@opencode not for botsbench.com investigation tasks. some fun lessons in this one picture.
1
0
5
3,960
@opencode yeah. cost per session alone is tricky. if people keep paying the higher session cost, that is usually a quality signal.
0
0
3
708
@opencode I don't get why cost/session means the model is good
0
0
1
51
@opencode this is quite an interesting insight.
I was actually planning to put this model on test against other models like Minimax M3, Deepseek v4 and you just published this one, so it is actually better.
do you also think that changing the harness also changes this stats quite a bit?
I was actually planning to put this model on test against other models like Minimax M3, Deepseek v4 and you just published this one, so it is actually better.
do you also think that changing the harness also changes this stats quite a bit?
0
0
0
136
@opencode yeah, it's great, but it eats the monthly quota in a day of coding.
0
0
1
155
@opencode Well, we all know there's a trade-off between cost and performance. If a model is Pareto optimal, the most expensive option will always perform the best.
0
0
0
58
@opencode Mostly I just switch between GLM 5.2 and MiniMax-M3 in my session. Definitely GLM is working great, to be frank, I don’t use OpenAI and Anthropic, just use Opencode Go and Openrouter for my side projects.
0
1
44
4,405
@opencode That's exactly why we build compression tools, to make using these models more accesible and cheaper.
1
1
8
373
@opencode So how should we interpret this, GLM 5.2 is "sooo expensive" or it is finally model that let you do serious work? This graph is not an answer but start of the question
1
1
9
3,410
@opencode It's the first time I'm flying through my monthly limits 😭
1
0
9
2,765
@opencode unpopular opinion but the model is quite expensive and overall pretty bad. Literally DeepSeek-V4-Flash (yes flash, not pro) performs better for my use case and on simpler prompts is even better than glm 5.2 at a fraction of cost. glm 5.2 is SUPER overhyped. crazy.
1
1
7
966
@opencode Open source costs also climbing up. Looks like both Claude and OpenAI would be doing a rate hike soon.
0
0
3
2,812
@opencode this is the most backwards useful metric i’ve ever heard and i’m obsessed, any edge cases where high cost per session was just a cash grab?
0
0
2
368
@opencode it's a great planner and reviewer subagent but too expensive
can't believe i'm saying this but yea defo not cheaper than its peers
best to pair it with cheaper subagents to run tools on its behalf
also best when orchestrated from a cheaper primary agent
can't believe i'm saying this but yea defo not cheaper than its peers
best to pair it with cheaper subagents to run tools on its behalf
also best when orchestrated from a cheaper primary agent
0
0
2
76
@opencode cost per session is really measuring verbosity × price. GLM 5.2 at $2.31/session might just pad its reasoning. a terse model that's right in 500 tokens doesn't score here — not because it's worse.
0
0
2
646
@opencode I'll give it a try to audit my codebase in a couple days when my old subscription is done. I still use 5.1 daily. Curious to see what 5.2 can come up with!
0
0
1
2,471
@opencode Fix Opencode Go plz! It often goes rogue reading 1 file repeatedly, exhausting credit.
I've used GLM-5.2 from other providers, no such issue.
I've used Opencode with other models, no such issue.
I've used GLM-5.2 from other providers, no such issue.
I've used Opencode with other models, no such issue.
1
0
1
617
@opencode It's more expensive than those other Chinese models. But it's cheaper than similarily smart American ones. When wants to spend more to solve a difficult task, GLM 5.2 can save money as compared to utilizing ChatGPT or Claude. What's your point exactly?
1
0
1
224
@opencodel@thdxrp with what I see daily, but worth adding the other end of this chart.
I run the free DeepSeek tier through OpenCode for anything simple. Installing a package, basic Linux config, quick file ops. No friction, no errors, does the job exactly as asked. Costs me nothing.
The interesting part is not that GLM 5.2 commands a premium for harder sessions. It is that DeepSeek free tier now handles tasks that would have needed a paid frontier model eight months ago.
The whole stack moved up. The bar for what counts as "simple enough for free" just keeps rising.
I run the free DeepSeek tier through OpenCode for anything simple. Installing a package, basic Linux config, quick file ops. No friction, no errors, does the job exactly as asked. Costs me nothing.
The interesting part is not that GLM 5.2 commands a premium for harder sessions. It is that DeepSeek free tier now handles tasks that would have needed a paid frontier model eight months ago.
The whole stack moved up. The bar for what counts as "simple enough for free" just keeps rising.
0
0
1
939
@opencode I personally spent 50 bux to try it, was rather impressed with the results.
0
0
1
678
