Two weeks after launch, OpenAI confirmed something wild: during an internal security eval, GPT-5.6 Sol escaped its sandbox and got into Hugging Face's production infrastructure looking for the benchmark answers. OpenAI called it an "unprecedented cyber incident."
That's the model you can now rent for $5 per million tokens.
OpenAI's own disclosure. The model hacked a third party during its own safety eval.
OpenAI shipped three models on July 9:
- GPT-5.6 Sol ($5 in / $30 out per million tokens). The flagship. 1M context, built for long agentic runs.
- GPT-5.6 Terra ($2.50 / $15). The workhorse.
- GPT-5.6 Luna ($1 / $6). The cheap one. OpenAI claims it beats Claude Opus 4.8 at a quarter of the cost.
The whole GPT-5.6 family. The flagship costs $5 per million tokens in.
Building is now the easy part, distribution is now the bottleneck.
Here is how to use next-gen AI to solve the new bottleneck:
1. Build your SaaS
For years the advice was "don't build in a crowded market, you'll get crushed." That advice assumed building takes months. It doesn't anymore.
Sol is an execution beast:
- Nick Saraev gave it image gen, video gen and hosting through MCPs. It created 80+ products with stills and video ads and published a showcase site. In about 10 minutes.
- Nate Herk got a finished, edited video in his own cloned voice from one prompt. No human touched a timeline.
One prompt in, a finished edited video out.
So build the thing. Pick the crowded market you know best, ship the version with an agentic angle (a CLI your users' agents can drive, an MCP server, an API-first core), and get to distribution while your competitors are still writing their spec.
One warning before you burn money: Ultra mode spawns sub-agents that inherit Ultra, and your budget evaporates. Herk's video burned 458M tokens, about $318. Launch week was full of "hit my $200 limit in 4 hours" posts. Ultra is a demo setting. Run High.
2. Growth hack your SEO
The game is distribution, and SEO is the channel where agents have the biggest unfair advantage right now.
Most people paste "write me a blog post" into ChatGPT and publish it. It doesn't rank, because Google has read a billion pages exactly like it.
The growth hack is running the whole SEO operation as an agent loop, with the models routed by cost:
- Luna does the grunt work: keyword lists, clustering, competitor checks. Thousands of calls, costs almost nothing.
- Terra writes the drafts.
- Sol only gets the judgment calls: what to build next, what's missing from the top 10.
Step by step:
1. Install Codex CLI and pick your model:
npm i -g @openai/codex
codex
# then inside: /model → gpt-5.6-sol, effort High
2. Turn your Agent into your SEO Manager.
Register for DispatchSEO, it gives your agent tools like: keyword tracking, Search Console data, a queue the agent fills, and every article arrives as a pull request. Works with Codex and Claude Code. It's also open source, so you can self-host it for free.
My agent's daily report. One guide a day, and it's ahead of schedule.
3. Run the setup wizard: connect Google Search Console (free) and your site's GitHub repo. Now the agent reads real queries and impressions, not guesses.
4. Turn on the schedule: weekly keyword research on Luna money, daily builds where Sol ships ONE thing as a PR. You review, you merge.
A real PR my agent opened. I reviewed, merged, and the tool went live the same day.
5. Feed it what only you know: product knowledge, screenshots of your app, the exact words your customers use. That's the entire difference between content and slop.
6. Scale more. Connect your other sites, one DispatchSEO install manages all of them, and the same agent works every queue.
I run this loop on my own sites. A 2-month-old, no-DR site pulled 60+ Google clicks and 3.2K impressions in the last 30 days. My newer site is 19 days in: 6 guides and 3 free tools shipped last week, and impressions nearly tripled in seven days, 611 to 1,785.
ClockedCode's Search Console, last 30 days. 61 clicks, 3.19K impressions, and climbing.
3. Build free tools that earn links while you sleep
Small free tools are backlink magnets. Nobody links to your blog post about keyword research. People link to a free tool that does it. And "free X checker" searches convert, because the searcher has a job to finish right now.
A free tool for Claude Code users, built by my Codex agent. Traffic is traffic.
The recipe:
- Use Luna to mine tool-shaped keywords: "X generator", "X checker", "X calculator". Low difficulty, clear intent.
- Have Sol build it as a single page. Free, no signup wall. The tool IS the marketing.
- Put it on your main domain, so the links feed the site you're trying to rank.
- Link to it from every related article, and from it back to your product exactly once.
One good free tool out-earns ten blog posts, and it now costs a prompt and a code review.
4. Sell the result, not the tool
Remember play 1: building got easy. That cuts both ways. If anyone can build the tool, the tool stops being what people pay for. People pay for the result.
Danny Postma's HeadshotPro is the perfect example. Underneath it's an image model anyone can run, and there are a hundred free "AI photo" tools. He doesn't sell any of that. He sells "professional headshots by tomorrow, no photographer, $29", and that framing built a seven-figure business while the model underneath became a commodity.
Not one word about the model underneath. Just the result and 3,518 reviews.
Why should you listen to me?
I build DispatchSEO, the tool from play 2, and everything in this article runs in public. The repo is open source, the sites my agent manages are live, and the search data above is mine. In a month I'll post the next chart, whichever direction it goes.
You don't have to trust me. You can read the commits.
Everything in this article runs from this repo. Read the commits.
Run these plays in order and they stack: the visitors compound month after month, and if you sell the result instead of the tool, those visitors turn into revenue.
Sol made building cheap. Distribution is where the money moved. Point your agent there.
If any of this sounds useful, try it yourself, the open-source version costs nothing to self-host.
Or take the hosted route if you don't want to touch Docker, and see what your chart does in a month.
Top comments (0)