Grok 4.6 与 DeepSeek V4 Pro 同日实测：准 Fable 级能力，价格打到底

Today can only be described as a magical day.

Two things happened:

One was DeepSeek V4 Pro's official release.

The other was Grok 4.6 officially launching.

Both happened overnight, within a two-hour window.

And the amusing part is that both are models of similar parameter scale — one at 1.6T, the other at 1.5T — yet both, tonight, edged close to the Claude Fable 5 experience.

Then, after trying them out tonight, I voted with my wallet.

For the first time in my life, I dropped $300 on a SuperGrok Heavy membership for Grok.

At the same time, I turned around and topped up the DeepSeek API with 1,000 yuan (~$140) of credit.

A little something for each side.

And with these two releases, I think the LLM market is about to get a razor-sharp kill line.

This round will likely eliminate a great many players.

It's been barely three or four months since Claude was so far ahead that the gap felt hopeless; now leadership changes hands every few weeks.

So, at this crazy moment, I'm also updating my everyday AI lineup:

First: for my daily main Vibe Coding work, I'm moving to Grok Build + Grok 4.6 — I'll explain why later.

Second: ChatGPT and Codex.

They'll keep handling the hardest, absolutely-cannot-fail mega-tasks, plus all my daily office work, creative and content production, and browser/computer control.

Third: DeepSeek V4 Pro (stable).

Because of its unbeatable value, I'll hand it a large volume of automation and async tasks that don't need real-time feedback, plus some automation Agent work inside products I build, like AIHOT.

That's the combination I've settled on after tonight's reshuffle.

Who knows whether it'll change once new models land — but for now, this genuinely is the setup that feels best to me.

The reason behind this split is simple.

I've always believed LLMs face an impossible triangle — three things that are brutally hard to satisfy at once:

Performance.

Price.

Speed.

This is the triangle that, historically, almost nobody could fill.

People talk about the first two all the time, but the last one — speed — is constantly forgotten.

Because with speed, you can only truly feel it if you're using these models every day. It depends not just on the model, but also on the vendor's inference capability and compute reserves.

If, like me, you spend close to ten hours a day with various Agents working nonstop, speed is sometimes the single most decisive factor.

I can barely tolerate the speed of Codex in normal mode with GPT-5.6 Sol anymore.

A task taking ten-odd minutes is common; half an hour as a starting point isn't rare; some tasks routinely stretch to a full hour.

What I hate most is getting feedback every ten or thirty minutes — it forcibly chops my time into little fragments, and while I wait, bored out of my mind, all I can do is scroll X...

Which is why my X-scrolling time has exploded lately...

So lately I've been forced to keep Codex in Fast mode around the clock.

And the way my quota burns — I'm on the $200 membership, and after a couple of casual days, the quota collapses.

This is a very, very, very real situation.

It's the impossible triangle I've been facing daily for a long time.

The smart ones — Claude Fable 5, GPT-5.6 Sol — are usually expensive and slow.

The cheap ones are often fast, but their performance falls short; DeepSeek V4 Flash, for example, is fine for everyday use, but a mediocre coder like me still needs stronger model capability to raise my ceiling.

Fast and powerful? That's basically what the various Fast modes sell you, and they're anything but cheap.

This is the impossible triangle of LLMs — and it's also the industry's kill line.

But in the early hours of today, DeepSeek V4 Pro and Grok 4.6 ripped a hole in this impossible triangle — from two different directions.

At current pricing, DeepSeek has squeezed Fable-class Agent capability down to $0.87 per million output tokens. They did announce a price increase ahead of time, and I don't know what it'll settle at, but I doubt it'll exceed $2.

Grok 4.6, meanwhile, turned near-Fable-5 comprehensive capability into a $2 input / $6 output price — and handed me a real-world development experience that's absurdly fast.

So a new kill line has formed. I've added this round's DeepSeek V4 Pro and Grok 4.6 to that famous LLM kill-line chart. Since DeepSeek V4 Pro's score hasn't been published yet, I estimated from the benchmarks and figured it lands around 56–58.

The LLM world is left with no survivors — total coverage.

First, let's look at DeepSeek V4 Pro's benchmark scores.

The numbers look dramatic, but out of all those scores, you only need to watch two.

The first is Terminal Bench 2.1. In simple terms, it makes the model enter a computer terminal by itself — installing the environment, typing commands, handling errors — and actually completing the task.

DeepSeek V4 Pro's preview version scored just 72.1; the official release jumped straight to 87.9, while Fable 5 sits at 88.0 — a gap of 0.1.

The second is DeepSWE.

This is the benchmark I focus on most. It's much closer to real software development: the model has to read and understand a code project on its own, then solve complex engineering problems inside it.

DeepSeek rocketed from 12.8 to 62.7 — a jump of nearly 50 points in one go.

Fable 5 scores 70, so DeepSeek still trails. But it's gone from being almost useless to dramatically stronger — it can actually get work done now.

The same trend shows up across codebase understanding, cybersecurity, automation, and tool calling: once it's given tools and actually allowed to operate in an environment, DeepSeek V4 Pro quickly closes in on Fable 5 — and even beats it on a few specific projects.

But its weaknesses are just as clear.

When it can't use tools and has to rely purely on the model's own knowledge and reasoning, it still trails Fable 5 by a dozen-plus points. On the hardest software engineering and complex full-stack tasks, it also lags by a few points.

So I'd rather call it a near-Fable-class model for Agent and coding work.

Its combat power inside tool environments is now first-tier, but its pure knowledge, hard software engineering, and long-horizon stability all felt a bit shaky in my own testing.

There's also one fatal point: it still doesn't do multimodal.

But that so-called fatal flaw doesn't matter much at this insane price and cache-hit rate.

Fable 5's API pricing is $10 per million input tokens and $50 per million output tokens.

DeepSeek V4 Pro currently charges $0.43 per million input tokens and $0.87 per million output tokens, with cache-hit input at just $0.0036.

That makes regular input roughly 23x cheaper.

Output roughly 57x cheaper.

And cache-hit input nearly 276x cheaper. Credit where it's due: DeepSeek's cache technology is straight-up black magic — not just cheap, but with an extremely high hit rate.

At this price, who the hell is going to complain, right? Of course, that assumes DeepSeek V4 Pro doesn't raise prices — and if the increase is modest, then paired with WorkBuddy in the future, I think that's the new god-tier combo.

I plugged DeepSeek V4 Pro straight into Claude Code and ran a bunch of real development and Agent tasks.

The capability is genuinely good — especially development and tool calling. I can clearly feel it's in a completely different state from the preview version, which I complained about quite a bit. Now at least it's usable; unless you're building something huge, you basically can't tell the difference.

Economically, this model is the new-generation kill line by itself.

But here's the catch — the impossible triangle still holds: DeepSeek V4 Pro's actual development speed is really quite average.

Especially when I'm using Grok 4.6 at the same time: over there, three tasks are already done; over here, one hasn't even finished. The contrast in speed is stark.

However, if you don't need real-time responsiveness and speed, or — like me — you're just running async automation, then I think this is a perfect model for today.

In the past, I'd habitually offload all those automation tasks to Codex.

But today, I've basically switched all of them to DeepSeek V4 Pro.

And then there's my new favorite, Grok 4.6.

Honestly, it's the only model out there today that I feel has nearly maxed out the impossible triangle: comprehensive performance rivaling Claude Fable 5, blazing speed, and a price that's higher than a price-slasher like DeepSeek, but several times cheaper than today's OpenAI and Anthropic flagship models.

Grok 4.6's API starts at $2 per million input tokens and $6 per million output tokens.

That's 5x cheaper than Fable 5 on input, and roughly 8.3x cheaper on output.

You might look at the price and performance and think it's about the same as Kimi K3 — but you can't do the math that way, because the Musk camp has subscriptions.

Today I signed up for $300 SuperGrok Heavy, and there's a promo right now: subscribing to SuperGrok Heavy gets you a Cursor Ultra membership worth $200... for free.

So for $300, you get the quota of two memberships.

Cursor alone gives you access to a pile of Claude models and so on — I won't even go into that. Just on the SuperGrok Heavy itself... do you know how absurd its quota is?

Do you know how absurd its quota is?

I subscribed just past 2 AM, then used their Grok Build and Coding straight through to 6 AM.

Four hours total — and the quota only dropped 2%. It went from 1% to 2%, and that tick only happened around 5:30. So for the first three and a half hours, I was developing nonstop, plus running N Agents on Heavy for deep research, and I'd used only 1% of my weekly quota. Even knowing this first week includes double usage, that's still absolutely insane.

At that burn rate, it's cheap to the point of absurdity — and I haven't even counted the Cursor quota.

Its overall performance is extremely strong, too. Compared with Kimi K3, apart from a slightly weaker aesthetic sense on frontend work, I don't really feel much difference.

And on real-world tasks — building financial models, PPTs, legal cases, and the like — Grok can even win a few. But its weak spot remains development and Agent work, where it does fall a bit short of Fable 5.

So that's why I call it a near-Fable-class model.

And its development speed is comfortable beyond words.

I developed continuously with Grok 4.6 for about four hours and barely ran into a task that took over 20 minutes.

Some tasks finished in as little as 3 minutes — with no drop in quality.

This is a speed I almost never experienced with Codex.

In the past, a task would run ten-odd minutes, half an hour, or even a full hour, and I'd be worn numb — spending most of the time just waiting. Today, developing with Grok 4.6, I felt, for the first time in ages, a genuine sense of interaction between me and my Agent.

Trust me: in many cases, speed is what truly turns a model's intelligence into productivity.

This impossible triangle — I think Musk pulled it off today. Grok 4.6 pulled it off.

That's why I said at the start that from today, I'm switching all my main daily development to Grok Build + Grok 4.6. I can sum it up in one word: bliss.

And Musk has thrown down another wild claim:

Grok 4.7 will surpass every model.

This time, I believe him.

So back to the very beginning: why did I say DeepSeek V4 Pro and Grok 4.6 would push all the LLMs onto the kill line?

Because the first thing a kill line severs is a model's right to keep being arrogant.

In the old LLM market, everyone could find room to survive.

The most capable could charge a premium and take their time.

The weaker ones could win users with low prices.

The fast enough ones could specialize in simple tasks.

Because the impossible triangle existed, every model had a shortcoming, and users could only accept all kinds of compromises.

But today, the market's minimum bar is different.

DeepSeek V4 Pro is telling everyone that near-Fable-class Agent capability can cost just $0.87 per million output tokens (and yes, I know it'll rise later, but I'd guess it settles around $2).

Grok 4.6 is telling everyone that a model with overall performance close to Fable 5 can simultaneously have blazing development speed and a subscription quota that drops only a few percent after hours of heavy use.

When both of these exist at once, the models caught in the middle are going to be in serious pain.

Grok 4.6, in particular, is basically the gatekeeper of Diamond rank in this game.

If you can't beat these two, what's the point of using you?

The entire kill line — pushed forward by DeepSeek V4 Pro and Grok 4.6 — has jumped a huge step ahead.

This is the best of times.

And also the worst of times.