~ 8 min read
At the beginning of June, I decided to embark on a bit of an experiment. I’d just come to the end of my Claude subscription for the month before a weekend so decided to cancel it and come back to it the next working week. I’m not using agents outside of working hours, so had no need for it at the weekend.
But then I got to thinking why use Claude at all? Why don’t I experiment with using open models and do it locally? This is something that I’ve talked about a number of times within videos on my YouTube channel. It’s been quite a while since I published those and models have come on leaps and bounds since then. I also wanted to be able to use tokens for other things as well and knew that Anthropic didn’t allow that with the subscription I had for Claude. I’d also like to know that I’m not building a dependency upon any single provider.
Provider & Harness #
I quickly realized that being able to do this with my M1 MacBook Pro Max, even though completely capable with 64GB RAM wasn’t feasible in my circumstances because because I didn’t want to have my fans blaring at full speed all the time while I was using my MacBook Pro. I quite like it being quiet.
I also found with some of the larger models they took a long time to respond on my Mac, despite it being fairly performant. It was anywhere from 30 seconds upwards for a first response, which was just completely unusable. I was doing this through Ollama, which I think possibly was adding a little overhead. But I don’t think I could have shaved off anything near to what I needed to be able to get a nice slick coding environment.
And so I went to look for other places in the cloud for inference. openrouter.ai is the one that I’ve settled on because I’ve heard of it the most and I liked the idea of being able to have access to a huge number of models. It’s also allowed me to experiment with using Hermes as well within the same account.
I have no idea which models are the best in terms of open ones. There’s so many right now - so I basically went for the most popular, which according to OpenRouter at this point is DeepSeek V4 Flash, which seemed fine to me. I’ve been using Claude Sonnet rather than any huge model up to this point and I figured that was a comparable replacement.
For my harness I’m using Opencode, Dax speaks a lot of sense and there’s no other harness on my radar that fits the bill. I started out with it working in vscode but found having it open fullscreen in ghostty a nicer experience.
Projects I’ve Built #
Throughout the course of June I’ve worked on three separate greenfield projects:
Dev News Site
The first one being n3ws.dev that allowed me to aggregate top/trending information from Hacker News, Github and Lobsters that were interesting to me for research. I used to get a similar daily summary from Mailbrew pulling together information from Reddit and Hacker News - I wondered if I could basically replace it with something of my own that I could then customize further. It’s a FastAPI app backed by sqlite and I used Hassan El Mghari’s Hallmark to create the initial design.
Added a ton of stuff to my minimal aggregator n3ws.dev
- Sorting (!)
- Made it sticky too
- An api so I can start getting a digest
- A load of backend boring stuff Thinking of filters next.
#buildinpublic
— Ian Wootten 🛸 ([[image or embed]][@ianwootten.co.uk])[6 June 2026 at 12:18] I put this together in about a week from start to finish - almost turned it into a mailing list using resend but decided that actually it was just a nice fun little project to have myself and that other people probably aren’t going to find it useful.
Desktop RSS Reader
The second project I worked on is a RSS aggregator. Yes, in 2026 RSS is still very much a useful thing. I’ve kind of got a bit fed up with Feedly that seems to be moving towards being a market intelligence tool, whatever that means. And I just wanted something very simple that I could pull together feeds. It’s sad to me that actually so many sites these days don’t have an RSS feed because it’s a really useful way of keeping up to date with people who you respect in the developer community.
I’ve actually seen a number of other developers working on similar tools to keep up with their own network. It is just a local desktop Tauri app that I’ll use myself. and haven’t actually gone too far down the line with it. It’s functional but I don’t know where I’m going to take it right now.
Show Notes for Podcasts
The final tool I worked on is a way of creating show notes for publishers of podcasts. I’ve been thinking of ways that I can leverage my https://despeech.com transcription API that I created a few months ago to better showcase it and this seemed like a a good way of doing that. Again, this is another FastAPI app backed by sqlite and kicking off with Hallmark. What’s mad about this project is that it’s essentially a fully fledged SaaS tool that’s pretty much ready to go. It has auth, accounts, uploads and a simple dashboard.
Having fun building a little show note generator for podcasts. Nice to see them all trickling through.
#buildinpublic #podcasts
— Ian Wootten 🛸 ([[image or embed]][@ianwootten.co.uk])[23 June 2026 at 15:44]
Token Spend #
Openrouter tells me that over June I’ve spent $16.64 over 5K requests and 282M tokens. 22M of that was me experimenting with pi as an agent so the costs of the projects are actually a bit less than that. So, less than my $20 Claude Pro subscription but I could see myself easily going over that if I was coding daily and running agents all the time. $20 is obviously peanuts in the grand scheme of things, but I’ve also spent my time back and forthing with Opencode to ensure it gives me something I’m happy with.
It’s worth mentioning that Opencode have their own optimised models through their own subscription Go, which seems very reasonable at $10/mo that I’ve signed up to compare against openrouter. It’s $5 for the first month - additionally if you sign up using my link we’ll both get $5 in usage credit 🙏.
Dev Experience #
My coding experience has been pretty much on par with what I’ve experienced with Claude already. For what I’ve been building, greenfield personal projects in a language I’m familiar with it’s been great - with me noticing little to no difference. YMMV.
It hasn’t been completely plain sailing though:
Early on with the buildout of n3ws.dev I was alerted to the fact a .git directory got exposed on the root of the site after publishing. This was actually down to a wide ranging COPY *
in the build process causing it being copied to a container image that hadn’t been caught by ignore files. This is something that comes from me reusing Docker images and not catching it in review rather than any fault of my setup. I’m not sure if Claude would of been able to catch it or a larger model maybe. I had it resolved in 30 mins of the original publish, but still.
The context window for Deepseek V4 Flash is 1M tokens, but I experienced slowdowns and strange behaviour as soon as I got to conversations above 100K on OpenRouter. This meant on certain days I was losing a large amount of time (and tokens) waiting for responses and trying to kill opencode. I found that if I immediately started a new session after a feature had been merged, I had a much better experience overall. NB: Having spent a short time using Opencode Go, it looks like the provider experience is much better there easily reaching windows over 200K without issue.
OpenCode experienced a number of rendering issues within ghostty, which when combined with the above made things pretty annoying when working on larger features. Once I’d
I’ve also missed being able to share a subscription with a typical chat interface. I dipped in to Claude free to ask a query every now and then over the month and only had a single session prompt me to upgrade.
Summary #
Overall, I’m really impressed - to get three side projects out of the door is no mean feat for any month. Open models are a far cry from when I was trying out tools like Aider a few years back, which really struggled with simple tasks.
I’d love to experiment far more with getting a better completely offline setup using open models so I could have a completely private environment able to be used for financial and more personal tasks. I’ll also definitely be diving into using OpenCode Go too.
For now, I’m going to continue rolling with Opencode with their Go service. I hope this inspires you to review your setup and possibly try an Open model or two.