the key thing to realize is the Haiku tokenizer consumes 1.5x the tokens for the same text compared to Luna.

So at minimum costing 1.5x for the same request and if input tokens cross 100k then all tokens cost 5x more

ok so they are comparing themselves to and claim to be beating GLM's last generation GLM 5.2 model, GLM 5.3 Flash is a monster, this is honestly embarassing

their static ip's were initially good and didn't get flagged, but now most sites are recognizing their ip ranges and blocking.

Muse's utility has significantly dropped with the blockages.

To become truly useful again they will need to use residential proxies, but I can't see them use those due to the risks and reputational damage.

Every time a new tool launches there's a good window where it can act as a scraping proxy. Back in the 2010s I used Google Translate for years to scrape hard targets like LinkedIn but these windows are much shorter these days as scraping is so much bigger.

One thing with Muse though is that you can scrape Meta's own sites which are currently all going under login walls and restricting discovery/search entirely.

similarly, gemini via google ai studio will happily run workloads over youtube that would be very annoying to run at scale. Especially if you needed to download the video.

Most residential proxies are already far more blocked and rate limited than any Meta IP. The internet is becoming a very weird place, where individual and "trusted" personal IPs are becoming a kind of commodity. Some sites are already scoring IPs based on usage activity - like a credit score. It's only a matter of time until this data is collated and commoditised. AI analysis is turning this up to 11.

That doesn’t track with reality as far as I can tell.

> Some sites are already scoring IPs based on usage activity - like a credit score. It's only a matter of time until this data is collated and commoditised.

Spamhaus is nearly thirty years old and the notion of electronic distribution of IP and domain "reputation" lists is at least that old.

I'll bet my hat that the Internet "advertising" [0] industry has been calculating and determining the reputation of individual households (if not individual users) for at least a decade.

[0] The scare quotes are because its primary purpose these days is for dragnet private-sector surveillance.

"risks and reputational damage"

Good one, I can't stop laughing. Thanks!

they can just use the user ip. i think grok already does this.

How can it do this? Wouldn't you see a hundred fetch requests in your browser network tab?

bruh have you heard of CSP?

I dont get it, why even make this announcement, nothing's available and only one real benchmark for comparison?

Only theory is team wanted this out before perf/promo reviews to kick it over the line and then its not their problem

Google (and the others) signed the White House Accord on Superintelligence yesterday. Basically a promise to self-regulate and release responsibly.

Maybe announcing it was delayed until the agreement, and releasing it was delayed to show compliance.

Likely just trying to appear relevant in the news cycle.

It's also kinda wild how the competition being at v6.1 makes 3.x feel ancient, at least saying you are at v4 now changes public perception a bit imo.

Remind me of Firefox changing version to catch up with Chrome

Judging by other comments here this model has been available internally for a few weeks. So the public release is likely imminent. The timing of the announcement might have to do with end of quarter for some obscure reason.

It was part of Sundar's Q3 OKRs. He is going up for promo this cycle.

Is he running for office in November?

Because none of the people who can use it (have been and will be using it) can talk about it if it hasn't been announced. NDAs and stuff.

By the time this is released to public it will already be outdated due to Opus 6 and Astra 6.5

promo already happened; perf is about 6 weeks away.

OpenAI released two model updates in the past week. 6 weeks from now is an eternity

Why would they want this out the day after openai dev day and after both anthropic and openai had major releases the previous week?

Sarcasm? Google's announcement today makes those other releases old news.

Did Jensen pass?

Nice write-up! The randomized SVD (Halko–Martinsson–Tropp) is a great practical fallback when you only need the top singular values/vectors.

Deepseek Flash is still much cheaper:

- lower input/output token pricing

- the cached token price is $0.0028/Million tokens, which is like 50-90% of tokens

DeepSeek Flash is a much worse model. Even DeepSeek Pro is much worse.

Oh how the turntables literally next day.

Slower/more verbose in my experience too. However, still self-hostable which is a major plus!

more like they were facing pressure from chinese models, and dropped prices and now their margins are squeezed

has claude escaped the lab???