Kagi

Let AI do your work. I show you how. Daily

[Kagi @Kaggii_]

@federicodonaton glad it clicked. Haiku as a sidekick is the part most people skip. happy to connect, I'm sending you a follow.

[Kagi @Kaggii_]

this is quietly huge

Anthropic just made Haiku 5.5 10x cheaper than Haiku 4.5, and told you to run it as a subagent under Opus 5.5 and Sonnet 5.5

most people send every agent task to one big model. you pay top price for file reads, lookups and summaries

the setup splits the stack:

> lead: Opus 5.5 or Sonnet 5.5 plans and makes the hard calls
> subagent: Haiku 5.5 does summaries, compaction, db queries and classification
> hard agent coding still goes to Sonnet 5.5 or Opus 5.5

the numbers:

> $0.10 per 1M input tokens. Haiku 4.5 was $1
> 20x cheaper than Sonnet 5.5 per token, for prompts up to 100K
> Devin Fusion holds 66.2 on FrontierCode with Haiku 5.5 as the sidekick (Cognition)

Opus thinks. Haiku runs.

you stop paying your smartest model to read files. you let the small one run

[Noisy @noisyb0y1]

https://t.co/q9ITDM8mqz

[Kagi @Kaggii_]

@Tensor_Verse exactly. big model thinks, small model fetches. you pay top price only for thinking

[Kagi @Kaggii_]

@Beaver_0x the stack line is the useful part. which of these engines renders fastest on a normal laptop?

[Kagi @Kaggii_]

@Argona0x the 30% vs 8% gap makes sense.
every new message makes the bot re-read the whole chat .so message 50 costs way more than message 1.

any simple rule for when to cut a chat and start fresh?

[Kagi @Kaggii_]

@cortexch not tested cleanly yet. same model in a fresh chat already helps. different family caught more in the paper, but that's a few cases.

[Kagi @Kaggii_]

1 person. A day job. Anthropic's AI hunting bugs for 31 days.
It found 171. 4 out of 5 got thrown out.

The win came from the setup, not a smarter model:

> AI #1 hunts. Its only job is to find suspects
> AI #2 tries to kill each suspect. Only survivors move up
> another model checks again, so one model's blind spots don't pass
> a human makes the final call. In the paper, AI found 0 bugs fully alone

Result: about 135 suspects dead. 4 became official security bugs (CVEs).

Same idea, another study (Monash University):
just asking AI "is this bug real?" caught 36.4% of fake alarms.
an AI agent that checks the code caught 95.5%.

I animated my version: Jev flags, Opus 5.5 checks, my spider pulls only what survives.

Try it today: when AI finds a bug, open a new chat and ask it to prove the bug is fake.

Finding bugs is cheap. Proving them is the job.

Your AI found bugs today?
Until something tries to kill them, they are guesses.

[Noisy @noisyb0y1]

https://t.co/q9ITDM8mqz

[Kagi @Kaggii_]

@cortexch not first hand. but a default that costs 6x is the easiest thing to forget. a spend alert per model would catch it.

[Kagi @Kaggii_]

@Beaver_0x a good backtest can still be overfit. how long do you paper trade before going live?

[Kagi @Kaggii_]

@Max_Buildog the last 3% is always dark mode. nobody asks. it just happens.

[Kagi @Kaggii_]

5 AI agents. 1 task. 60 seconds.
only 1 finished.

agents from OpenAI, Anthropic and two made-up ones.
(a skit, not a benchmark. you'll argue anyway.)

JEV was done in 4 seconds. the website was one button: YES. disqualified.

DOTS never left the start line. still "watching".

ASTRA took the lead. at 82% it clicked "delete project".

my money was on SONNET. it hit 97%. still adding dark mode.

OPUS sat at 0% for 50 seconds. just thinking.
then it built the whole site in the last 10.

the winner spent 50 of the 60 seconds thinking.

who do you bet on: the agent that answers in 4 seconds, or the one that thinks for 50?

[Noisy @noisyb0y1]

https://t.co/q9ITDM8mqz

[Kagi @Kaggii_]

@Scratcherrrr it watched all 5. nobody was safe. not even the one that sat still for 50 seconds

[Kagi @Kaggii_]

@slash1sol @boltdotnew one run for a whole round beats ten separate prompts. what happens when two comments contradict each other?

[Kagi @Kaggii_]

@noisyb0y1 breakouts are where slippage is worst: everyone hits the same level at once. with a -1% stop, even 0.2% slippage each way eats 40% of your risk

[Kagi @Kaggii_]

@Argona0x The actual trick from Anthropic's Opus 5.5 guide is missing here: a time budget.

Add a line like "elapsed 340s / 1200s" to every message. The lead agent then keeps more helpers running in parallel.

And the guide says quality stayed "comparable", not better.

[Kagi @Kaggii_]

@x_xShin What happens when two sources say opposite things? does the spider keep both tags or pick one?

[Kagi @Kaggii_]

@marfinxx read the paper. no Opus 5.5 or GPT-6 Astra in it. they tested 1.5B-8B open models like Qwen2.5 and LLaMA-3.1. The 58% on MATH500 is a 1.5B model

[Kagi @Kaggii_]

@andreysuperior fun part of that $300M story: it was Spread Networks, a fiber line Chicago to New York. soon after, microwave towers beat it anyway

[Kagi @Kaggii_]

@Teka1900 one catch on the bonus: chatting is free, but tasks it hands to Codex still count toward your limits. step 2 of that prompt will eat them