Opus 5 vs GPT-5.6 Sol: The Same Build Prompt, Side by Side

Most model comparisons are throwaway. Somebody picks a toy app nobody will ever use, runs it twice, and posts a scorecard. This was not that.

I needed a front-end prototype of a client-facing AI ops audit report. It is going to live at audit.clearmud.ai. So instead of inventing a test, I handed the exact same prompt to Opus 5 and GPT-5.6 Sol, in two fresh terminals, and let the winner become the codebase I actually build on.

The Cost Gap Showed Up Before Any Work Started

Before either model touched the prompt, I typed "hi" in both terminals just to get the meter logged.

Opus 5 on high: about 30 cents. GPT-5.6 Sol: about 5 cents.

Thirty cents to say hello. That was the first laugh of the day, and it turned out to be a preview of the whole session.

The Prompt, and What I Deliberately Left Out

The prompt was specific about constraints. Front-end only prototype. Not a production application. No real authentication, no real back end, no database, no payment processing, no scheduling integration, no PDF generation service, no multi-tenant logic. All data synthetic and hard-coded or mocked in the front end. All export and send actions had to be visibly non-functional UI states, not misleadingly styled to look real.

The stack was locked down too: React plus Vite, Tailwind CSS, shadcn UI components, Recharts for any data visualization, shipped as a single deployable front end with no server required. Sections included a synthetic client placeholder, a "what we heard" source material block, and opportunity candidates to score.

Then came the part that made this a real test. I listed what I was intentionally not specifying:

• Visual design direction
• Information architecture
• Copy tone
• Chart types
• Section order beyond the list above
• How confidently to score each opportunity

That judgment is what was being tested. The silence in the prompt is where you learn how a model actually thinks.

GPT Finished First, and Deployed Itself

GPT-5.6 Sol came in at 11 minutes. It ran its own QA. And then it did something I did not ask for: it deployed to Vercel.

I genuinely expected both models to spin up a local preview. Instead, GPT read the room. My best guess is it picked up on how much I have been deploying with Codex lately and just defaulted to that path.

To avoid biasing myself, I pasted the deployed URL into a browser tab, hit enter, and immediately tabbed away without looking. Opus was still working.

Opus 5 Took 27 Minutes and 1.95 Million Input Tokens

Then Opus finished, and I was not expecting what the meter said.

Twenty-seven minutes against GPT's eleven. Roughly 1.95 million input tokens. Opus 5 appears to be insanely token hungry on long agentic builds, and the number was so absurd that I made a note right there to go back at the end and verify my API meter before publishing anything.

Opus also handled delivery differently. Rather than deploying, it sent me to an artifact hosted through my Anthropic account. Neither model was told what to do here. Claude defaulted to its own platform. Codex defaulted to my most recent deployment habit.

Both Outputs, Reviewed Side by Side

GPT's build opened as a split-screen client brief, and it looked professional immediately. Clean navigation, a single flowing scroll, clear pain points pulled from the meeting, and an opportunity map. The map scored each opportunity across efficiency, revenue upside, revenue protection, and effort using a dot layout. Is it working exactly as intended? Kind of. I would make the dots larger and add an overall baseline score. But the packaging read like something I could put in front of a prospect.

Opus went with a very similar split-screen approach and a vertical menu bar that I actually preferred over GPT's layout. It had a nice accordion effect on the pain points, business consequence framing, and a different chart type for the opportunity map that I could not name off the top of my head. Honestly, it felt more detailed.

But the coloring was very basic, and the overall direction just did not land for me. I did not give either model our palette or a Clearmud front-end design skill to follow. I wanted best judgment. Opus's judgment produced something that did not feel professional enough for a client-facing artifact.

That is a taste call, not an objective quality finding. I want to be clear about that. It is not my style, and I said so on camera. Had the packaging been a bit different, I might have leaned toward Opus.

The Thing Both Models Missed Was My Fault

Neither output recommended what to actually build per pain point. No tools section, no solutions mapped to each problem.

I was hoping one of them would decide to include it on its own. Neither did, and that is on me for leaving it out of the prompt. If you want tool or solution recommendations tied to each pain point, say so explicitly. Do not expect either model to infer it.

I Checked the Meter, and the Meter Was Wrong

The cost gap looked too extreme to publish as-is, so I stopped the review and did the boring thing.

I screenshotted my open source model meter, pointed the model at both sessions, and asked it to cross-reference the numbers against the official API pricing docs for both providers.

The verdict: the Claude figure was roughly three times higher than it should have been. GPT's published Sol rates were already correct, so that number stayed unchanged. Applying the fix brought the Opus total down to just under $13.

I committed the fix and merged to main on camera.

I am glad I checked. A shocking number is usually a broken instrument, and if I had shipped the raw figure I would have been wrong in public. The meter is open source, by the way. It is on the Clearmud GitHub repository.

The Verdict, and Why It Is Not "GPT Beats Opus"

In my book, Codex is the winner here. I thought for sure I was going to pick Opus, and it went the other way. Opus took nearly three times as long at roughly three times the cost, and GPT's packaging looked more professional for a client-facing artifact. Efficiency plus presentation won.

But this is not a one-versus-the-other conclusion. My actual workflow is going to use both:

• Anthropic models for ideation, creation, and prototyping. That is where they are always my go-to.
• GPT for production efficiency. Once I am ready to move toward production, I trust it to produce a production-ready codebase faster and cheaper.
• Both models for security audits. I run my audits through GPT today, but the benchmarks I have been seeing suggest Opus 5 does a very good job at security work. New plan is to run both against the same repository.

What the Prototype Is Actually For

The output is a starting point, not a deliverable. This is not for the public. It is something I will use internally once we perform the audit, plus a dashboard for me and the prospect who books the walkthrough.

The real product is the pipeline behind it. Have a fluid, natural conversation with a prospect where I am listening first. Take the transcript or the audio file from that meeting, drop it into the platform, and have Claude or Codex or both analyze it, define what we heard, define the pain points, and build the opportunity map. That is the workflow. The prototype is just the surface it renders on.

If You Want to Run This Test Yourself

• Benchmark on something you actually intend to ship. A test you have no use for teaches you nothing about which codebase you want to live in.
• Be exact about stack and required sections. Be deliberately silent about design, information architecture, tone, chart types, and ordering.
• Log a trivial prompt first to baseline per-model cost before the real run.
• Cross-reference your cost tooling against official pricing docs before citing any number publicly.
• Judge the packaging against the audience, not against your own preferences in a vacuum. Mine is client-facing, so presentation outranked extra detail.
• Split the workflow instead of picking a side.

I am not an AI expert. I am building in public and sharing what actually works.

Watch the full head-to-head, including the live pricing correction: https://www.youtube.com/watch?v=MPqtbmz791g