Three frontier brains. One month. One routing rule.
Three frontier AI models are live right now, and they all just landed within about a day of each other.
One of them is the smartest money can buy.
One of them costs so little that you can leave it running all day and forget the meter exists.
One of them went from nowhere to the frontier in a single month.
I ran all three on my own bench and put the builds side by side, so you can see exactly where each one wins and where each one falls apart.
And there's one number in this comparison almost nobody is talking about. It isn't the intelligence score, and it decides everything.
What you're watching: one operator, three frontier brains, and every job going to the machine that should be doing it.
0frontier models
0side-by-side builds
0× cheaper output
0× cheaper cache
same brief · three models · played at the same time
Outrun, built by all three.
Grok's has light-gates and traffic to dodge, DeepSeek's is the widest track, Fable 5 goes hardest on the colour. Same one-line brief, one shot each.
§1I ────── the problem
The Wrong-Brain Problem.
Imagine hiring one person to do every job in your business.
They plan your strategy. They also type your data entry.
You pay senior rates for both.
That's what almost everyone does with AI right now.
One model, one subscription, every single task.
So the hard thinking and the mindless grinding cost you exactly the same.
And now that three frontier models are live at once, picking the wrong one for a job can cost you fifty times more than it should.
The Three-Brain Engine ends that for good.
Left: what most people do today. Right: the same jobs, sorted by which brain they actually need.
same brief · three models · played at the same time
Doom, built by all three.
Grok's is the moodiest and by far the darkest, DeepSeek's corridor is the most readable, Fable 5 washes the whole screen red.
Thinking it?"Three models sounds like three times the work."
It's one setting, not three workflows.
You pick which brain answers, the same way you pick which staff member takes a job.
§2II ────── the week
Three frontier models, about one day apart.
DeepSeek V4 Pro went full release on 12 August with no launch video and no countdown — they updated the pricing page and let the model speak. Grok 4.6 landed hours earlier, and Claude Fable 5 was already sitting at the top of nearly every leaderboard.
same brief · three models · played at the same time
Open-city driving, built by all three.
Three takes on 'steal cars, outrun cops'. Watch the speed readouts and the minimaps — all three wired a real driving model from one sentence.
Tweet 1 · the launch
xAI shipped frontier intelligence at the old price
This is the announcement itself — 16.9 million views. The key line is "at the same price" as Grok 4.5, and the thread underneath pins it down: two dollars per million in, six dollars out, which xAI calls half of what other frontier models charge.
One is the smartest. One is the cheapest. One is the fastest riser.
§3III ────── raw intelligence
Fable 5 is still the smartest model you can buy.
That part isn't really in dispute. On the hard reasoning and long-build tests, Anthropic still holds the top of the table — and this week didn't change it.
DeepSWE numbers are each lab's own published figures. When the work is genuinely hard, Fable wins.
same brief · three models · played at the same time
The twilight world, built by all three.
The test that really separates them: the hero models. A multi-part character with a cape and a drawn sword, versus a coloured shape with legs.
§4IV ────── the fastest riser
Now look how close Grok got.
One month ago Grok 4.5 was sitting in thirteenth place on the blind-vote coding board. Grok 4.6 walked into the top seven and landed within nine points of Fable 5 — a jump most labs take three to six months to make.
same brief · three models · played at the same time
The voxel world, built by all three.
Block worlds from one sentence. Terrain generation is where one-shot builds usually collapse into a black screen, so a lit walkable world is the bar.
Tweet 2 · the leaderboard
The blind-vote board put Grok 4.6 level with the best
This is Arena — real people voting on builds without knowing which model made them, which makes it the hardest kind of test to game. Grok 4.6 came in at 1618 against Fable 5's 1627. Watch the last line: they warn the picture will sharpen as more votes land.
§5V ────── deepseek's claim
Then DeepSeek published enormous numbers.
On their own testing, DeepSeek V4 Pro nearly caught Fable 5 on agent work — and their DeepSWE score went from 12.8 in the April preview to 62.7 in this release. That's a fifty-point leap on the same test.
Tweet 3 · the claimed jumps
Every number here comes from DeepSeek
Terminal-Bench 2.1 from 72.1 to 87.9, CyberGym from 52.7 to 83.3, DeepSWE from 12.8 to 62.7. Note what Chris says at the end — every lab reports its own separate benchmarks, so he'd rather wait for the independent index. Hold that thought.
Tweet 4 · the builders react
The coding-agent crowd called it instantly
Cline builds one of the big AI coding agents, so they judge a model by what it costs to run all day. Their read: Fable 5 performance at roughly fifty-seven times cheaper, and the best price-to-performance model on the market. They had it live in their own product the same day.
§6VI ────── the independent check
Then the independent testers landed.
Vals AI run their own evaluations, and their DeepSeek V4 Pro results published this morning tell a split story: the cheapness is completely real, and the agent-benchmark claim does not survive contact.
This is why you never buy a model on its own launch-day slide. The price claim survived. The agent score didn't.
same brief · three models · played at the same time
The arcade shooter, built by all three.
Enemy waves, radar and hull damage in all three. The difference is the juice — trails, tracers, and how much the screen reacts when you hit something.
Tweet 5 · the reality check
Independent testing, published this morning
Read the fourth post down — that's the one that matters. Terminal Bench 2.1 at 54.68 percent, thirty-third of fifty-two, against the 87.9 percent DeepSeek published. But look at the rest of the thread too: their Proof Bench score went from 10 to 49, and it runs at seven cents a task against a rival's one dollar sixty-seven.
Thinking it?"So how do I know which numbers to trust?"
You don't trust any of them — you watch the same model do your kind of work.
That's exactly what I built the next section for.
Launch slides are marketing. Builds are evidence.
§7VII ────── the arena
So I gave all three the same jobs.
Same prompt, same rules, one shot each, no hand-fixing — then a vision judge scored every build against the identical rubric. Click any tile and the real thing loads and runs in your browser.
How to read this. Every tile is a real one-shot build from this week, playable right here. Nothing was scored from a still frame — each build was driven for about twenty seconds like a player would drive it, watching whether the meters actually moved, and judged on what it did. Anything scored 5 or below is a broken build, and I've left every wreck in, because a comparison that hides the failures is a brochure.
8.26Fable 5 · avg
6.80Grok 4.6 · avg
6.35DeepSeek · avg
0Fable builds broken
01Skyrim-lite
first-person open-world fantasy explorer.
Claude Fable 59.0
AAA rebuild is showcase-grade: a dramatic Nordic dusk valley with layered snow-capped peaks, pine forests and ruins, a…
DeepSeek V4 Pro6.8
Renders a clean low-poly first-person fantasy world with terrain, trees, rocks, shards, HUD meters and full control hi…
Grok 4.6failed · 2.0
Broken build — shown deliberately.
02Twilight Vale
3D open-world RPG with combat, terrain, weather.
Claude Fable 59.0
AAA rebuild is showcase-grade: an authored hooded hero with a glowing trailed weapon on a moody twilight field of glow…
DeepSeek V4 Pro7.4
Renders a polished twilight-toned 3D world with stylized low-poly trees, rocks, a knight avatar, shadows, minimap and …
Grok 4.68.3
Renders a cohesive twilight RPG world with terrain, trees, rocks, shard pickups, enemies, a rigged hero with sword, mi…
03Voxel Craft
Minecraft-style sandbox, place + break blocks, day/night cycle.
Claude Fable 59.0
AAA rebuild is an authentic textured Minecraft: procedurally-textured grass/dirt/stone/wood blocks (not flat colors), …
DeepSeek V4 Pro7.8
Strong voxel world renders with trees, blocky terrain, a held pickaxe tool, polished HUD/hotbar and day/night indicato…
Grok 4.67.8
Renders a polished voxel world with detailed character model, clean HUD, hotbar, and trees; altitude changed (012→007)…
04Crypt
torch-lit dungeon crawler.
Claude Fable 58.8
AAA rebuild via the threejs-game-director skill transforms it into a real torch-lit crawler: warm-lit cracked-stone wa…
DeepSeek V4 Profailed · 2.2
Broken build — shown deliberately.
Grok 4.66.3
Renders a full 3D dungeon with a polished HUD, minimap, character model and torch/vitals meters, but the lighting read…
05Outrun
synthwave horizon driving game with pseudo-3D road.
Claude Fable 58.7
Gorgeous synthwave palette with banded retro sun, parallax mountains, glowing palms and neon-edged pseudo-3D road with…
DeepSeek V4 Pro8.4
Gorgeous on-brief synthwave scene — gradient sky, striped sun, mountains, neon grid, and a chunky pixel car with worki…
Grok 4.68.6
Gorgeous synthwave scene — glowing pink grid, neon-edged pseudo-3D road, arches, mountains, city skyline and a detaile…
06Neon Blaster
juicy arcade space shooter, waves, bosses, power-ups, screen-shake, synth music.
Claude Fable 58.4
Screenshot shows a polished, on-brief neon shooter with glowing enemies, a firing player ship, a colorful starfield, a…
DeepSeek V4 Pro5.5
Renders a polished 3D ship with clean HUD and waves progress in text (wave 2→3), but pixel-change is near-zero (~0.02)…
Grok 4.66.8
Renders a polished 3D scene with detailed ship, HUD meters, minimap, crosshair and asteroids, and hull damage clearly …
07Top-down RPG
top-down RPG with sprites, combat, inventory.
Claude Fable 58.4
Iterated rebuild renders a polished top-down Emberfall RPG: a checkerboard overworld with trees, a dirt path, a lake, …
DeepSeek V4 Pro7.8
Renders a polished low-poly 3D top-down world with clean HUD, working combat/XP/potion loop (KILLS, SCORE, POTIONS all…
Grok 4.67.8
Strong 3D top-down world with polished HUD, minimap, low-poly trees/enemies/pickups, and working combat (vitality drop…
08Arcade
classic arcade-style game (pick: tetris, breakout, snake).
Strong render: coherent city grid, textured buildings, traffic, working minimap, and a polished HUD; playtest confirms…
19Dogfight
air-combat shooter.
Claude Fable 57.4
Renders cleanly with a solid 3D scene—chase-cam plane, clouds, mountains, HUD, health bar and crosshair all present an…
DeepSeek V4 Pro6.5
Renders a polished 3D flight scene with a detailed aircraft, HUD, minimap, and objective rings, and SPD/ALT meters pro…
Grok 4.65.5
Polished HUD, radar with blips, and a nicely modeled 3D fighter render well, but score/kills stayed frozen at 000000 t…
The pattern in the wreckage. Grok 4.6 broke three of these, and it broke them the same way every time: a beautiful, complete HUD drawn over a completely empty world. Its Skyrim build has health bars, a stamina meter and a gold counter — and no landscape at all, because one duplicate variable kills the script. Nordic Crypt died the same way on k is not defined. Neon City threw no error at all and still rendered no city. DeepSeek broke more often — five of nineteen — but differently: usually something half-finished rather than something hollow. Fable 5 broke nothing.
Skyrim — the job Grok blanked hardest, scoring 2.0 — is one of the three Fable scored a 9.0 on.
same brief · three models · played at the same time
The crypt crawler, built by all three.
Torch-lit dungeons, three ways. Look at how each one handles light falloff — it's the whole mood of the level.
Thinking it?"I don't build games, so why should I care?"
Games are just the hardest possible version of "follow a long brief without being reminded."
A model that forgets the world while drawing the dashboard will do the same thing to your report.
Now the number I promised you at the start.
§8VIII ────── the number
The number is 276.
On the headline sticker price, DeepSeek is roughly 23 times cheaper than Fable 5 to read and 57 times cheaper to write. Grok sits in the middle at about half of what other frontier models charge. But none of those are the number that matters.
The headline says 57 times cheaper. For agents that run all day, the real gap is far bigger.
same brief · three models · played at the same time
On foot, built by all three.
The third-person street brief. Ammo counters, pedestrians and a working minimap in all three — the gap is the lighting and how alive the street feels.
Why re-reading is where the money actually goes
An agent doesn't read your instructions once. It reads the whole job folder again before every single action.
Open the folder, read everything, do one thing. Open it again, read it again, do the next thing — hundreds of times.
same brief · three models · played at the same time
The parachute jump, built by all three.
Freefall and canopy. The phase change — freefall into an open chute — is the part that catches models out.
Tweet 6 · the penny drops
Someone else spotted it the same day
"A Fable level model at one fiftieth of the API price, but with a much higher cache hit rate." That's the whole thing in one sentence. And read the reply underneath — someone says the cache hit rate is the part nobody talks about, and that it's what actually decides the bill. That's the number I just showed you.
Thinking it?"This is starting to sound like accounting homework."
You never do this maths — the setting does it.
All you need to know is that leaving an agent running all day stopped being a decision you have to think about.
§9IX ────── in plain terms
Here's what that buys you.
If you're a freelancer, you can leave a research agent running on client work all day and stop watching the meter. If you sell online, your product descriptions and customer replies can run nonstop on frontier-adjacent intelligence.
Beforeyou ration it
You open one chat and babysit it
You think twice before a long job
Repetitive work still costs top rate
Agents feel like a luxury you switch on
You close the tab when you leave the desk
Afteryou forget it
Agents run in the background all day
Long jobs stop being a decision
Volume work routes to the cheap brain
The hard calls still get the best brain
Work happens while you sleep
§10X ────── already happening
The agents are already on it.
The number one app sending traffic to DeepSeek V4 Pro on OpenRouter right now is Hermes Agent — the open-source agent from Nous Research — with over two billion tokens. The agents that quietly do real work in the background moved within days.
What you're watching: my own Agent OS with all three of these models wired in as engines — the same job handed to a different brain by changing one setting.
Thinking it?"Doesn't running Agent OS burn a fortune in tokens?"
No — that's the biggest myth about it. The everyday ninety percent runs on a free local model on your own machine, nothing leaving it, and free APIs slot in for more.
For the frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and Agent OS plugs straight into it. It's a layer on top of what you already own, not a new meter. There are full token-efficiency tutorials inside the Boardroom too.
§12XI ────── the third factor
A model isn't a score. It's a thing you have to use.
This is where the three of them split completely — and it might decide more real outcomes than intelligence and price combined.
DeepSeek expects you to bring your own harness. If you already run agents, your setup is the harness — DeepSeek is just a very cheap brain you drop inside it.
same brief · three models · played at the same time
The frozen open world, built by all three.
The deep-build prompt: walk into the snow, draw your sword. Watch how much world each one actually puts in front of you.
§13XII ────── grok's edge
Grok has one thing nobody can copy.
It's wired directly into X. No plugins, no workarounds — if your business depends on knowing what people are saying right now, Grok pulls that live. Fable can't. DeepSeek can't.
What you're watching: Grok 4.6's Outrun build from the arena, running live — 8.6, one of the five jobs it won outright.
And it admits when something isn't there
There's a trap test where a model gets ten problems to fix, and five of them are fake.
Most models invent answers for the fake ones. Grok 4.6 said "this doesn't exist" and moved on.
Making things up is the number one reason people stop trusting AI with real work. A model that admits a gap is one you can leave alone with a job.
same brief · three models · played at the same time
The 3D racer, built by all three.
A track, a car and obstacles. The tell here is whether there's a real racing surface under the vehicle or just a coloured plane.
Databricks found the same pattern on their office-work benchmark — reading reports, pulling numbers out of files, making sense of messy data. Grok set the top score there. That's office work, not code.
§14XIII ────── fable's edge
Fable's edge is the hardest work.
Long, ambiguous, high-stakes jobs — planning a whole campaign, untangling a messy client problem, designing a system from nothing. That seven-point lead isn't a rounding error; on genuinely difficult multi-step work it shows up as fewer mistakes and fewer restarts.
What you're watching: Fable 5's Skyrim build from the arena — the 9.0, on the exact job Grok scored 2.0 on. This is what "follows a long brief without being reminded" looks like.
§15XIV ────── deepseek's edge
DeepSeek's edge is volume.
It has 1.6 trillion total parameters but only 49 billion switched on at any moment — picture a company with 1.6 trillion staff where only the relevant 49 billion turn up for each task. That's the engineering reason it's cheap, not a company selling below cost.
same brief · three models · played at the same time
The flight sim, built by all three.
Take off, climb, bank a turn. Flight models are the single hardest thing to get right in one shot — watch the altitude and speed readouts.
What you're watching: DeepSeek V4 Pro's Neon City build from the arena, running live — it scored 8.6, its best of the run, and beat both other models on that job.
§16XV ────── the catches
DeepSeek has three catches. Hear them now.
It can't see.
No vision at all. If your workflow depends on screenshots or reading documents as pictures, it's out — and DeepSeek have said vision doesn't advance the research they care about, so don't wait for it.
A price rise is coming.
They've posted a notice about a significant increase across the whole API. No date, no amount. Today's pricing is real but not guaranteed — though they'd have to raise it many times over before the value maths flips.
Their terms let them train on what you send.
That applies to the official API. Other providers host the model without that condition and more are coming online — but on day one, if your work is sensitive client data, factor it in.
Thinking it?"Then I'll just skip DeepSeek."
You don't have to hand it anything sensitive to get the win.
Point it at the high-volume work that isn't confidential — research, drafting, sorting, monitoring — and keep the rest where it already is.
Nobody wins everything. That's the most useful finding here.
§17XVI ────── the framework
The Three-Brain Engine.
The pattern the smartest operators landed on doesn't pick one. It uses all three like a team, and each brain has exactly one job.
i.
Fable is the planner. It runs once.
It designs the workflow, makes the hard calls, and reviews the final output. You pay top price — but only for the moments that deserve top price.
ii.
DeepSeek is the workhorse. It runs ten thousand times.
Research, drafting, sorting, monitoring, follow-ups. Every repetitive high-volume task routes to the brain that costs a fraction as much, with cache pricing that makes long runs almost free.
iii.
Grok is the specialist. It watches what's happening now.
Live market awareness through X, document and data work where it set the top score, plus images and video. It's also the value pick when you want near-Fable quality at a fraction of the price.
The expensive model runs once. The cheap model runs ten thousand times. That's where the leverage lives.
same brief · three models · played at the same time
The dogfight, built by all three.
Air combat with a radar and enemy fighters. Spawning somewhere survivable is half the battle in a one-shot build.
§18XVII ────── what it looks like
For a business owner, it looks like this.
Fable designs your lead follow-up workflow once. DeepSeek runs it on every single lead, every day, forever. Grok watches your market in real time and feeds what it finds back in.
Fable writes the playbook.
One session. It designs how a lead gets qualified, what gets said, when to follow up, and what a good reply looks like. You pay top rate once.
DeepSeek works the list.
Every lead that arrives gets researched, drafted for and followed up by the cheap brain, running in the background all day.
Grok watches the market.
It pulls what people are complaining about and what's landing right now, and hands that back so the messaging stays current instead of going stale.
Fable reviews the output.
At the end, the expensive brain checks the work and catches what the cheap one missed. Plan with the best, execute with the cheapest, review with the best.
Developers all over Hacker News described exactly this split last week. It works because the expensive model runs once and the cheap model runs ten thousand times.
§19XVIII ────── three beliefs to drop
Three things you might be thinking right now.
Wrong: "Three models and routing tasks — this is for technical people."
Right: Look at what we actually covered. The Databricks test Grok topped was reading reports and pulling numbers out of files. The agents running DeepSeek take instructions in plain English sentences. If you can write an email, you can run this.
Wrong: "I already pay for one of these, why complicate it."
Right: If you run a handful of tasks a day, one model is genuinely fine. The moment you run agents in the background, you're paying frontier prices for work that doesn't need frontier intelligence. It's the same call every business makes about people — you don't put your most senior person on data entry.
Wrong: "It's moving too fast. I'll wait until it settles."
Right: In one hour last week DeepSeek V4 Pro went live, Grok 4.6 dropped, and Qwen released the weights for their 3.8 Max model. It isn't going to settle. But the chaos only hurts people without a system — the ones who had their setup running plugged Grok in the day it dropped and DeepSeek the day after. The models change monthly. Your setup doesn't.
Tweet 7 · one hour
All three of those landed inside a single hour
Not one week. One hour. DeepSeek V4 Pro, Grok 4.6, and Qwen releasing the weights for their 3.8 Max model. This is the tweet to look at when you catch yourself thinking you'll wait until things calm down — there is no calm point coming.
Don't take my word for it
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
The best model didn't change. The best strategy did.
Old way1 brain · top rate
One subscription does every job
Data sorting costs the same as strategy
You ration long jobs to control spend
A new release means starting over
No live market data at all
You babysit one chat window
Result: brilliant, running rarely
New way3 brains · routed
Each job goes to the brain it needs
Volume work runs 57× cheaper
Long agent runs are nearly free on cache
A new release is a free upgrade, swapped in minutes
Grok feeds live market signal back in
Agents run in the background all day
Result: smart-enough, running always
§21XX ────── what's next
And it isn't slowing down.
Grok 4.7 is already trained. DeepSeek jumped their own DeepSWE score from 12.8 in the April preview to 62.7 in this release. Anthropic will answer, because Anthropic always answers.
Tweet 8 · the next one
Grok 4.7 is three to four weeks out
Elon replying to a post saying Grok is back on the frontier menu. The detail that matters: initial training is already finished, and they're now feeding in a mass of SpaceX company data. This is the pace you're planning around — not a yearly cycle, a monthly one.
This three-way fight is the new normal. And every round of it drives the cost of intelligence down, which means every round makes your agents cheaper to run — whichever model you picked.
§22XXI ────── the landing
So let's land this.
Claude Fable 5 is still the smartest.It earned that. On my own bench it averaged 8.36 across these sixteen jobs — and it won the two hardest outright with 9.0s.
Grok 4.6 proved xAI can reach the frontier in a month.Thirteenth to seventh on the blind-vote board, nine points off Fable, with live X data nobody else has.
DeepSeek V4 Pro made frontier-adjacent intelligence nearly free.57× cheaper output, 276× cheaper cache reads — for the volume work most businesses actually need.
You stopped picking a model.You started routing jobs. Hard work to Fable, volume to DeepSeek, live signal to Grok.
You stopped paying senior rates for data entry.The expensive brain runs once. The cheap brain runs ten thousand times.
You stopped fearing releases.Grok 4.7 lands in weeks. With a system, that's a free upgrade instead of homework.
Smart-enough running all day beats brilliant running rarely.
And now you can afford both.
The models stopped being the bottleneck.
Knowing which brain to point at which job — that's the bottleneck now.
