Hansen Liang

Building AI agents in public. Sharing what I learn: agents, moats & the AI-era SaaS shift. Multi-model stack, shipping daily.

[Hansen Liang @HansenIsSo]

Agree? #anthropic #claude #opus55

[Hansen Liang @HansenIsSo]

@AnthropicAI cooked with #Opus 5.5 so hard that I'm switching back from gpt-6-astra.

In a number of complex coding tasks in my testing on a production code base, Opus 5.5 smoked gpt-6-astra in both speed and quality. The only place where astra leads for me is computer use.

Anyone else feel this way?

[Hansen Liang @HansenIsSo]

I’m looking forward to trying @typesafeai myself and in production. We’ve known for a while now that LLMs alone may not be the answer to AGI… stay tuned on whether this is hype vs real

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Hansen Liang @HansenIsSo]

Definitely felt the reduced usage immediately

[Matthew Miller @matthewmillerai]

Anthropic just rug pulled every Claude Max user.

The +50% weekly limits promo is over. Replaced with +25% permanently. That is a usage CUT disguised as a perk.

30 minutes of Fable 5.1 this morning. 90% of my session limit gone. 13% of my weekly gone.

It is noticeably worse. Immediately.

This is a $200 plan. 30 minutes should not cost you 90% of anything.

Anthropic keeps finding new ways to give you less while charging the same.

[Hansen Liang @HansenIsSo]

Today, I officially moved my agent product off of @AnthropicAI . Here's why.

1/ The cost is adding up. A couple of months into a growing product, the bill from Anthropic in inference (I've been using Sonnet for prod) has become one of the MOST expensive parts of the product.

2/ Frontier intelligence isn't needed for 90% of the job. What you get out of @deepseek_ai, @Alibaba_Qwen et al is more than enough.

3/ You can quite literally get MORE for LESS if you were aiming for "mainstream intelligence" (aka Sonnet-level). How much less? See my report below.

I've done extensive testing with my own benchmark. I'm sharing it here if you want to dig in: https://t.co/dSfnsK8b1t

[Hansen Liang @HansenIsSo]

Congrats @Arjunjain1017 and team on the raise! Hardware remains an incredibly exciting frontier for AI. Can’t wait to see what you cook up.

[Arjun Jain @Arjunjain1017]

We've raised $4.9M to build @tndminc - The AI system for hardware engineering.

Engineering ambition has outgrown the systems built to carry it. Tandem connects requirements, design intent, CAD changes, reviews, and validation.

The next great machine is waiting on its system.

[Hansen Liang @HansenIsSo]

The writing style is a bit exaggerated, but the point is interesting. I think he’s right.

[Pascio @IAmPascio]

https://t.co/5b2m7xiDPx

[MIKE @mikenevermiss]

this is f**king insane.

a solo dev just open sourced a 100% FREE ElevenLabs replacement that runs entirely on your own machine.
the GitHub repo is at 19.4K stars.

it lets you:

→ clone a voice from one clean reference clip
→ dub any video into 646 languages
→ generate audiobooks, dictation, transcription
→ pick from 14 TTS engines instead of one

ElevenLabs supports 32 languages. this does 646.

no per-character billing. no usage caps. no audio ever leaves your computer.

save this for later.

repo below

[Hansen Liang @HansenIsSo]

@petersykim Yeah. Hooks are very useful

[Hansen Liang @HansenIsSo]

Crossed over PR #2000 last night. Here're a few things I learned:

1/ Nail what “done” means before it starts.
2/ Do proper user testing. Agents will get most right, but rarely with the polish you need.
3/ Writing down a rule doesn’t mean the agent will follow it. Enforce it in code.
4/ Prioritization still matters. Even though building got faster, saying "no" is still the most important art of shipping a good product.

What have you learned recently about AI coding?

#aicoding #agentic

[Hansen Liang @HansenIsSo]

I did not expect @claudeai to send me a video today...

I've been working remotely as I travel, and use Claude Code's Remote Control feature to work across my agents to ship while I'm on the road. Part of the job required Claude to build out a progress bar component. So, as I required in my repo prompts, it needs a UX review from me.

When Claude realized I couldn't access my local to see the draft components, without me prompting, it just uploaded a video to the Linear ticket... here's that video.

I'd seen people demo this type of stuff before (e.g. for agents to attach a demo video to each bug fix / feature), but was still pleasantly surprised when it did so without me explicitly asking! (Opus 5)

#opus5 #claudecode

[Hansen Liang @HansenIsSo]

Who else has an agent that swears?

[Hansen Liang @HansenIsSo]

@jetfam @AnthropicAI @DarioAmodei Literally thinking about doing the same. Just posted about this.

[Bradley Bernard @bradleybernard]

PSA: GPT-6 Astra is out and it's time to audit your AGENTS.md and SKILL files, due to the new model!

Here is the Model guidance doc which should help: https://t.co/92b9UQgskB

Point Codex at that doc and let it rip in your repo, to improve Astra's performance!

[Hansen Liang @HansenIsSo]

🛑 PSA: if #gpt6astra just showed up in your model selector...

DO NOT UPDATE your #Codex client. Whatever mechanism @OpenAI is using to do the rollout seems to be client-dependent. 🫡

[Hansen Liang @HansenIsSo]

@Anthropic's pricing on Fable 5.1 is driving users away.

I have the 20x plan. I've long been a loyal customer of #ClaudeCode. But this is silly.

I can put ~1 major task on #Fable5.1 PER WEEK. I have to avoid running other agents in parallel (so tasks don't stall due to 5-hour limits).

Yes, Fable 5.1 is a great model. But with #GPT6 Astra on the horizon and open models not far behind, this type of pricing and limits will drive many power users away.

[OpenAI @OpenAI]

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

[Hansen Liang @HansenIsSo]

hello

[John Ternus @johnternus]

hello

[Hansen Liang @HansenIsSo]

We asked @ChatGPT how it feels to be an AI. 🤔

And it just kept getting crazier. 😳

#gptlive #voiceai

[Hansen Liang @HansenIsSo]

@AnthropicAI please add this #codex feature that hides thinking / working steps upon completion to #claudecode! 🙏

Seriously, nobody wants to read every step when we're running this many agents.