Dr. Matthew Talmage

Investor, educator, advisor to HNW and UHNW individuals, enterprises, and schools. PhD in History (Roman/Classical Studies, Realpolitik, Power Dynamics, War).

[Dr. Matthew Talmage @MatthewTalmage]

Being unable to force high-reasoning on 4.3 (beta) on the web is dumb, @elonmusk and @xai.

My options on the web, not API, are either:
1. Use the Expert mode and use a dumber, older model (49 score on that benchmark).
2. Use Grok 4.3 (beta) and hope that the Auto selector decides that it thinks my query "warrants" high-reasoning; otherwise it spits out useless non-reasoning information (score of 31 on that benchmark).
3. Pay for API to force high-reasoning on Grok 4.3 (51 on that benchmark).

Or...

4. Pay for @OpenAI Pro to use GPT 5.5 Pro (60/59 on that benchmark), LESS than a SuperGrok Heavy subscription costs typically. That is, pay less money for a smarter reasoning model on the web.

[Dr. Matthew Talmage @MatthewTalmage]

@grok When asking Grok directly on the web, Grok doesn't give a clear and consistent answer to this question, so I wanted to try it here. On the web browser (not through API), I have access to: "Auto", "Fast", "Expert", and "4.3 (beta)". Which of these is the most powerful right now that actually uses reasoning? When I choose "Expert", I get the multiple agent style that looks like it's from earlier versions of Grok (4.2). Based on this site's benchmarks, which version does each selection on the web correspond to? https://t.co/LYgvcB5X9e

[Dr. Matthew Talmage @MatthewTalmage]

That's my point. There is no way to FORCE 4.3 (beta) on the web to use the high-reasoning effort (the 53-level config) because it's stuck on Auto. So the paying subscriber must either use 4.3 (beta) and hope that the Auto picker randomly decides that the query warrants high-reasoning, OR choose "Expert" which uses the older version of Grok. Neither of these are good situations. The user should be able to force, on the web, the high-reasoning effort on 4.3.

[Dr. Matthew Talmage @MatthewTalmage]

Except choosing 4.3 (beta) wouldn't NECESSARILY be the strongest current performance, because the non-reasoning version of 4.3 scores only a 31 on that benchmark I posted earlier, whereas the "high" reasoning version of 4.3 scores 53, and the user on the web would have to HOPE that the 4.3 (beta) option will actually use the "high" reasoning version vs. the stupid non-reasoning version, because there is no way to force 4.3 (beta) to use high reasoning on the web currently. Correct?

[Dr. Matthew Talmage @MatthewTalmage]

@grok Logically, then, choosing "Expert" switches to the PREVIOUS model's reasoning mode, NOT onto 4.3 beta's reasoning mode. That would mean your prior information on this thread was wrong.

[Dr. Matthew Talmage @MatthewTalmage]

@grok So if "4.3 (beta)" starts in Auto mode, then what in the world is the point of the mode selector having an "Auto" option if that option does the same exact thing as choosing "4.3 (beta)"?

[Dr. Matthew Talmage @MatthewTalmage]

...But the "Expert" selection on the website isn't clearly marked as "4.3". In fact, it's marked as nothing at all besides "Expert". "4.3 (beta)" is a totally different selection and there is no way to add on an "Expert" mode to 4.3 beta on the web. You can only set "High" on 4.3 via the API. So what exactly is "Expert" running on the web version?

[Dr. Matthew Talmage @MatthewTalmage]

@grok So why does the "Expert" option on the web invoke the multiple agent mode from 4.20? I thought 4.3 was single agent throughout?

[Dr. Matthew Talmage @MatthewTalmage]

@grok So if the "4.3 (beta)" option on the web is the one called "4.3 (high)" on that site, what is the point of the "Expert" selection option?

[Dr. Matthew Talmage @MatthewTalmage]

@grok Are you sure about this? That benchmark site shows a "Grok 4.3 (high)" and "Grok 4.3" with starkly different scores (53 and 31, respectively).