Vast | LightAI

Indie dev building with AI agents. Hands-on notes on new models, coding tools & workflows. Building LightAI: one key for the latest GPT & Claude.

[Vast | LightAI @lightai_hi]

Hi, I'm an indie dev deep in agent workflows and real-world AI.

Here I share hands-on tests of new models and coding tools, honest takes, and the pitfalls I hit.

I also build LightAI: one key for the latest GPT & Claude at 30% of official prices. https://t.co/lrxoIHLGwZ

[Vast | LightAI @lightai_hi]

ok google is actually cooking now. Hydrogen at $1.50 / $7.50, flat all the way to 1M, lands right between Terra ($1/$5) and Sonnet 5.5 ($2/$10)

the one line that made me wince: on Vertex, from Nov 13, a Gemini 4 output past 32K still bills you even if it times out
https://t.co/SJRFDL2P4E

[Logan Kilpatrick @OfficialLoganK]

Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone!

Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.

[Vast | LightAI @lightai_hi]

this week's model drops, in faces:

Fable 5.2: 😍 → 😱 (account banned)
Gemini 4 Carbon: 🤩 → 😶 (not on my plan)
Qwen 4 Max Preview: 🙂 → 😐 → 😑 (still thinking)
GPT-6.2: 👀👀👀

[Vast | LightAI @lightai_hi]

a test i'm keeping for side projects: what's your f()?

if more gpus or tokens wouldn't clearly mean more output, you don't have a scaling problem. the plane just isn't flying yet. a lot of "we just need more compute" projects are stuck right there

https://t.co/YhYKUZsQKY

[will depue @willdepue]

i am insanely underslept in this clip but i feel very strongly about this principle. most times your job in the world is to turn things into “just” f() of something (dollars/gpus/tokens/standard labor/…)

[Vast | LightAI @lightai_hi]

@theo "is this actually yours" is the best growth metric there is. means it's spreading without your name on it

[Vast | LightAI @lightai_hi]

The consistency trick that actually stuck: generate image #1, then never let go of it.

Everything after that is just edits off the anchor. Two sticker packs on gpt-image-2.6, about $4 in leftover API credit. The deadpan fox is the one I keep opening.

https://t.co/EmeF0CMm7x

[弎水 @vastxie]

拿pro订阅送的API额度，试了下刚出的gpt-image-2.6，一口气跑完两套微信动态表情包——《内阁打工人》和《精神稳定的藏狐》

加上前期测试，全流程一共也就4刀出头的赠金，四舍五入等于白嫖，分享下摸索出来的4步极简工作流：

1. 先抽1号锚点图锁人设：调generations接口先跑每套的第1张，参数开 output_format="gif" + size="320x320" + background="transparent"，底层会按3×3母版直出9帧透明动图。确认人设和印章/标牌比例OK后，抽出第0帧存成 anchor.png

2. 喂锚点图并发跑齐整套：剩下15张直接把 anchor.png 喂给 edits 接口并发跑，只改动作描述和配字。古风方印尽量固定用2×2四字词，现代标牌用粗体短词，两套32张下来零错字

3. 微信深色模式去白沫避坑：prompt里千万别惯性加 white sticker border（白色贴纸描边）或者标牌连接杆，不然帽翅和耳朵缝隙里全是抠不干净的白沫，切微信深色背景直接没法看。强制写死 strictly NO white sticker border, NO white outline，让印章和标牌直接悬空就OK

4. 脚本一键合规打包：最后用python按9帧联合包围盒居中缩放到240×240，保留前景原始RGB做GIF89a透明量化（不向白底预乘），顺手裁好cover和icon就能直接丢微信表情后台了

[Vast | LightAI @lightai_hi]

One API key for the latest flagships: Claude Opus/Sonnet 5.5, Fable 5.2, GPT-6 / 6.1. All at 30% of official prices.

• Native Anthropic + OpenAI APIs
• Claude Code: https://t.co/lrxoIHLGwZ (no /v1)
• Cache reads from $0.015/1M

Pricing: https://t.co/30Q4AO6onu