DISCLOSURE: If you buy through affiliate links, I may earn a small commission. (disclosures)
GPT 5.6 came out several weeks ago and the initial benchmarks look good - capable, fast, and cheap.
Claude and Claude Code have been my daily driver for the past year, and I thought it was a good time to test alternative models + harnesses to see what the hype is about.
I've been experimenting with GPT 5.6 + OpenCode as my daily driver for coding and my personal assistant, so here I'll share my first impressions.
Why Try Alternative Models + Harnesses?
Obviously the benchmark results are what triggered this experiment. A faster, cheaper, and just as capable model is a no-brainer for swapping out the brains of my setup and is in line with my predictions that AI will be a commodity sooner rather than later. Plus cost is becoming an increasing issue at work and I'm spending ~$50k per year on AI - so if I can get better ROI that's better for me and the company.
But I've been thinking about building a model-agnostic setup for a while because the AI labs tend to one-up each other every few months. So an ability to have one setup and swap out different parts of it when things change is, in my opinion, a competitive advantage. Anthropic definitely felt like it was winning by a long shot the past year or so, but I think Fable was a bit of a letdown - very capable but not very practical for 99% of applications.
What I've already got that's cross-harness:
So here I wanted to try a capable harness that was model agnostic and allowed me to use some of the heavily subsidized first-party subscriptions. That landed me on OpenCode (though this is also part of the experiment, so not tied to it).
I'm not fully sure if sharing my setup is that useful to people, but my understanding is that harnesses and applications do tend to impact how good a model feels, so it felt worth putting in here.
GPT 5.6 vs Claude Fable + Opus
GPT pros:
- Faster - Things I would kick off and wait 30 seconds to a few minutes for with Claude come back almost immediately in most cases. I know how fast software can be, so this shouldn't surprise me, but I've been consistently impressed with the speed of inference on these models. Maybe this is the norm and Claude got me used to slow turns, idk.
- Cheaper - From light glances at my subscription usage (which itself is not a very precise measurement), it seems like the models are 2-3x cheaper and I'm using Sol a good bit more than Fable since I found Fable to be way too expensive and extremely slow.
- Task + skill following - GPT is reliably reading and following my skill instructions. This shouldn't be a surprise, but I guess I got used to how little Claude was - even when I'd explicitly ask it to. This is mostly a positive but has required a bit of a learning curve as now I'm finding it is following instructions more than Claude, so I need to tweak them to allow for more flexibility or fix areas that may actually be inefficient as written.
GPT neutral:
- Quality - Feels about the same as Claude. GPT seems to have a little less personality but seems just as accurate. Caveat: I haven't done a large volume of coding with it yet, so I'll need to see.
GPT cons:
- Writing is mechanical - The writing has been a lot more mechanical. This is fine for the most part, but if trying to draft a post or small communication, it definitely reads as less human. But in some ways this is good because it kind of feels more like using an impressive tool and then I do the human translation myself.
OpenCode vs Claude Code
OpenCode pros:
- Open source - Doesn't get me that much today, but I'm generally in favor of choosing open source when comparable with a closed-source version. Typically it has less lock-in, is cheaper, and is easier to modify. The AI race makes this a little different because everything is bleeding edge, so obvs the labs with $100Bs are going to move faster than open source, but I'm typically okay with moving fast on stable infrastructure.
- Fewer rendering issues - Claude Code has had sporadic rendering issues which make the harness annoying at best and unusable at worst. I think it has gotten better in that I've run into these issues less, but still some rendering bugs / issues here and there. OpenCode apparently doesn't have them as much and anecdotally I haven't seen them yet, but I have to caveat it's only been a week so we'll have to see.
- Model agnostic by default - I hear you can get the first-party harnesses to hook up to other providers but haven't tried it myself. But OpenCode has a nice picker for various providers and I believe most use whatever protocol OpenCode uses under the hood, so changing out models is just a matter of a few keystrokes, which is great for experimentation when the next model comes out.
OpenCode neutral:
- Quality + cost - I hear harness has a decent impact on quality and cost, but I haven't spent enough time / money with it to tell for sure. I think after a month using it at work I may have a better idea and data to share, but it's too early to say for now.
OpenCode cons:
- More work to set up remote access - Claude Code has built-in remote control which I've used for the past several months to power my personal AI agent and to build software from my phone. OpenCode does offer a similar experience which is great, but it's more manual and I don't know if I'd recommend it to someone less technically savvy - you need to have a secure network port, automate setup, and think a bit about machines + terminals. May work on making this easier / simpler, but it's working for me for now.
GPT 5.6 + OpenCode Is My Daily-Driver Experiment
For now, GPT 5.6 + OpenCode seems faster, cheaper, about as smart, and better at staying on task for long-running tasks. That makes it a good daily driver for me, while Claude is still the likely escalation path for subjective writing, UI work, and difficult trust-critical tasks.
- Model - GPT 5.6 terra for typical work (like Opus), Sol for harder work (like Fable)
- Harness - OpenCode terminal for focused work, web for accessing on the go (secured via Tailscale)
- Subscriptions - Open AI max plan in Open Code, fallback to Claude in Claude Code
This is still early though so will have more data in about a month to make a more solid decision but I'm guessing OpenCode's model agnosticism will be its killer feature.
Next
I'm going to keep running GPT + OpenCode for a few more weeks and see how it feels.
If you're experimenting with a similar setup, I'd be curious what you're running and how it's working out for you.
If you liked this post you might also like: