A Coding Subscription Tier List
Brokk has stopped publishing its Power Ranking after larger labs (finally) picked up the torch of useful coding evals [1], so here's my informal ranking of today's models (and subscriptions) through the lens of writing novel static analysis tooling in Rust [2].
Value for your Money
Obviously, the main reason to go through the effort of subscribing is to save money. Frontier models are fantastically expensive now, including at Chinese labs and third party open-weights providers. 2026 models are just bigger than they were ago, and apparently massive margins on inference is something that enterprise customers will happily pay.
The first way to look at how much value you get for your subscription is, "What would my tokens cost if I paid list price for them?" To answer this, I normalized each lab's top plan to what you'd get if you could magically pay $200 for each. xAI is the only one whose top plan is more, at $300; Zai's is a bit less at $168, and Qwen's is much less at $80.
The fine print: rather than include ridiculously huge ranges for DeepSeek, Zai, and Qwen, I've constructed this with the assumption that your usage is 100% off-peak. If you only use these models interactively during American working hours, this will be reasonably accurate; if you like to leave your agents working overnight, it will not be.
And as is well-known, none of these labs actually publish the numbers we're interested in, so we have to reconstruct them from user reports. The numbers I've used here are ballpark-consistent with what SemiAnalysis found in June, and with my own logs from OpenAI, Anthropic, and Moonshot.
(Finer print: Zai actually does publish their subscription discount, but it varies across input, cached input, and output tokens.)
Productivity for your Money
The other way to look at these subscriptions is, "how much can I get done with each?" Raw token counts are misleading; Claude models usually need more tokens to solve the same tasks than most other models, while GPT models usually need fewer. This is not a small difference! The difference between most- and least- efficient models is close to 3x!
To account for this, I've I've estimated task-level output from the DeepSWE data, which includes token counts per task and reasoning level. Again, normalizing to an imaginary $200/m level playing field, here's how many DeepSWE tasks you can solve with each subscription. Where the data is available (notably, not yet for Qwen 3.8 Flash Next), I've included both the flagship smartest model, and the best smaller/cheaper model from the same lab. I've also attempted to show the reasoning level before the "bend in the knee" of diminishing returns.
The headline result, and why OpenAI stands alone in S Tier, is how badly Luna stomps everyone else for value. This is a log scale graph: Luna solves more than 10x as many tasks per $200 of subscription, than anyone else! And if you take away the second most-efficient (Luna's big brother Sol) and 3rd (GLM 5.3 Flash) it's over 20x!
So yes, it is apparently simultaneously true that Sam Altman wishes OpenAI had more compute, and also that they have enough compute that they can afford to sell a subscription where you can solve more problems with their expensive flagship Sol, than you can with anyone else's cheap-and-cheerful model. And with Luna you can solve more than 10x as many.
I'm not an OpenAI fanboy, but that's just a remarkable flex.
The rest of the rankings
I'll post a more detailed analysis on Monday, but besides the quantitative value where OpenAI just mauls everyone, here's why I've ranked the models where I have:
- A Tier: Fable is the smartest model available today, and Claude Code is the best harness. No, Sol isn't as smart. No, GLM 5.3 Flash isn't as smart. (LOL.)
- B Tier: K3 is the smartest model that doesn't sometimes reject my requests for being too suspicious, and Zai has the best smart/cheap model subscription outside of OpenAI. DeepSeek is the "you'll probably use this if you max out your subscriptions and have to pay API prices" option.
- C Tier: Grok 4.6 isn't quite smart enough.
- D Tier: Qwen 3.8 isn't smart enough, and their subscription is the stingiest.