High End and Workhorse Models Appear: GPT 5.6 Sol and Grok 4.5

OpenAI released their Mythos/Fable competitor GPT 5.6 Sol. Word is that it’s not quite AS good at the most difficult tasks as Anthropic’s Fable 5, but for almost all tasks it’s just as good and 70% cheaper to run and faster. xAI released Grok 4.5 that is now in the hunt for the top. It’s not QUITE there, but it’s allegedly very good and a LOT cheaper. Cheaper per token and more efficient in getting work done using less tokens.

We are well into the “labor pyramid” of AI models. You have your expensive human solutions architect, supported by a cheaper project manager, then a set of mid level workers and some cheap newbies. Nobody staffs a project with just the highest priced most capable humans. Likewise with AI these days.

I wanted to make significant refactor of my personal website. I used GPT 5.6-Sol to take my long list of changes I wanted to my website and asked it to evaluate it’s use as a public portfolio and my own personal operating os/dashboard for my apps etc.

I told Codex on my Linux box to setup the ability to run Cursor with Grok 4.5 medium. It’s wonderful how you can just direct these things that way. I didn’t have to know how to set that up.

I gave Sol’s output to Codex using GPT-5-Terra (the not as expensive, use it for most work a day tasks). I had Terra act as the supervisor giving the work to Cursor/Grok 4.5 medium. I setup a loop. Here’s the mission, use Grok for the work, check his work, if it doesn’t meet mission requirements, direct Grok to fix it. Keep going until everything is complete. And it did.

Then I asked Terra - How’d Grok 4.5 Medium do as a worker?

———————————————

• Grok 4.5 Medium is capable, but I’d rate it as a supervised workhorse rather than an autonomous one.

In this task it produced a strong first implementation: coherent visual refactor, correct card structure, responsive fixes, and clean archive
pages. It also responded well to narrow corrective prompts.

Its weaknesses were operational and security-detail oriented:

• Broad prompts sometimes appeared to complete without a usable report, so I had to split work into smaller slices.
• It didn’t reliably recognize workspace/write boundaries until explicitly probed.
• Its first “sanitized” Cloudflare pages still included credential-path and token-header examples—technically placeholders, but contrary to
  the requirement. It fixed them cleanly once the violation was named.

My practical take: excellent for fast implementation, refactors, CSS/UI work, and iterative fixes. Use it with explicit acceptance tests, a
small initial write probe, and mandatory human/agent review for secrets, deployments, permissions, and edge cases.

———————————————

Result? All the changes were made as desired.

I think I’ll try Grok 4.6 High next time.

I’m not quite sure where I’ll eventually draw the line between GPT 5.6 Sol and Terra.

Oh - I also signed up for the $20/mo plan for Cursor to try this out . The $30/mo Super Grok was a bust. I got one day’s worth of coding with Grok Build and used up 30% of my MONTHLY quota.

I’m setting up my own orchestrator with roles now and each role will have a list of coding harnesses/models to use given a priority and given my limits from subscriptions.