Adrian DONG @AdrianDongtech - Twitter Profile

Pinned Tweet

about 1 month ago

gemini 4 argon is the most hyped model almost nobody can use announced sep 30. today it's google's own teams and vetted cyber defenders. paid api and ultra: "rolling out soon," no date and from oct 9, free gemini users reportedly lose flash and pro. the 2026 story isn't capability, it's access WHAT GOOGLE ACTUALLY SAID (its sep 30 blog post) - access: trusted cyber defenders in the fairwind program, google's internal teams, and wiz. paid api customers and google ai ultra subscribers come next - price: $2 in / $10 out per million tokens as an introductory price, then $4 / $20. cached input is 95% off. it doesn't say how long the intro price lasts - output limit: 1 million tokens, up from 64k. the context window isn't stated - benchmarks with numbers: deepswe v1.1 77.9% (called state of the art), automationbench 51.3% (#1), lvbench 91.7% (state of the art), cwe-bench v1 68% (tied for first). the full table is an image, and all of it is google's own data - cyber: google says argon can find, validate and patch serious software vulnerabilities on its own, and that trusted defenders get a version without cyber guardrails WHAT IT HASN'T SAID - a date for paid api or ultra - the context window - scores for the comparison models, in text THE DOUBTS bloomberg reported on sep 30 (i've only seen recaps) that some google employees think argon is strong on benchmarks but weaker at real coding, with front-end design named. two people reportedly used the word "benchmaxxing." google disputes it, and deepmind's koray kavukcuoglu has said he's encouraged by the model. it's anonymous and unconfirmed THE PRICE, NEXT TO DEEPSEEK - deepseek-v4-pro, from deepseek's official pricing page: $1.32 in / $3.96 out at peak, $0.66 / $1.98 off-peak (the page is undated; third-party trackers show the same rates) - at argon's intro price, input costs about 1.5x v4-pro's peak price and output about 2.5x. after the intro period, about 3x and 5x - that's price only. i couldn't find a head-to-head on quality THE RATIONING - per officechai, from oct 9 free gemini users keep only flash-lite, and flash and pro move to paid plans. ai plus keeps flash but not pro. ai pro and ultra keep all three - so the best model is gated for cyber defenders, the next tier is gated for paid users, and the free tier shrinks to the smallest model WHAT HAPPENS NEXT (my read, not google's) 1. a paid api and ultra launch within weeks, probably with independent tests attached. if argon holds up outside google's own table, the "benchmaxxing" story dies. if it doesn't, it hardens 2. the intro price ends quietly. $4 / $20 is the number to judge it by 3. frontier prices stay high while small-model prices collapse. haiku 5.5 launched oct 7 at $0.10 / $0.50 per million. the gap between those two tiers is the thing to watch WHAT TO WATCH - a release date for paid api and ultra - independent scores against the best models from anthropic and openai - what developers say about real coding after a week, not the launch table - whether the intro price gets an end date - what chinese labs do at the top end sources: google blog, gemini 4 argon (sep 30); deepseek api docs, pricing page; officechai on gemini access changes (oct); bloomberg via recaps (sep 30)

AdrianDongtech's tweet photo. gemini 4 argon is the most hyped model almost nobody can use

announced sep 30. today it's google's own teams and vetted cyber defenders. paid api and ultra: "rolling out soon," no date

and from oct 9, free gemini users reportedly lose flash and pro. the 2026 story isn't capability, it's access

WHAT GOOGLE ACTUALLY SAID (its sep 30 blog post)
- access: trusted cyber defenders in the fairwind program, google's internal teams, and wiz. paid api customers and google ai ultra subscribers come next
- price: $2 in / $10 out per million tokens as an introductory price, then $4 / $20. cached input is 95% off. it doesn't say how long the intro price lasts
- output limit: 1 million tokens, up from 64k. the context window isn't stated
- benchmarks with numbers: deepswe v1.1 77.9% (called state of the art), automationbench 51.3% (#1), lvbench 91.7% (state of the art), cwe-bench v1 68% (tied for first). the full table is an image, and all of it is google's own data
- cyber: google says argon can find, validate and patch serious software vulnerabilities on its own, and that trusted defenders get a version without cyber guardrails

WHAT IT HASN'T SAID
- a date for paid api or ultra
- the context window
- scores for the comparison models, in text

THE DOUBTS
bloomberg reported on sep 30 (i've only seen recaps) that some google employees think argon is strong on benchmarks but weaker at real coding, with front-end design named. two people reportedly used the word "benchmaxxing." google disputes it, and deepmind's koray kavukcuoglu has said he's encouraged by the model. it's anonymous and unconfirmed

THE PRICE, NEXT TO DEEPSEEK
- deepseek-v4-pro, from deepseek's official pricing page: $1.32 in / $3.96 out at peak, $0.66 / $1.98 off-peak (the page is undated; third-party trackers show the same rates)
- at argon's intro price, input costs about 1.5x v4-pro's peak price and output about 2.5x. after the intro period, about 3x and 5x
- that's price only. i couldn't find a head-to-head on quality

THE RATIONING
- per officechai, from oct 9 free gemini users keep only flash-lite, and flash and pro move to paid plans. ai plus keeps flash but not pro. ai pro and ultra keep all three
- so the best model is gated for cyber defenders, the next tier is gated for paid users, and the free tier shrinks to the smallest model

WHAT HAPPENS NEXT (my read, not google's)
1. a paid api and ultra launch within weeks, probably with independent tests attached. if argon holds up outside google's own table, the "benchmaxxing" story dies. if it doesn't, it hardens
2. the intro price ends quietly. $4 / $20 is the number to judge it by
3. frontier prices stay high while small-model prices collapse. haiku 5.5 launched oct 7 at $0.10 / $0.50 per million. the gap between those two tiers is the thing to watch

WHAT TO WATCH
- a release date for paid api and ultra
- independent scores against the best models from anthropic and openai
- what developers say about real coding after a week, not the launch table
- whether the intro price gets an end date
- what chinese labs do at the top end

sources: google blog, gemini 4 argon (sep 30); deepseek api docs, pricing page; officechai on gemini access changes (oct); bloomberg via recaps (sep 30)

2

7

0

1K

Adrian DONG

@AdrianDongtech

Last Seen Users on Sotwe

Trends for you

Most Popular Users