Fran

[Fran @FranVidiCheck]

@yuhasbeentaken Besides speed and lower cost

[Fran @FranVidiCheck]

@yuhasbeentaken Why would flash be better at those things? It's the worse of the two models... No?

[Fran @FranVidiCheck]

@TimJayas Gemini 4, maybe, 3.5 absolutely no chance.

[Fran @FranVidiCheck]

@trikcode That's an "agent loop", but we referr to things like claude code, codex, antigravity , ect. As agents, which is even more confusing, because those are harnesses...

[Fran @FranVidiCheck]

@zeeg That is a different use case though. He was referring specifically to using "be through" in the use case of writing commit messages. Of course anyaddition to prompt will change the nature of an AI, but is it really necessary for the specific use case?

[Fran @FranVidiCheck]

@Prathkum Isnt GPT 5.5. Medium both cheaper and smarter thougg?

[Fran @FranVidiCheck]

@Angaisb_ Note, you are comparing medium thinking effort to max thinking effort. We don't know if GLM 5.2 beats 3.5 flash at max thinking.

[Fran @FranVidiCheck]

@levelsio @jonoringer What is WAL mode?

[Fran @FranVidiCheck]

@MrAhmadAwais Aren't there SOTA models like GPT 5.5 medium that are both better and cheaper than GLM 5.2?

[Fran @FranVidiCheck]

@rauchg Missing comma here: "so shadows stay subtle."

[Fran @FranVidiCheck]

@noman_huge @CircleCI Current looks better imo

[Fran @FranVidiCheck]

@rachpradhan What are some alternatives to this and how does this differentiate?

[Fran @FranVidiCheck]

@AiBattle_ Where do Qwen 3.7 Max and Kimi 2.7 code stand on this bench?

[Fran @FranVidiCheck]

@pankajstwt Cost difference?
Let's not ignore the real benefit of AI - affordability at scale.

[Fran @FranVidiCheck]

@theo It would be interesting if authors of benchmarks built skills/instructions/harnesses that bump the scores maximally so that we can diagnose what exactly increases the scores in the specific benchmark.

Obviously, the instructions should be general, not task specific.

[Fran @FranVidiCheck]

@dabit3 @YayaSoumah @DevinAI Personal, free plan

[Fran @FranVidiCheck]

@dabit3 @YayaSoumah @DevinAI The post says Devin Desktop is compatible with OpenCode, but i don't see it as an option. Neither do I have codex, claude or cursor.

[Fran @FranVidiCheck]

@magical_ux @benln Pretty cool.
Feedback: I selected sign up with google and was still forced to enter my email and set a password, not cool.

[Fran @FranVidiCheck]

@TheAhmadOsman M3 for frontend/design. Qwen 3.7 Max for coding.

[Fran @FranVidiCheck]

@victoriawu77 B is best