Hey @mid_gen do you use the advisor setting with Claude Code? I've been tending to run Sonnet 5 on high with Opus 5 set as advisor and it works fine, but I don't really have much of a feel for what these settings really change and when to use one thing or another.
-
Notice regarding image hosts and attachments. See here.
Claude code or something else?
- Thread starter Cromulent
- Start date
I haven't tried it tbh, I just use Opus as I never have a problem hitting usage limits. If I was trying to be more conservative with my usage I'd probably use Sonnet with Opus as advisor.
I ended up working with Sonnet5 as my only model for a few days last week without noticing. As the models evolve, these cheaper ones become 'good enough' for most tasks.
I've actually just cancelled my Claude sub as it was going to renew imminently, and am going to give Deepseek a go and see what API pricing works out at.
I tried it out when it was free. I didn't really notice any quality difference to Opus. Although that probably depends on what kind you are giving it to do; I understand that it's significantly better at "end to end" tasks (i.e. build a whole project) whereas I tend to give the AI defined subtasks.
It is good, but prohibitively expensive. It’s really for when you’ve got tasks to do where you’re happy to burn 5 figure sums on inference. It eats through even Claude Max 20x in minutes.
I enjoyed using Cursor a lot last year, just started using it again as I get to grips with and test some of the newer models.
I had to write up a website discussion doc/brief for an ecommerce web client recently who have been a two year slow burn. They approached me ready to actually pull the trigger, but all of our communication, notes, and file assets they'd sent me were scattered all over the place, it was a bit daunting trying to get myself back up to speed with it.
So I tested Gemini in Workspace and used the Gemini UI in Google drive to attach all the clients emails, meeting notes, image assets, and excel sheet with inventory list to revise the brief and highlight discussion points we need to finalise to solidify the initial build brief.
Once I had the document prepared, I signed back up for Cursor for $20 and created a gorgeous 1 page website to act as a reference document for the meeting. It one shotted about 90% of what I needed, a few more hours of tweaking and I had a very professional website / doc I could share with them and they were very impressed.
I'm quite looking forward to exploring how I change my workflow for creating a site like this in the modern AI age over the next few weeks.
I had to write up a website discussion doc/brief for an ecommerce web client recently who have been a two year slow burn. They approached me ready to actually pull the trigger, but all of our communication, notes, and file assets they'd sent me were scattered all over the place, it was a bit daunting trying to get myself back up to speed with it.
So I tested Gemini in Workspace and used the Gemini UI in Google drive to attach all the clients emails, meeting notes, image assets, and excel sheet with inventory list to revise the brief and highlight discussion points we need to finalise to solidify the initial build brief.
Once I had the document prepared, I signed back up for Cursor for $20 and created a gorgeous 1 page website to act as a reference document for the meeting. It one shotted about 90% of what I needed, a few more hours of tweaking and I had a very professional website / doc I could share with them and they were very impressed.
I'm quite looking forward to exploring how I change my workflow for creating a site like this in the modern AI age over the next few weeks.
Have stuck with OpenCode and OpenRouter this week.
Interesting process trying out different models.
I've spent $4.50 ish.
Majority on GPT 5.6 Terra which I've quite liked for planning work. It's *really* fast, nice to have discussions with. Once I have a working plan, Deepseek V4 Flash for the implementation.
Interesting getting used to making per-token charging work. Short answer is it can absolutely work, with a bit of practice.
Interesting process trying out different models.
I've spent $4.50 ish.
Majority on GPT 5.6 Terra which I've quite liked for planning work. It's *really* fast, nice to have discussions with. Once I have a working plan, Deepseek V4 Flash for the implementation.
Interesting getting used to making per-token charging work. Short answer is it can absolutely work, with a bit of practice.
Whoops, don't try Gemini 3.6 Flash! Burned £6 in a few minutes.
Honestly Deepseek V4 Flash is doing sterling work at the moment, think I'll stick with this for now.
It is tempting to just take advantage of Anthropic's pricing....until it gets more expensive. But I'd really like to be able to move to a locally hosted solution.
Honestly Deepseek V4 Flash is doing sterling work at the moment, think I'll stick with this for now.
It is tempting to just take advantage of Anthropic's pricing....until it gets more expensive. But I'd really like to be able to move to a locally hosted solution.
There is probably no better deal atm than using GPT-5.6-Luna with Max reasoning via any subscription tier atm. At 5x pro, even running multiple sub agents at once etc in Hermes agent, it feels unlimited. Effectively at the pareto frontier for performance/price at API (bar Deepseek V4 Flash which just edges it) but with usage and cost multiplied as it is in the subscription.
Have to admit, I don't have the time to learn the foibles of Deepseek V4 Flash right now....so have gone back to my Claude Max sub.
Opus 5 is just that much more trustworthy I can just get on with with and not be faffing with tools so much.
It's good to evaluate the cutting edge distils for when the economics of self-host become more pressing, but for now, sure am just gonna make use of the subsidised inference.
Opus 5 is just that much more trustworthy I can just get on with with and not be faffing with tools so much.
It's good to evaluate the cutting edge distils for when the economics of self-host become more pressing, but for now, sure am just gonna make use of the subsidised inference.
I use Claude Code at work and for programming, it's definitely the best right now. I've also been trying out DeepSeek V4 Flash locally over the last week for general use and its been insanely competent. Gave it my pfsense config to audit and also diagnose an issue and it gave me effectively the same response that I got from ChatGPT