LASCHUK

Vibe Coding Until R$ 1.000.000

[LASCHUK @laschuk]

obrigado pela concorrência, @OpenAI

[LASCHUK @laschuk]

grok 4.6 > fable 5 in frontend design.

[LASCHUK @laschuk]

12 heavy benchmark runs on GPT burned the same 5% weekly limit that Claude burned in 4.

3x the work, same budget. That math happens on your machine every day and you can't see it.

So I built Redline, every provider's rate limit in one menu bar. Shipped today.

[LASCHUK @laschuk]

Everyone is about to post that Opus 5 matched Fable 5 at half the price. 51 vs 50 on Artificial Analysis.

I ran it. That's true... and it's incomplete.

11 real tasks. Opus went 9 for 9, faster and at half the cost. Then I gave both the same landing page. Fable won outright, at 2.5x the price and twice the time.

Both, side by side 👇

Benchmarks have a column for "finished". They don't have one for "good".

[LASCHUK @laschuk]

@maarcoofdezz Boris explain in 30 minutes how to burns your Fable 5 limit in 15 min.

[LASCHUK @laschuk]

Anthropic had the best model in the world. Had. The facts, in order:

1. Jul 9: OpenAI ships GPT-5.6: 3 sizes, transparent pricing, no fine print.
2. Jul 16: China's Moonshot releases Kimi K3: open-weight, 3x cheaper, full weights free for anyone by Jul 27.
3. Jul 16: K3 debuts at #1 on @arena's coding leaderboard: 1,679 vs Fable 5's 1,631. First place in 6 of 7 categories.
4. Jul 16: My own benchmark reached the same verdict: 9 real coding tasks, K3 won 7×2 while spending less than half ($3.43 vs $7.43). Fable only won the 3D
game (the exact same lone category it held on Arena).
5. Jul 17: Anthropic's response: Fable returns to paid plans... at half the usage cap, plus a $100 credit that evaporates in a weekend.
6. That's the 4th deadline change in 1 month (Jul 7 → 12 → 19 → 20). "Demand has been challenging to predict" is the polite version of "we're out of chips"... the CEO himself admits the company is compute-constrained.

Bottom line: they lost the coding crown to a free model, and their answer was to charge more and deliver less.

You're no longer selling intelligence. You're selling a waiting line.

[LASCHUK @laschuk]

Fable 5 vs Kimi K3. Same prompt.

One cost $2.11. The other $1.19.

Which one won? Follow me + reply "FPS" and I'll DM you the exact prompt I used.

[LASCHUK @laschuk]

Create a single HTML file that recreates the Apple homepage (apple dot com). A fixed Navbar at the top displays the Apple menu (logo, products, search, shopping bag). Featured product hero: large centered image, large title, subtitle, and two links ("Learn More ›" and "Buy Now ›"). Below, a grid of product brochures alternating between light and dark backgrounds, each with a centered title, subtitle, and CTAs. Typography, weights, and spacing faithful to Apple (San Francisco / -apple-system), responsive desktop and mobile, subtle hover and scroll animations. You can use gradients or SVG as placeholders for product images. Do not reference any other files.

[LASCHUK @laschuk]

I put Kimi K3 up against Fable 5 in my benchmark.

The test: clone Apple's homepage. Single prompt, one shot, zero help.

Kimi K3: $0.44 (at left)
Fable 5: $0.94 (at right)

Same test. Same prompt. Half the price.

Fable runs $10/$50 per MTok, the most expensive tokens on the market. K3 costs $3/$15 and shipped the site anyway.

So the question is: what exactly am I paying for in that premium?

[LASCHUK @laschuk]

Full game, zero lines of code written by me.

Built by Kimi K3: Chinese, launched today, beating Fable 5 at 1/3 of the price.

[LASCHUK @laschuk]

Claude Fable 5 acabou ter seus limites resetados pela Anthropic.

Tudo isso diante do lançamento do GPT 5.6.

Sentiram.

[LASCHUK @laschuk]

Sol Ultra liberado, yeeah!

[LASCHUK @laschuk]

@claudeai Claude tá com medo do Sol Ultra!

[LASCHUK @laschuk]

contratando a 3a conta de Claude 20x pra usar Fable 5. que absurdo esse modelo.

[LASCHUK @laschuk]

Overclock v1.0 hoje às 18h.

[LASCHUK @laschuk]

O lançamento do Fable 5 é uma cortina de fumaça.

Anthropic acabou de disponibilizar a versão Next do Claude Code CLI 2.1.198 que agora por padrão dispara subagente em background e herda o modelo do pai.

Fiz um teste rápido: contar os .ts do repo:
- Fable 5 herdado‚ US$ 2,29
- Haiku forçado‚ US$ 0,15

15x pra uma ação simples.

Roteamento de modelo devia ser trabalho da ferramenta.

[LASCHUK @laschuk]

acabei de forçar a atualização do Claude Code CLI pra 2.1.198.

Fable 5 muito perto de ser desbloqueado.

[LASCHUK @laschuk]

confesso: eu tava melhor com o fable 5 banido. quando lançaram, o tps caiu pra 2 tokens/s e travou tudo. sem ele o opus 4.8 voltou a voar com 58 tokens/s. não senti falta.

[LASCHUK @laschuk]

o preço por token não é o custo.

claude sonnet 5 tem o token mais barato dos 3. mesmo desafio, ele saiu MAIS caro que o opus.

26 idas e voltas. releu 1,3M de cache pra conferir o próprio trabalho.

você paga pelo comportamento, não pela tabela de preço.

[LASCHUK @laschuk]

@TarcisoRenova vaamo!

[LASCHUK @laschuk]

acabamos de bater nosso record de ARR. isso graças ao banimento do Fable 5, que descongestionou os servidores do Claude e restaurou os 60 tps. acelerar o vibe coding não é sobre o melhor modelo e sim fazer o Overclock da sua stack combinando LLM, effort, CLI, Skill e Agent.

#vibecode #overclock