The 57X Brain Swap.

DeepSeek V4 Pro went toe to toe with the most expensive model on the market.

The full version landed on August 12th, 2026.

On DeepSeek's own agent tests it sits a tenth of a point behind Claude Fable 5 — and costs 57 times less to run.

Both prices are live right now — I pulled them off the API before writing this.

This is what the cheap model built while I was writing that paragraph. No retries, no fixes from me.

Most people got this story half right.

V4 Pro has been sitting in preview since April — usable, but unfinished.

On August 12th the finished version landed, and it is a different model.

No blog post. No launch video. No countdown. They updated the pricing page and let the model talk.

The April preview could not do this. Same prompt, same one shot.

Three agent tests. Two of them are basically a tie.

These come from DeepSeek's own testing, so hold them loosely — I'll show you what the independent testers found further down.

Mint is DeepSeek V4 Pro, violet is Claude Fable 5. Humanity's Last Exam: Fable takes it, 63.0 to 60.0.

Benchmarks are one thing. This is the same model flying a plane it wrote from a sentence.

Six more, playing together: an RPG with loot, a wave shooter, a brick-breaker, a dogfight, a lap racer and a snow RPG. Every panel is a separate one-prompt build.

23 times less to read. 57 times less to write.

Same neighbourhood of performance on agent work, a tiny fraction of the cost.

Rose is Claude Fable 5, mint is DeepSeek V4 Pro. I checked all six numbers against the live pricing API today.

What you're watching: the two ways to point an agent at V4 Pro — and the exact error you hit on OpenRouter if your data-policy setting is still locked down.

An agent doesn't read your instructions once. It reads them hundreds of times.

Picture a worker who opens the whole job folder, reads everything, then does one thing — and repeats that before every single action.

Every trip round that loop re-reads the same context. That re-read is called a cache hit — and it's where the money actually goes.

What you're watching: three real steps of the same agent. Step one pays full price. From step two, 7,552 of 7,616 tokens are cached and the step costs $0.00009.

Over 90% of what an agent pays for is nearly free now.

Cache reads on V4 Pro cost 276 times less than on Fable 5, and this model is running at about a 92% cache hit rate.

The headline gap is 57×. For a long-running agent, the real gap is bigger than the headline suggests.

Cheap enough to leave running. Good enough to build this.

And three more — a particle galaxy you can swirl, a synthwave sunset, and a snake game that plays itself. That is fifteen builds filmed live, side by side.

You can leave an agent running all day and stop thinking about it.

Checking every step, retrying when it fails, working while you sleep.

Like having an agent sort and draft replies to your whole inbox overnight, without watching the meter.

What you're watching: my own Agent OS. I flip the model from V4 Flash to V4 Pro with one tap, type one line, and it builds — sped up, because V4 Pro reasons for minutes before it writes.

The number one app on this model is an agent that never stops.

Hermes Agent — the open-source agent from Nous Research — is sending more traffic to V4 Pro than anything else, over two billion tokens.

Coding tools like Pi and the Claude Code harnesses sit right behind it.

The agents I keep talking about on this channel are already running on this model. Today.

Same engine those agents are running on. This took one sentence.

What you're watching: the Agent Kanban board inside my Agent OS — a Planner, a Builder and a Reviewer working one board, with every Done card previewing live. This is what you point a cheap frontier brain at.

So I handed it twenty games and one prompt each.

Same GoldieBench tasks I run every new model through — one shot, no retries, no hand-patching.

What you're watching: one of my real build runs, replayed at reading pace — ten games in parallel, 652,220 bytes of working code. I metered one of them — 22,619 tokens, $0.0197. The same tokens on Fable 5 output pricing would be $1.13.

These are real, and you can play them right now.

Every clip below is the actual build being driven — no mockups, no edits.

Seventeen of the twenty came back working on the first try.

"The preview became official" undersells it badly.

Compare April's preview to this build and the jumps are enormous — DeepSWE went from 12.8 to 62.7.

Rose is the April preview, mint is the August build. Nearly 50 points on the same software engineering test.

They grow separate specialists, then merge them into one brain.

A researcher, a builder, a planner — trained apart, then combined through a process called distillation.

The April preview hadn't finished that process. This build has.

The independent testers have now weighed in.

Vals AI ran it, and their results cut both ways — huge gains on some tests, and a Terminal Bench score nowhere near DeepSeek's own.

This is why the lab's own numbers get held loosely. The cheap agent story survives it — the "ties Fable on everything" story doesn't.

1.6 trillion parts, but only 49 billion show up for each job.

Picture a company with 1.6 trillion employees where only the relevant ones walk into the room — that's how it stays cheap to run.

Hear these from me now, rather than find out later.

They'd have to raise prices many times over before this stops being the value play. They know that — it's probably why they can afford to.

Plan with the expensive brain. Execute with the cheap one.

Claude designs your lead follow-up workflow once. DeepSeek runs it on every lead, every day, forever.

The cost excuse is gone. What's left is the knowledge gap.

Smart-enough just became nearly free.

Chinese labs are shipping frontier-class models nearly every month now — Kimi K3, MiniMax, Qwen, GLM — and DeepSeek just planted its flag at the front of that pack.

Whatever you think about where these models come from, the competition is driving the cost of intelligence toward zero, and every business benefits from that whether you use DeepSeek or not.

One tap swaps the brain from V4 Flash to V4 Pro, and the same workspace keeps building. That's the whole point of running an operating system instead of a pile of tabs.