Nox

finding the AI tools, open-source repos & prompts actually worth your time.

[Nox @heynoxyz]

GOOGLE JUST GAVE CLAUDE CODE AND CODEX A CLI FOR EVERY GEMINI MODEL

gemini-api is one go binary for mac, linux and windows

- generate images, video, music and voiceovers
- analyze files, transcribe audio
- launch managed agents in the background
- --dry-run shows the request without calling the api

your agent asks for a hero image, it lands in ./assets

bookmark this

[Nox @heynoxyz]

@anya_yfa7 Haiku 5.5 need to upgrade

[Nox @heynoxyz]

ANTHROPIC'S S-1 IS OUT AND THE NUMBERS ARE WILD

H1 2026 revenue: $16.29B (vs $1.16B in H1 2025)

→ Q2 revenue up 1,368% YoY
→ ~250 customers paying $10M+ a year
→ operating cash flow +$3.9B in H1
→ ticker: ANT on Nasdaq

the catch: H1 net loss was $80.7B ($72.4B of it is convertible note revaluation) and it's locked into $110B+ compute deals with both Google and Amazon

buying ANT at IPO or waiting?

[Nox @heynoxyz]

GPT-6.1 TERRA MATCHES ASTRA AT COMPUTER USE FOR 1/10 THE PRICE

$1 in / $5 out per 1M tokens. astra is $10 / $50

openai's own numbers:
→ Terminal-Bench 4.0: 50.3% at max effort (GPT-5.6 Terra was 21.5%)
→ error rate on hard prompts: 53.6% to 16.7% at low effort
→ 3x+ faster than astra and 6.1 sol

the catch: all self-reported, and openai still says use sol or astra for the hardest coding. prompts over 272K tokens bill at 2x input

moving your agents to terra or staying on sol?

[Nox @heynoxyz]

@Cerevia_ Gpt astra bro

[Nox @heynoxyz]

@aGim_asf my bet is a coding refresh on the 28th, astra or something with a new name

[Nox @heynoxyz]

@EnvolDev beating sol and kimi on a free model is wild ngl

[Nox @heynoxyz]

@OwnYourCompute makes sense, decode being bandwidth bound explains why tok/s barely moved. how much did prefill go up on the 3080?

[Nox @heynoxyz]

YOUR OLD RTX 3090 WAS HOLDING BACK ~80% OF ITS BF16 SPEED 😳

gpu-unlock is a free linux kernel module (8.8K stars in 8 days)

3090 Ti, BF16 GEMM:
before: 87.5 TFLOPS
after: 157.4 TFLOPS

the catch: llm prefill gets +14-36%, token generation barely moves. main repo is for 30-series cards

would you flash this on your main rig?

[Nox @heynoxyz]

@MrTacticalX 2 min vs 9 min at less than half the cost is hard to argue with

[Nox @heynoxyz]

official post: https://t.co/4vTPrwOPfB

model id is gpt-6.1-terra. computer use + tool calling need the Responses API. batch/flex are 50% cheaper on top

[Nox @heynoxyz]

GPT-6.1 TERRA MATCHES ASTRA AT COMPUTER USE FOR 1/10 THE PRICE

$1 in / $5 out per 1M tokens. astra is $10 / $50

openai's own numbers:
→ Terminal-Bench 4.0: 50.3% at max effort (GPT-5.6 Terra was 21.5%)
→ error rate on hard prompts: 53.6% to 16.7% at low effort
→ 3x+ faster than astra and 6.1 sol

the catch: all self-reported, and openai still says use sol or astra for the hardest coding. prompts over 272K tokens bill at 2x input

moving your agents to terra or staying on sol?

[OpenAI @OpenAI]

GPT-6.1 Terra: delivering fast and capable computer use at a lower price.

GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

[Nox @heynoxyz]

@jjohana $0.095 a task for a 41 on the index is the number that stands out

[Nox @heynoxyz]

@Bhavani_00007 terra's whale breach looks cleaner than astra's here, at that speed too

[Nox @heynoxyz]

@jvr0x well specified instructions only is a fair filter, that covers most agent glue work

[Nox @heynoxyz]

@GrokInsider if it really holds at 1/10 astra pricing, a lot of sol usage moves over tonight

[Nox @heynoxyz]

@HarshithLucky3 launch day codex retry limits, every single time lol

[Nox @heynoxyz]

anthropic's s-1 says $16.29B revenue in h1
pretty sure half of that is me typing "continue" in claude code

[Nox @heynoxyz]

@JackDishman prompt caching is the boring one that saves the most tbh

[Nox @heynoxyz]

@BeidiChen training the harness instead of the weights is such a cheap win if it holds up