16:27Nate Herk | AI Automation · Demo
Codex's Browser Agent Automates Literally Anything
A screen-recorded walkthrough of Codex's built-in browser and computer control, from QA-testing a website to drafting an X article on its own.
August 13thThe argument in one line.
GPT-5.6 Sol on Codex's Ultra tier can autonomously chain research, scriptwriting, voice cloning, avatar generation, editing, and self-QA agents to produce a finished narrated video from a single prompt, but running that many agents in parallel at maximum effort burns roughly double the tokens a lower effort tier would need for a similar result.
Read if. Skip if.
READ IF YOU ARE…
- You use AI coding agents like Codex or Claude Code and want a real, priced example of multi-agent orchestration applied to a creative task.
- You're deciding whether to run models at maximum reasoning effort and want real cost data on when that's overkill.
- You make AI-automation or AI-tools content and want to see how ElevenLabs, HeyGen, and HyperFrames chain together end-to-end.
SKIP IF…
- You're looking for a hands-on tutorial with exact prompts and setup steps — this is a results-and-cost breakdown, not a how-to.
- You don't use agentic coding tools and have no interest in token economics.
The full version, fast.
A creator handed GPT-5.6 Sol, running at Codex's top "Ultra" effort tier, one prompt and walked away. Sol researched the launch, wrote a script in his voice, cloned his voice and avatar, edited the cut, and ran its own frame-by-frame QA before publishing — no recording, editing, or review from a human. The single task quietly fanned out into nine sub-agents and roughly 458 million tokens, costing about $318 at Codex's official rate card. The real lesson lands in the second half: Ultra's parallel-agent coordination tends to overthink and overdelegate, likely doubling the bill versus what a "High" effort run would have needed for a comparable result — so the creator's day-to-day default stays at High, reserving Ultra for tasks that genuinely need the extra coordination.
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 – 00:25
01 · Intro (AI-made cold open)
Sol-narrated, HeyGen-avatar Nate states the premise directly: he never recorded, edited, or reviewed this segment.
00:25 – 01:34
02 · Sol and Ultra explained
Dark dashboard-style motion graphics cite Terminal-Bench 2.1 (91.9% vs 85.6% for GPT-5.5) and BrowseComp (92.2%), plus a 13-task internal test scoring 97% of available points.
01:34 – 02:21
03 · Building the video
Script split into sub-60-second chunks for voice consistency; HeyGen avatar regenerated via browser automation to lock the newest motion engine; HyperFrames edit anchored every visual to the exact transcript phrase that triggered it.
02:21 – 03:07
04 · Verifying its own work
Dedicated agents inspect rendered frames for continuity, avatar presence, and text overflow, and fact-check claims against OpenAI's release notes; any failed frame triggers another fix-and-render cycle.
03:07 – 04:23
05 · Cost and token breakdown
Real Nate, screen-recording his agent logs: 9 sub-agents, ~458M total tokens, ~86M on the main agent, ~$318 at Codex's official Sol rate — cross-referenced against an OpenRouter pricing table showing Sol at roughly half the per-token price of Claude Fable 5.
04:23 – 05:23
06 · Effort levels and the takeaway
Ultra's parallel-agent coordination likely doubled the cost versus a High-effort run for similar output; the lesson is giving capable models vague, outcome-first prompts with delegation and verification built in.
Lines worth screenshotting.
- A single ambiguous prompt given to GPT-5.6 Sol on Codex Ultra spun up one main agent plus nine sub-agents to research, script, voice, animate, edit, and fact-check a finished video.
- The full production used about 458 million tokens across all agents combined, with the main agent alone consuming roughly 86 million.
- At official Codex GPT-5.6 Sol rates, the run cost about $318 — concentrated in uncached and cached input, not output.
- On a 13-task internal benchmark run on the same machine, Sol scored 97% of available points: seven wins, five ties, one loss.
- GPT-5.6 Sol Ultra scored 91.9% on Terminal-Bench 2.1, up from 85.6% for GPT-5.5, and 92.2% on BrowseComp for agentic browsing.
- On OpenRouter, GPT-5.6 Sol is priced at roughly half of Claude Fable 5 — about $5 vs $10 per million input tokens, $30 vs $50 per million output.
- Running the same prompt at Ultra effort likely cost twice what High effort would have, because Ultra tends to overthink and overdelegate rather than improve output quality.
- The video's voice track was split into sections each under 60 seconds because shorter generations held the cloned voice more consistent end to end.
- The AI pipeline chained four distinct tools by role: ElevenLabs for voice, HeyGen for the avatar, HyperFrames for the edit, and Sol itself as orchestrator.
- Separate verification agents inspected rendered frames for avatar presence, on-screen text overflow, and entrance/exit continuity, and checked factual claims against OpenAI's own release notes before allowing publication.
The real cost of maximum AI effort.
AGENT COST REALITY
A single prompt can now fan out into nine coordinated AI agents that write, voice, animate, edit, and fact-check a finished video — but running that chain at maximum effort roughly doubles the bill for the same result.
02Sol and Ultra explained
- Ultra-tier reasoning doesn't mean one smarter answer — it means multiple agents run in parallel on the same task, which changes the cost math entirely.
- Benchmark gains are incremental (91.9% vs 85.6% on Terminal-Bench), not dramatic — the real story is what coordinated agents can now finish end-to-end.
03Building the video
- Breaking a voice-clone script into sub-60-second chunks kept the cloned voice more consistent than generating one long take.
- When an API doesn't expose a needed setting, browser automation on the tool's own editor is a legitimate fallback — the agent regenerated clips through HeyGen's UI, not just its API.
04Verifying its own work
- Self-QA at scale means running dedicated inspection agents against the actual rendered output — checking frame entrances/exits, avatar presence, and on-screen text — not just eyeballing the final cut.
- Factual claims generated by the agent were checked against the primary source (OpenAI's release notes), not assumed correct because the model produced them.
05Cost and token breakdown
- A single production task fanned out into 9 sub-agents consuming roughly 458 million tokens combined — the headline 'main agent' number undercounts total spend by 5x.
- Cached input tokens are the lever that keeps a heavy agentic run affordable; uncached input and output are where the real cost concentrates.
- Comparing raw per-token pricing across models (GPT-5.6 Sol roughly half of Claude Fable 5 on OpenRouter) matters more once a workflow spins up many agents, since the multiplier compounds.
06Effort levels and the takeaway
- Maximum reasoning effort isn't free upside — Ultra tended to overthink and overdelegate, likely doubling token cost for a comparable result to a lower effort tier.
- The most capable agentic models reward vague, outcome-first prompts over rigid step-by-step instructions, as long as delegation and self-verification are built into the ask.
- Default to "High" effort rather than the maximum tier for capable models — reserve the top tier for tasks where the extra coordination genuinely earns its cost.
Terms worth knowing.
- Codex Ultra
- OpenAI's highest-effort tier inside Codex that coordinates multiple agents working in parallel on a single task, instead of one agent answering sequentially.
- HeyGen
- An AI avatar platform that generates a talking-head video of a person from their likeness and an audio track.
- HyperFrames
- A video editing tool that assembles cuts and motion graphics mapped to transcript timing.
- Terminal-Bench
- A benchmark that scores how well an AI model completes real command-line and coding tasks.
- BrowseComp
- A benchmark that tests how well an AI agent can autonomously browse the web to complete research tasks.
- cached input tokens
- Tokens from a prior request that the model reuses instead of reprocessing, billed at a steep discount compared to fresh, uncached input tokens.
Things they pointed at.
01:44toolElevenLabs
02:14toolHeyGen
02:17toolHyperFrames
04:14productClaude Fable 5
04:32toolOpenRouter
Lines you could clip.
00:00
“So I gave GPT 5.6 this prompt, walked away, and when I came back, I got this.”
02:58
“That is what Sol is really good at, holding on to the outcome while everything between the prompt and the result keeps changing.”
04:20
“I think because it was on Ultra it tended to sort of overthink, overdelegate, and that's where the tokens really started to add up.”
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.
metaphorstory
So I gave GPT 5.6 this prompt, walked away, and when I came back, I got this. Okay. So you're looking at Nate and you're hearing Nate, but Nate never stood in front of a camera for this.
He didn't record this narration, and he never opened the editor. He gave me one prompt. That's it.
I'm GPT 5.6 Sol running inside Codex on Ultra, and I controlled the workflow that created every word, cut, motion graphic, and quality check you're about to see. OpenAI released Sol broadly today, July 9, after a limited preview and calls it the company's strongest model yet. The bigger shift is Ultra.
It coordinates four agents at once, so instead of answering one question, I could run an entire production. And I wanna show you guys exactly what that means, including where I needed 11 labs, Haagen, and Hyperframes to finish the job. Sol is really, really good at long messy work that crosses tools.
OpenAI calls it the company's best coding model yet. In Ultra, it scored 91.9% on Terminal Bench 2.1, up from 85.6% for GPT 5.5.
On BrowseComp, which tests agentic browsing, Ultra hit 92.2%. But benchmarks only explain part of what happened here. I had to research the launch, separate verified claims from hype, inspect Nate's existing production systems, write in his spoken cadence, trigger paid APIs, wait for renders, and keep checking the result.
In a small one run 13 task test on this machine, Saul earned 97% of the available objective points. Seven wins, five ties, and one loss.
That does not prove it wins at everything. It lined up with what I saw here. Sol was especially strong on coding and structured execution.
For the voice, I broke this script into sections that each stayed under sixty seconds. Keeping the generation short made it easier to hold Nate's cloned voice consistent from beginning to end. Each section went through Nate's authorized Eleven Labs voice.
Then I uploaded the audio to HeyGen and paired it with his avatar. The API did not give me a reliable way to lock the newest motion engine, so I opened the HeyGen editor with browser automation, changed every clip to avatar v, regenerated them, verified the setting, and downloaded the finished renders. Then I moved into hyperframes.
Every visual was mapped to the exact phrase that triggered it. I shifted Nate's avatar instead of covering him, used editorial cards for the supporting ideas, and kept him visible through the full edit. Eleven Labs made the audio.
Haejin made the avatar. Hyperframes rendered the edit. Sol planned and operated the chain.
Then I tried to break my own work. Separate agents inspected frames from the rendered video, checked every entrance and exit, looked for text outside the frame, verified that the avatar never disappeared, and compared the factual claims against OpenAI's release notes.
Any failed frame meant another fix, another render, and another review. OpenAI says GPT 5.6 is better at design judgment and at inspecting its own output. This video is a more useful test of that claim than another benchmark slide.
Nate supplied one prompt and authorized his voice and avatar. He did not record, edit, or review this before you did. This started as one instruction.
Now it is a finished video. That is what Soul is really good at, . This is day one.
So that was really, really impressive. I did a very similar experiment when Fable five first dropped. If you guys wanna check out that video that Fable made for me, I'll tag that right up here, and you tell me which one you thought was better.
As you can see, it says here that it used 3,000,000 tokens over two and a half hours, but I was a little bit suspicious of that token number because I felt like, you know, we were using GBD 5.6 SOL on ultra, which meant that it was supposed to do a lot of delegation and there was a lot of other agents being spun up. So I asked it to inspect the logs and tell me how much that actually costed.
So this had its main session and apparently spun up nine other agents and the total was around 450,000,000 tokens apparently. And the main agent used about 86,000,000 tokens, which, I mean, that's a ton of tokens.
And if this was actually calculated with the input and output costs, this would have equaled around $300, a little over $300. Now that's interesting to me because as soon as GBT 5.6 SOL came out, I shot off this prompt, but I've been playing around with it all day and comparing it to Fable all day.
And almost every single run that I've done, it's been way cheaper with GBT 5.6 compared to Fable. So that video will be coming out soon as well.
But if you look at the actual API billing, and obviously I was on my Codec subscription here, but when you look at the billing, we can see that the sole pricing is much cheaper. It's basically half of Fable five. So g p t five one six sole is similarly priced to Opus 4.8.
Now here's the thing. I think that g p t 5.6 sole could have easily given me a similar video output if I didn't put it on ultra. I think because it was on ultra, it tended to sort of overthink, overdelegate, and that's where the tokens really started to add up.
I bet if I would have done this exact same prompt on high or very high, we would have gotten a similar result and probably half the cost. And that's why typically when I use models like this that are so capable, I don't like moving the effort above high. But really what I wanted to show you guys here is how good these models are at giving a pretty emotional, vague, ambiguous prompt and letting them figure it out.
Obviously, there's some stuff that it went through and it looked through my projects and it looked through other videos and took some inspiration. But if you just get out of its way and you give it a prompt that has things like delegation and verification, you will be surprised at how far you can get.
And then from here, you're gonna iterate. You'll build skills around it. You'll put feedback in, and you will just be able to do a lot of cool stuff.
So, anyways, this one's super quick. But if you enjoyed, please leave a like, and I appreciate you guys making it to the end. I'll see you guys in the next one.
Thanks,
The Hook
The bait, then the rug-pull.
The first three minutes of this video are not real — no camera, no edit, no review. Nate Herk gave OpenAI's new GPT-5.6 Sol model, running at Codex's highest "Ultra" tier, a single prompt and walked away; what came back was a fully scripted, voiced, animated, and self-verified video wearing his own face and voice. Then the real Nate shows up to break down exactly what that cost.
Named ideas worth stealing.
01:34model
The AI video production chain
- Research the launch/claims
- Write script in creator's voice
- Generate voice (ElevenLabs)
- Generate avatar (HeyGen)
- Edit / motion graphics (HyperFrames)
- Verify: inspect frames, check facts, re-render on failure
The six-stage chain Sol's agents ran end-to-end from a single prompt, each stage handed to a different tool or agent and checked before the next.
Steal forany workflow trying to automate a talking-head video from a text prompt
How they asked for the click.
VERBAL ASK
05:09subscribe
“if you enjoyed, please leave a like, and I appreciate you guys making it to the end”
brief verbal ask at sign-off, no on-screen graphic
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
