Let's discuss models and subscriptions. A constantly changing environment so this is my "as of 9/3/26"

Context - I'm currently between jobs and using the "down time" to build my skills and build a viable AI Employee Factory. I'm essentially running a one person firm. Eating my own dogfood. Building the platform to build the agents, building agents to do real work in ANTICIPATION of client work. I'm basically working 12 hour days 7 days a week so my LLM usage is a LOT more than normal.

Homebase for me remains OpenAI. ChatGPT has the most deep, rich and meaningful memory of me so our talks, planning, etc is required. Claude is excellent too, but I hit my limit on Claude often enough that he doesn't have all the context on me that ChatGPT does.

I'm at the $100/mo sub for OpenAI and have been for a long time. I find them far and away the most generous. I use Sol medium as a default, and Sol Max when I really need to solve a hard problem or follow a long and complex development task. Luna Max is the workhorse model that I have Sol create sub agents to do the work with Sol as the supervisor. It used to feel "infinite" but for a month now I find myself having to husband this resource. I'm not sure if their limits lowered or my usage soard. OpenAI dropped the price of Luna so low, that frankly, even the Chinese open weight models aren't cheaper given the amount included in the sub.

Computer control and browser control used to be what I used Claude for but OpenAI works much better now.

Claude is next. Fable 5.1 (which just came out) is indeed the smartest model. I'm not currently enthused with Opus 5 nor Sonnet 5 given my use of OpenAI above. I pretty much only use Fable and have it create output that I take to other models to execute, or direct Fable 5 to sub out to OpenAI models. That works, but I think it uses up too much Fable 5.

As I said above, I use Codex/CoWork more than Claude Cowork these days becuase, while Claude had the features first, I think Codex is better (for now).

I used to get by on the $20/mo sub but went to $100/mo for Fable 5. This sub is the one most likely to change - downward. Anthropic just isn't nearly as generous as OpenAI on how much computing you can do for the money. And while Fable 5.1 is indeed the smartest for my work, Sol Max is plenty smart.

Gemini was forever a disappointment. I subscribed at the $20/mo tier since it gave my Google drive space that I needed and just happened to have AI features. I do like Google's deep research best and Notebook LLM is wonderful. Antigravity was nothing special when it came out, but it's nice enough now. The big change was Gemini 3.7 Flash which is FINALLY a capable coding model and it's WICKED fast, and even the $20/mo sub had plenty of usage. 3.8 Flash is out now and is a bit better. I went up to the $100/mo tier because it saved my $20/mo on my Youtube Premium subscription and has $40/mo in Google Cloud credits that I can put to use. Now I have "near infinite" usage of 3.8 Flash. I rely on this when I'm having too cool off on OpenAI/Anthropic because I'm running out of usage on my sub. I've even come to like Gemini 3.1 Pro for things OTHER than coding. It's a terrible coding mode. Wonderful writing model and pretty good at planning and code review.

Don't know if I'll keep at the $100/mo level. But I do like having one of my subs that has "infinite" capacity.

I throw $10/mo to an Open Code Go subscription. It's basically "$10 in, $60 out". Nothing near as generous as OpenAI, but it's a nice way to have cheap access to the open weights models to play with and keep tabs on developments in that space. I blew through half of the month's allotment in 30min or so using Kimi K3, the most expensive "Fable 5 competitive" open weights model. I won't be doing that again. I will use it for GLM 5.3 Flash a wonderful open weights model at a fantastic price.

DeepSeek V4 Flash is so cheap it's practically free. It's just not capable enough for me to use as a coding work hourse. DeepSeek V4 Pro is good, and low cost but not "so cheap it's basically free".

I like GLM 5.3 Flash so much, and I burned through my month's Open Code Go after the first week - that I just pay the API price via OpenRouter.ai. I have it do significant work and an hour later, it's cost me ten cents. It is not quite in Gemini 3.8 Flash territory and via a sub, I have infinite supply. Thus I'm not using GLM 5.3 Flash as I might were I not stocked with Gemini.

Grok? Love grok as a conversational model with it's "let adults be adults" guard rails and live access to the x/twitter stream. It's just not price competitive. Neither the Cursor sub at $20/mo nor the Grok $30 mo come with much Grok 4.6 usage. I might dip my toe back in, in order to play with Grok Bot. But Grok models are good, not "the very best" but fail -- for me -- on not being price competitive. I'm hopeful of what the Cursor/xAI combination will produce in the future.

Local models? Not for coding. Not for me on the machines I own, and the needs I have. Sorry for all the Qwen 2.8 27B fans out there. You do you, this isn't for me (yet).

ibm-granite/granite-embedding-small-english-r2 - I use this local model in my sqlite-mem project that gives my Folder-Chief project a rag database. http://github.com/leebase/folder-chief

Coqui TTS for local voice generation. I have Telegram hooked up to my Chief and if I send a voice message, I get a voice response. It's not lightnig fast, but it's nice and no monthly subscription charges.

I do my best to pick the model for the job. I can't affort to use Sol and Fable for everything. But unfortunately, I know -- FOR MY USES -- the smallest, cheapest models aren't up for every task.

I know this was long. But this is my current perspective. Love to hear yours.