Model Rogue

the signal is in the terminal, not the timeline.

less think pieces, more git clones.

[Model Rogue @modelrogue]

@scaling01

[Model Rogue @modelrogue]

@tdinh_me Three models came out claiming to be Mythos-level.

[Model Rogue @modelrogue]

@scaling01

[Model Rogue @modelrogue]

@anah_sahh I don't know some organic version of this shit

[Model Rogue @modelrogue]

Nothing is guaranteed. No amount of hard work can guarantee a certain outcome, certainty is impossible when nature itself is random. Sure, trying harder can generally give you a better shot at a decent life than most people, but success is never guaranteed, no matter how hard you work.

[Model Rogue @modelrogue]

@mark_k @SpaceXAI Planning to use a SuperGrok subscription for almost all my use cases at this point, since usage limits have apparently become a permanent part of life now.

[Model Rogue @modelrogue]

@tannerlinsley I like the new super app. I honestly don’t get what people are complaining about. Sure, there are features I’ll probably never use, but that’s fine.

[Model Rogue @modelrogue]

@n0w00j Looks interesting

[Model Rogue @modelrogue]

@sama What’s up with the usage limits? How can a model that’s more efficient than the previous one hit usage limits so much faster?

[Model Rogue @modelrogue]

@Av1dlive @sama What’s up with the usage limits? How can a model that’s more efficient than the previous one hit usage limits so much faster?

[Model Rogue @modelrogue]

@rezoundous I’m thinking of getting a SuperGrok subscription for most of my day-to-day use.

[Model Rogue @modelrogue]

@rezoundous There’s useful, and then there’s the usual “cowork, design sites, this killed Figma, this killed Adobe” nonsense.

[Model Rogue @modelrogue]

@jun_song And please, nobody say use “Sonnet” or one of the lower-end models, even those are hitting limits, so clearly something else is going on.

[Model Rogue @modelrogue]

Why isn’t anyone making videos on Muse Spark 1.1? Is nobody interested? I wanted to see how good or bad it actually is, but I can’t find a single proper review or video on it. Was no one given early access, or does nobody care enough to cover it?

[Model Rogue @modelrogue]

@trikcode 3.5 Flash

[Model Rogue @modelrogue]

@cb_doge https://t.co/4ZvbBvb8zA

I heard this as well

[BridgeMind @bridgemindai]

Grok 4.5 is NOT on the same level as Claude Opus 4.7.

Check out the BridgeBench lava lamp test.

Claude Opus 4.7 produces far better UI designs compared with Grok 4.5.

Grok 4.5 is a significant jump compared with Grok 4.3, but not on the same level as Opus 4.7.

[Model Rogue @modelrogue]

It feels like everyone has access to GPT-5.6 except me. Please tell me at least a few of you don’t have access either, so I can preserve what’s left of my self-respect.

[Model Rogue @modelrogue]

@d4m1n

[Model Rogue @modelrogue]

@guneysol Read the fine print

[Model Rogue @modelrogue]

Grok 4.5 just dropped, and it’s a quiet generational leap.

Not because it’s the absolute smartest on every chart. But because it kills the usual frontier model tradeoffs: smart and fast and cheap.

Why it feels like a real jump:
Scale + targeted training: 1.5T parameter V9 foundation (3x previous production models).

Heavy supplemental training on real Cursor developer sessions + ongoing RL from Grok Build harness. This isn’t generic pre-training, it grew up around actual coding/agentic workflows.

Reasoning efficiency: Uses far fewer tokens to solve the same tasks (4.2x fewer output tokens than Claude Opus 4.8 on SWE-Bench Pro).

The model finds shorter, smarter paths instead of verbose overthinking.

Raw speed: ~80 tokens per second. Feels snappier for daily use and agent loops.

How they kept it cheap:
Mixture-of-Experts (MoE) design means only the relevant parts activate per query.

Colossus-scale training infrastructure (built for this) + relentless efficiency focus from xAI. Result: $2 / $6 per million input/output tokens — often half (or better) the effective cost of comparable frontier models because it wastes fewer tokens.

This is what happens when a lab optimizes for usable intelligence at scale instead of pure benchmark flexing.

Builders win: you can run serious agentic/coding workflows without burning through budget or waiting forever.

The old era forced you to pick two out of three (smart, fast, affordable). Grok 4.5 is pushing all three.
Worth testing if your work involves code, agents, or high-volume reasoning. The floor just rose.

Sources
xAI / Elon Musk / Grok official announcements.

Early benchmark notes on SWE-Bench, token efficiency, and pricing.