Reasoning level
- A setting (low, medium, high, x-high, max) that controls how much internal deliberation an AI coding agent does before responding, trading cost and speed against output quality.
- Fast mode
- A Codex setting that speeds up responses by about 1.5x but consumes usage-limit budget roughly 2.5x faster than normal.
- Five-hour limit / weekly limit
- The two usage caps on Codex/ChatGPT subscription plans — a rolling five-hour allowance and a larger weekly allowance equal to roughly four to five five-hour limits.
- AGENTS.md / CLAUDE.md
- A markdown file of hand-written instructions and preferences that an AI coding agent reads at the start of every session, used to steer its behavior without changing code or settings.
- Sub-agent
- A secondary AI agent instance spun up by a main coding agent to handle a delegated piece of work in parallel or isolation.
- Context window / auto-compact
- The token budget an agent can hold in memory at once; when a session grows past a threshold, the agent 'compacts' (summarizes and trims) older context to keep going, which itself consumes tokens.
- Stop point
- An explicit instruction written into a prompt telling the agent exactly when to pause and wait for human input, rather than continuing indefinitely.
Things they pointed at.
14:30toolDeepSWE benchmark
16:01toolCursor Bench
16:40toolArtificial Analysis (intelligence vs cost index)
13:22channelNerd Snipe podcast
28:50toolOpenCode
Lines you could clip.
00:00
“GPT five six is an incredible model, but it's also incredible at burning through people's rate limits.”
05:05
“Ultra is best avoided not used.”
11:29
“So if you're hitting limits and you're using fast mode, please stop.”
19:01
“Only use sub agents if the user explicitly requests them.”
22:48
“The model needs to know when to stop.”
27:33
“This is not correct. Do not do this if you do not understand exactly what you're doing.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00GPT five six is an incredible model, but it's also incredible at . It used to be really hard to hit five hour limits on GPT five five, even x high. Now a five six solo, it's draining super fast even on medium reasoning.
00:12I've already burned through three five hour limits. Seriously, this has to stop. I've now set five six from high to medium, not fast mode, and I'm still burning through my rates at an insane rate.
00:21My five hours are almost gone again. I've hit the limits every single time since Soul was released, and I don't think I ever hit it with five five. Bruh, I've waxed a weekly and a half on one message in Xi.
00:31With GBD $5.05 on the $100 plan, I could barely use up my quota. Now, one small PR with $5.06 Soul consumes about half of my five hour limit. Not gonna lie, GBD five six Soul is great and all, but the usage drain is whack.
00:43Used to be able to work consistently all day with five five on x high. Now I can't even get through a single three hour session without hitting limits. Whack.
00:51Not good for long term work. As you can tell, people aren't happy. And I'll be frank, a lot of this is OpenAI's fault.
00:57There are some things in Codex right now that just don't make sense. I wanna do my best to help you guys out, to make sure you can get the most out of five six without hitting these limits. I made a bunch of these changes myself and I've actually noticed the quality of outputs going up while also using as little as a fourth or a fifth as many tokens as I was before.
01:14I took the time to write an article sharing my advice here, but since then, quite a bit has changed, even though it's been twenty four hours. I've also seen a bunch of really bad advice going around that has been confirmed by many, including the actual team at OpenAI, to not be good. So if you want to get the most out of your subscription plans with Codex, I hope this is helpful.
01:33And even if you don't, maybe you're using something else like Cursor or Claude Code, this video will still have a lot of tips that could help you be better at using those in an efficient, effective way. I have never burned as many tokens as I have with Five Six Soul, and I've really come to know how this model works, what its strengths and weaknesses are, and what gets it to burn.
01:50I wanna take the time to share this all with you, but this experimentation was not cheap, so I hope you can pardon a quick break for today's sponsor. As AI writes more and more of our code, the chances of it getting things wrong goes up too. And as smart as the models are, there are certain places you just shouldn't take the risk.
02:05And if you're trying to land enterprise customers, the auth layer is one of the worst places to take that risk. That's why WorkOS has always been a great option. They really understand enterprise and they don't compromise on the developer experience to get there.
02:17They have really good SDKs, really good integrations, really good docs, and everything else you need to set up your apps in a great way. They also give you up to a million users for free, which is just insane, and I can't imagine many of us having that problem.
02:30I wish I had enough users to hit the limits and need to pay money, but I don't, so it's a really generous offer. But nowadays, users aren't your only users. Your users have agents.
02:40How are those agents going to Auth? And if a company like Microsoft wants their agents to use your service, how are they gonna get in? WorkOS is one of the few companies thinking about this, which is why they partnered with many others to build AuthMD.
02:51They're working with companies like Cloudflare and Firecrawl to get all the pieces necessary so your agents can auth for you, can create accounts for you, and then let you attach them to your service account, whatever, in all of the logical ways that are needed. To be very, very clear, AuthMD is not a WorkOS feature. It's an open standard that they helped build.
03:10They're the authors of the protocol, and obviously that means you can one click turn it on in your WorkOS apps, but this is a standard that anyone could implement, and I'm excited for everyone else too as well. WorkOS is great because they let you sell to enterprises now, but they also set you up for a future where the agents are signing up instead.
03:26Get enterprise proof and agent ready at soydev.link/workos. Time to talk about how to use SOL without hitting limits. The model is great and the $200 plan is still great.
03:36Even the $100 plan is pretty generous. I'm not sure what the state of the $20 plan is right now, but the $100.01 is fine, then $200.01 is really good as long as you don't make certain mistakes. And again, to your credit, OpenAI is not doing this right.
03:51There are a lot of little things that exist in Codecs right now that are kind of forcing you to overuse your usage. And this is their fault.
04:02I feel bad. If people in chat are saying the $20 plan is actually pretty good for coding, that's good to hear. I One person said it's horrible, but others are saying it's fine.
04:11So, yeah. Take it with a grain of salt. Apparently, it's okay.
04:15I am on the $200 plan, and I was hitting limits aggressively when they made this move. Also, full disclosure, when I was testing five six, it didn't count against my usage, which is why I was able to push the model so hard and do a genuinely absurd amount of inference on it. But as soon as it came back, just a day or two before it went public, I started hitting limits immediately.
04:37It was twenty minutes after they said, okay, you have it again, that I hit a limit for the first time. And the reason I hit that limit was a certain new feature called Ultra. I have a dedicated video about Ultra coming very soon.
04:50It might even be the next video after this one. So as always, make sure you subscribe and hit that bell if you wanna better understand all of these pieces. If I was to talk about Ultra in-depth right now, this video would be two plus hours long.
05:01I want this to be to the point and easy to apply to your work and to send to coworkers and whatnot. So I won't do that here. But know that Ultra is not worth using at the very least until you wait for that follow-up video.
05:13So for now, Ultra, no. Ignore this. Don't touch it for now.
05:17This might change in the future. I am hopeful that it will. But for now, Ultra is best avoided not used.
05:23With Ultra out of the way, I do wanna talk about a thing that just changed because it's quite important. OpenAI is aware of the issues. I've been working with them a bunch on getting these things fixed.
05:32Hopefully, more of them are fixed before this video's live. Probably not too many though. We'll still have useful tips long term, but I wanna give you guys context on where things are at the moment of filming.
05:41Tivo posted this morning the following update, which is very relevant to what we're talking about here. The last forty eight hours of Codex and ChatGPT work have been intense. There's three important updates.
05:50First is that they've temporarily removed the five hour usage limits for all Plus business and Pro plans. If you're not familiar, your sub is broken up into two limits. You have a five hour limit that resets every five hours starting from when you send a message.
06:05So let's say you do a bunch of work at the start of the day, you start at 9AM, you have a limit to how much you can use between 9AM and 2PM, and then let's say you hit that limit at 1PM. You have an hour until you can use it again.
06:16But if you don't start until three, the next five hour limit doesn't start until 03:00. One of the tricks I have is I have a cron job running that does something every five hours so that I'm always burning one of those five hour limits. It uses a very small model with a really simple like hello, hi prompt.
06:32But now you don't have to at least in the interim because the five hour limit is gone for now. The weekly limit is the one to be scared about. The weekly limit is equal to roughly four or five of the five hour limits.
06:45So if you go hard enough to max out the five hour limit four plus times, you will hit the weekly limit. Weekly is the only one applying right now, which is good because the five hour limits were a little too easy to hit, but it's bad because if you accidentally ran an ultra fast run that is way heavier than expected, the five hour limit would have stopped it before it used more than 25% of your weekly.
07:06Now it won't. So it's possible to run one prompt and blow through your whole weekly limit. I'm tempted to do it as an example and then burn a reset, but I'm I'm hoarding my resets.
07:15I'll be real, guys. So with all of this known, let's go through the rest here.
07:20The second point he made is rolling out changes that will make Five Six Soul more efficient across the board and it will be reflected in less usage being used so it can take you further. Exact impact to be quantified and shared. Still working on that there.
07:32And his most exciting point here is that they have 6,000,000 active users and they're about to land a usage reset. That one already landed. So as was established prior, they got rid of the five hour for now and we just have the weekly limit.
07:42I fired one prompt off before streaming, so I don't have much used here yet. But I'll probably burn through this in the next day or two, and I'll be sure to share how, why, and what I'm doing to minimize and maximize usage. With all of this established, it's time to give practical advice.
07:56Part of how I wanna think about this is with Five Five versus Five Six. When Five Five came out, I was far from the biggest fan. I saw how capable it could be, but I didn't love using it because it felt like it would just lose track of what I was doing.
08:11It was really easy to screw up its context. If it read the wrong file and had something in its history, it would fixate on that instead of doing what I asked it to do. And this happened a lot.
08:20It also was really really bad at compaction which was just obnoxious when you had long running threads. I've never had more threads than I do with Five Five because I felt like I had to in order to keep it on task. None of that is why Five Five was cheaper though.
08:35The main reason Five Five was cheaper is that it stopped and asked for permission or feedback all of the time. If you asked Five Five to write a plan, it would. It would say, here's step one, here's step two, here's step three through 10.
08:48And then you would say, okay, go do all of the steps. It would get through step one, it would get halfway through step two, then and it would stop and say, okay, I finished step one and I'm halfway through step two. Let me know when I can keep going.
09:00It's like I never told you to stop, bro. The result of this behavior is that Five Five didn't really use your limits heavily on a per message basis. From my experience, a single message on Five Five even on the highest reasoning levels would use between like point 12% of your five hour limit per message.
09:18It wasn't that bad. Five six fixed these problems, but in doing such, massively increased how much usage you're getting.
09:28Five five was a price bump from five four. They went from $15 per mil out to 30 per mil out, which would have hurt a lot more if it wasn't for this sudden stopping behavior from five five.
09:38Five six fixed that behavior, but as a result, it can go much longer per message, which can result in much bigger usage per message. For my experience on higher reasoning levels, especially x high and max, five six can use up to 15% of my five hour limit in one message.
09:56This is a big jump obviously, but honestly, it's not too too bad because I would end up using similar amounts of five five, but it would be prompt after prompt after prompt. Because each message only would use so much of my limits, I didn't find it too brutal and I would often use it with fast.
10:15Fast mode is pretty cool because you get 1.5 x faster speeds. But as a catch, you go through your limit 2.5 x faster. And this wasn't too bad when I was using five five because when I sent one message with five five, the 2.5 x more burn would make that 2% into 5%.
10:34That wasn't too bad. 2.5 x burn on a message that was 15% though, that's a little more brutal.
10:43That is closer to half of your five hour limit from one message. And that's not even with Ultra. That is scary.
10:51And that is a problem I think a lot of people are having right now is that they are continuing to use the model the way they did with five five. And with five five, since each message was less usage, fast mode didn't feel too bad.
11:02So if you're hitting limits and you're using fast mode, please stop. It's not that much faster. The reason I think fast mode felt so good with five five is because it would stop itself constantly and you had to talk back.
11:15I never watched my agents run quite as much as I did with five five, both because its context was constantly getting screwed up and I had to prune it, clean it, make a new thread, but also because it stopped all the time, so I would have to respond and tell it what to do next. As such, the speed mattered a lot because I was sitting there waiting for it to respond.
11:33With five six, I spin it up and then I go do something else. Maybe I spin up another thread, maybe I go do code reviews, maybe I check email, maybe I play Power World, I don't care. With five six, I do other things.
11:44With five five, I have to sit there and watch, so fast mode felt important. With five six, it's running for so long anyways that the inference is barely the slow part. It's usually the tool calls, the test runs it's doing, all the other things it has to do.
11:57I've barely noticed a difference in speed since turning off fast mode, but I have noticed a massive shift in the amount of usage that I'm hitting with it. For context, there's lots of people in chat including those on my team like Maria that says almost every prompt and thread they've done with five six has lasted over eight hours.
12:14Yeah. You can get these models going for a while. Even without using slash goal, to be clear.
12:20I almost felt like slash goal was necessary with five five because otherwise, it would stop. And the goal would basically tell it to keep going over and over instead of having a human do it. With five six, not necessary.
12:31I never use slash goal anymore unless I'm just trying to burn tokens. Generally, with five six, it will complete the thing without needing extra encouragement. So now establish two things.
12:40Don't touch ultra, and you should probably turn off fast mode at this point because it doesn't matter as much as it used to. But there's more that we can learn from. Reasoning level should be an easy enough one to cover.
12:50I'll dive through this super quick. You have five options. They've been rebranded, but they used to be low, medium, high, x high, and max.
13:00You might think alter is a reasoning level. It's not. Again, video crashing out about ultra coming very, very soon.
13:06Hit that button if you haven't at the bottom. Subscribing makes it more likely you see that when it drops. So with Ultra removed, we have these five options.
13:14I did a really good analogy about what these are and how they work on my most recent podcast episode that should hopefully be out in the next day or two. If you're not already listening to the podcast, check that out as well. Ben and I just nerd out about the details here and it lets me go a little longer than I try to on the videos.
13:28Trying to keep these shorter lately. Hope you appreciate the effort. So instead of breaking down all of the details here, I'll give it to you simple.
13:35Ignore everything past this line for now. X high can be really cool. Max is less likely to be.
13:42Low, medium, and high are all very good options. I mostly just default to high right now because the model's so efficient. Even on high reasoning, if the task is simple, it will stop really quickly and not reason too much.
13:53When I was benchmarking the model, I thought there were bugs in how I was passing the reasoning level, because the difference between low and high was like 50 to a 100 tokens at most, which just didn't seem right at all. Turns out, it is actually just good at that. It will use less tokens if the task is simple.
14:10So leaving it on high has been fine for me. I do wanna look at the numbers quick though. So let's do that with deep SWE.
14:18This is my favorite code benchmark right now, very much subject to change eventually. I'll throw in Fable for a comparison here. So with GPT five six SOL, on low, you score 45% and in this bench, it costs a dollar per task.
14:32Medium, it bumps all the way from a 45% to a 61% for a dollar and 86 per task.
14:39And then high bumps you up to a 69. Nice. Which puts you neck and neck with Fable on this benchmark, which gets a 70.
14:47But the cost is $3.47 per task versus Fable at the same score being $13 per task.
14:55So that is super efficient. All three of these are meaningful jumps in score. You get from a 45% on low, to a 61 on medium, to a 69 on high.
15:05But then we see the next two levels like x high, which goes from a 69 nice to a 71 less nice, but also bumps the price meaningfully going from $3.47 to $4.70 per task. And while we're not paying API rates when we use it through a subscription, the API rates are a good measure of what you're burning.
15:24So yeah, that's not great. And then when we bump up to max, the price doubles to $8.39 for an additional two percentage points.
15:35I don't think that's good. Medium is a dollar 86 and gets a 61. Max is $8.39 and gets a 73.
15:43And again, high is $3.47 for 69%. That's a more than double cost for a four percentage bump.
15:50Not worth it. Just stick with high. If you disagree, that's fine.
15:54I just hope you have the budget to handle your disagreement there. And for full transparency, there are other benches that have shown different numbers.
16:01For example, in cursor bench, high only scored a 63.5 and max scored a 67.2, and the price gap there was two seventy nine to five sixty nine.
16:11So it wasn't quite a doubling of price and it was a much more meaningful percentage bump. So this will depend on the work you're doing. Also worth noting that in cursor bench three two, Fable ended up scoring quite a bit higher than Soul did, which we'll talk more about Soul versus Fable in the near future.
16:26That will be a big video too. I don't wanna be sidetracked by that. The point I'm trying to make here is high is really good, and past high is when you lose this like vertical line where the cost jump isn't that big and the success jump is.
16:41I just decided to quickly check the artificial analysis intelligence versus cost index, And when I turn off everything other than five six soul and Fable five, five six soul high is the only thing in the good quadrant of cost to intelligence.
16:56It does meaningfully go up for each bump, but again, five six SOL seems to be this really solid end of the vertical curve. So reasoning levels are now covered. Now we need to talk about what is probably the biggest behavioral difference with the five six model, specifically both SOL and Tyra.
17:14We need to talk about sub agents. There's a lot of different ways to implement sub agents, but you generally need a tool that allows for your main top level agent to spin up other work for other agents to complete. I could say that Codex's implementation of sub agents isn't good, but that wouldn't be fair because they have two implementations of sub agents called v one and v two, and they're both not good.
17:35I have a lot to say about that. Again, the Ultra video will go more in-depth there. I will resist the urge to crash out hard about it.
17:43But what I will say with relatively high confidence is use with caution. There's a problem though.
17:50You might want to use sub agents with caution, but Five Six was trained to use them with less caution. As such, it's not uncommon for Five Six to spin up sub agents for things where it doesn't really need it. If you notice this happening, I don't think you should rush to go change anything about this just yet, because the sub agents can be pretty good.
18:08And I'm also hoping that the Codex team fixes the issues I have with it as they currently stand. But if you are noticing limit burn after applying all of these other recommendations and you're seeing sub agents spinning up a bunch, I would recommend doing a slight adjustment to your global agents md file.
18:24The easiest fix for now is to add the following to your agents md. Only use sub agents if the user explicitly requests them.
18:31This change will pretty much stop them entirely unless you request, and I usually request when I want them. You'll eventually figure out when to or not to use sub agents for work, like what work justifies splitting things up that way.
18:46So don't be hesitant to experiment with them. I think they're really cool, especially when applied carefully. But right now, Five six is a little too eager.
18:54If you need to tone it down, here's the strategy to do such. It is worth noting that the sub agent implementation and other tools like ClaudeCode and Cursor is meaningfully better, so much so that I've actually done some hacks in order to get ClaudeCode working with my Codec sub, and I've been really, really happy with the results there.
19:11And crazy enough, Thibault even blessed it, saying that if anyone is banned for using my weird hacks to get my Codec sub into Claude Code, that he will give you a reset. So yeah. This is a blessed solution.
19:23Very fun video on that coming in the near future. Just wanted you to know that it's worth exploring other options. I know PY is a decent one as well if you're curious enough to explore.
19:31But I do expect codecs to fix this in the not too distant future. So if you're doing it out of fear, just wait. If you're doing it out of curiosity, go nuts.
19:40One other thing I forgot with the reasoning level bit, and I'll just be super quick about this, is model selection. TLDR, don't use Luna.
19:47It's just not for us to use for code. It's really useful as a thing that you access programmatically, like you're using it over API to filter data and whatnot.
19:55Not really what you should be using for coding. Tyros seems to be a good middle ground, but honestly, I would rather just use sole on lower medium. They're really efficient.
20:03I just recommend using Soul High as your default. And if you find things that it goes longer than it should for, try out Solo, maybe play with Terra. But I have been totally fine just using Soul for everything.
20:14And once I made these other changes, I was not coming close to my limits. But there is one last change we have to talk about, the most important tip in this whole video. This is definitely the change that affected how Soul worked for me the most, and I'm really excited to share more about it right after a quick break for a sponsor.
20:29Today's ad's gonna be a little different. You've already heard me talk about g two I. They help you hire world class engineers at whatever scale and time frame you need.
20:36Normally, tell you all about how crazy it is that they can get you engineers in under a week that are actually experienced and ready to go, but instead, I'm gonna tell you about one of their customers, BattAround, because they've hired nine engineers through g two I and they learned about g two I through me. So I think that's pretty cool.
20:51These guys needed to hire and if you've ever hired, you know how grueling it is. You just lose all of your spare time and no work gets done. If you're spending time hiring, you're not spending time working.
21:02And it's this weird unintuitive thing where when you realize you need more help, you end up needing more help sooner because you're stuck doing other stuff. And they wanted to hire, but they couldn't afford to lose bandwidth. Battalion's leadership team had decades of experience hiring engineers in other sectors, but their network in the media space was small.
21:19It was getting hard to find qualified candidates, which is why they reached out to G2I. They wanted fewer resumes from better candidates so they could more easily evaluate. G2I already had a large network of engineers ready to go, and they filtered to the best ones fit for the specific needs of Battaround in this mixed media space.
21:35With G2I, Battaround has contracted nine engineers, had a 90% success rate, reduced leadership time spent screening candidates, integrated engineers directly through their existing Slack workflows, and they've minimized disruption through the clean transitions and fast backfills when they need to shift around their staff.
21:50Battaround themselves said that a 50% success rate would feel like a superpower to them, and now they're up to 90. Waste less time and hire better engineers at soydev.link/g2i. It is worth noting that a lot of analysis is showing that Terra isn't really a great option, which is disappointing.
22:06I was quite excited about it. Artificial analysis said the following. Five, six, Soul, and Luna are ahead of Terra at every point on the intelligence versus cost per task chart.
22:15This is the chart that I had just shown before. And here you could see that Luna on max scores slightly better than Soul on low, while being just barely more expensive. It almost looks like a pretty consistent line from Luna low mid high x high, and then Sol low mid high x high.
22:31But Terra underneath here, not looking great. The model needs to know when to stop.
22:38As I was hinting at earlier, five six is very eager to get work done. It will keep going as far as it possibly can unless you give it a reason to stop going.
22:49I gave some good examples in my article, so I'm gonna reuse those here. This is the style of prompts I've learned to write when using Soul because without it, it can go further than I want it to pretty often. Example one is the following.
23:01I want you to build this new feature. Start by writing a plan. When you finish the plan, stop and ask for feedback before proceeding.
23:09I have told the model explicitly, do this thing and be done at this point. That doesn't mean you have to give the model prompts that are less work though.
23:18You can have the stop point be way, way, way further on. For example, the plan looks great.
23:24Let's build it out. Use computer use to test your implementation. Keep going until the code works and you're happy with the implementation.
23:30Put up a PR, babysit it for the first set of review comments and address them. Stop after the first set of review comments, I'll handle it from there. Previously with Five Five, the model stopped way too often.
23:41With Five Six, it's so unlikely to stop that I feel like I have to put up the stop signs myself. But now that I've been doing that, it feels much better. It's nice that I could start building my own intuition for how far the model can go without my intervention, and also that I can kind of choose when to stop it.
23:59Rather than trying to do this through reasoning levels or changes to your harness, so much of this can just be done via prompting. And I will say I am proud of myself for getting this far in the video without my instructions being prompt better.
24:11Because your prompt can absolutely cost you more money, but there's other things that I thought were more important. This piece though, this can change how you use the model fundamentally. When you realize that the end is a thing you can define in the prompt rather than a thing you have to set up with tools, you can let the model go way further.
24:28And the challenge I would present to you is see how far off you can make that stop sign without the results in the quality going down. I didn't talk about this one too much here, but it is really fun. You can let other agents steer.
24:41For example, I set up my Claude code to allow for it to call Codex. I've went much further with this since and that's the Codex in Claude code video coming soon. We're actually using Sol as the model in Claude code.
24:53Seriously, Claudex is way better than I expected it to be. And as silly as it is seeing this five six Sol there, it's been working great for me.
25:01But most importantly here, the thing I wanna encourage is more experimentation. You shouldn't be trying to copy my skills, my agent m d, and my prompts directly, because this should be different for everyone.
25:14You should experiment more with these different things. Try prompting different ways. Try different reasoning levels.
25:20Try adjusting your codecs AgentsMD and your CLOUDMD files a bit to see how it changes behavior. You should be spending a good bit of time in the dot Codex and dot Cloud directories on your computer if you're using these agents and models. It's so customize because it's not like you have to be deep in the weeds of the details of how it works in all these config files.
25:41It's a markdown file and changes to that can fundamentally change the experience that you're having. And the more you move things around, the more you channel our inner Dan Abramov to experiment and shift until things feel the way you want them to and have your own feelings change in the process, that's what's gonna give you the best experience using these things.
26:00If you blindly follow my recommendations or you go set up somebody else's recommendations like you install any of the oh my whatevers, you're never going to learn what does and doesn't work, and you're going to be limited by the creativity of someone else, go play with it yourself.
26:15I don't have a bunch of other people's skills and things installed. I just dick around and change things and read the traces from my agents when they don't do what I expect and make adjustments until they do. I am currently deep in the process of rewriting my agents m d and this is all by hand.
26:32Not a line of this was LM generated, and I think that is one of the best places to take the time, because you'll be amazed how much these things can change how the model feels to you. What I'm trying to say is that you need to stop looking for for you to buy, copy, paste, and hope for the best.
26:49You should use resources like this video and like these other examples people have as inspiration to get it the way you want yourself. I mentioned at the start that I've been seeing some bad advice.
26:59This is probably the worst one. While I don't trust OpenAI to configure everything perfectly in Codex for us, they absolutely covered the context windows themselves just fine.
27:10I've seen a lot of people making recommendations like this, specifying that you should limit the context window and when it auto compacts manually inside of your config, don't do this. The model was trained on specific levels of compaction in tools like Codecs.
27:26This is going to make the model dumber and it's not going to save you any money. If anything, it's gonna cause compaction to happen more than it should and compaction is expensive. Thibault jumped in and said as much.
27:35This is not correct. Do not do this if you do not understand exactly what you're doing. We do not charge extra above two seventy k context and the context threshold has been tuned for five six sold to be perfect with a default limit.
27:44It is, trust him, don't just blindly follow random advice on Twitter, at the very least make sure the people giving the advice aren't getting their advice countered by people at OpenAI. I hate filming live sometimes because literally right after I finished that part and was ready to go offline, Thibault tweeted the following.
28:03We've laid inference optimizations to make it cheaper to run. It's gonna be 10% savings. Cool.
28:08Awesome. They also noticed that changing the context size limit in the product to three seventy two k, up from two seventy two k, resulted in more usage being charged than intended. This means that their usage tracking was broken even though they planned to let this stay.
28:22They've temporarily reverted to two seventy two k and will roll back in the days to come. This should be a big change in how fast your usage drains. Cool.
28:29Awesome. Stupid that we're here, but here we are. They confirmed that the leaks around juice values being changed were true, but they've been reverted since.
28:37And they also call out that there's been more use of multi agent than intended in high and high reasoning efforts, and they're fixing it going forward, also fixing small other things they noticed with auto review where they can be more efficient. Reporting on the fly is annoying, but you now have the additional context.
28:51That all said, every piece of advice I gave in this is still true, so don't think this means my advice is invalid, just know that this is a moving target. Best practices are still best practices.
29:01One more point in favor of lower reasoning levels, the OpenCode folks are obsessed with Five Six.
29:08They barely even use Fable, they like Five Six so much. They like it so much that most of OpenAI has been promoting their posts throughout.
29:15I assume they were probably using it on higher reasoning levels, because who wouldn't initially? They hit an issue though. One that I've hit when I was testing these things over API.
29:24They changed the key in the JSON for reasoning levels at some point, and a lot of tools don't pass through reasoning level properly. I ran into this a lot during benchmarking. Turns out, OpenCode did too.
29:34And during all of their usage, a month of usage, they realized that they had it misconfigured to always use Medium, even though they thought they were using x high or other levels. And despite that, the whole team agreed it was their favorite model.
29:48Definitely an endorsement for Medium as the default. So if you do like OpenCode and or you trust the OpenCode team, they have confirmed Medium is incredible. Medium and High are both really, really good options.
29:59So go play, go configure, go customize, and see what you can do. I have a feeling you'll be surprised. Keep playing and keep prompting, and until next time, peace nerds.
The Hook
The bait, then the rug-pull.
GPT-5.6 Codex users are blowing through five-hour and weekly usage limits days after launch — some burning half a limit on a single message. This breaks down exactly why, and the five concrete changes that stop it.
Named ideas worth stealing.
11:29list
Reasoning level ladder
- low
- medium
- high
- x-high
- max
Five reasoning-depth settings for GPT-5.6 Codex; cost and score both rise with each step, but the value curve flattens hard past 'high'.
Steal forbudgeting any AI coding agent's default reasoning setting
How they asked for the click.
VERBAL ASK
02:10link
“Get enterprise proof and agent ready at soydev.link/workos.”
read live over a WorkOS site screen-share, tied thematically into the video's AI-agent-auth angle rather than bolted on
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
