mixed.dev

Happy to share I’ll be speaking at React Paris 2026 !

@julien_c I seriously wish @MistralAI had a come back with their upcoming model

Who will be the “deepseek of the west”?

@devinjameson @manicule And you’re doing a great job at it

@manicule is neat. Docs are really hard. So much content to create, prose and code snippets to keep in sync, diverse audiences to consider… Docs are also pretty fun to work on if you give yourself the time. There’s a particular joy in explaining things well.

@devinjameson @notch

@notch What

TypeScript is an insane quagmire for absolutely no benefit beyond documentation.

@imnotchalk @icanvardar Which is why I specified “in this chart”

@mixeddev @icanvardar well it scores well on other benchmarks as well tbh

grok 4.5 high is so efficient it genuinely doesn't feel real

@KatieMiller The benchmark was included in the training data

Grok 4.5 High: $1.51 per task Fable 5 Max: $17.32 per task Grok is 1/10th of the price for roughly same performance.

@jshobrook And token efficiency

Grok 4.5 is on the Pareto frontier of cost and intelligence

@JordanHall_dev @vmg If I understand correctly it’s only available in the USA currently

@vmg @mixeddev What about the UK as it's not inside the EU

This is a major breakthrough in both intelligence and cost efficiency. The team really outdid themselves! Sadly it’s a large enough model that it cannot ship in the EU right away, but it’s coming very soon.

@Rasmic it's raining models, unfortunately us europeans aren't invited to the party...

Everyone’s dropping models

@cursor_ai Available for EU customers when ?

We've partnered with SpaceXAI to train Grok 4.5. It’s our most powerful model yet and the first we've built for more than software engineering.

@billyuchenlin Will it be in Cursor too when it drops?

Install grok-build now. Try Grok 4.5 as soon as it drops. > "curl -fsSL x.ai/cli/install.sh | bash"

@dillon_mulroy @jonathan_wilke thx, will try doing more of this ! it really is very manual, but yeah creating reusable skills to steer it away from the things it does that I don't like seems like a big unlock

specs that focus on types, interfaces, input/output types, call stacks/data flow start keep track of the things you regularly have to correct it on during local code review and build a skill or code review agent around them to take a first pass - or feed them back into what goes into the tech spec i have not cracked automating this process and is very manual, but i am more productive and it’s gotten easier to steer and have consistency over time

Be honest: are you still reviewing every line of code your agents generated?

@dillon_mulroy @jonathan_wilke it feels like I spend more time trying to steer the model and clean up slop than it takes to actually write the code myself... any tips on what works for you ?

@jonathan_wilke yes for anything that ships to production with users

OH: "Instead of AI taking my job, I now work two jobs to pay for Fable" 🙃

@jediahkatz But composer 2.5 scores higher and is much cheaper ?

For those who are not rich enough for Fable 5, we also just launched Kimi K2.7 Code in Cursor! Try it out and let me know what you think.

@robj3d3 Anthropic would need to fix the capacity issue…

GPT Sol Ultra via the API will cost nearly 50% less than Fable 5. For effectively the same result (according to benchmarks). Mark my word: once Sol Ultra is released to the public, Fable will be re-included into subscriptions in order to compete.

Beautiful way to put it.

@rauchg This is actually very cool

Next.js "Ways to fix this" with "Copy prompt" buttons is a work of (agentic) art

@fortysevenfx If it’s real, I’m pretty sure Claude wrote it

By their calculations, I have 80 years of experience in TypeScript