vorty (@vorty279) header
vorty (@vorty279) profile picture

I'm going all-in on the risks I just want to grab all those millions ✪ player @polymarket coo @okanedao

Joined February 2026
Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)Photo by vorty (@vorty279)
someone left 7 claude code agents running overnight. the panel said 2.5M tokens the logs said 30.3M. 12 times more, and the panel never said a f*cking word two lines in settings fix most of it do you leave agents running overnight?
6
1
15
808
subagents in claude code forget everything after 5 minutes. even on max one night, 7 agents, mac asleep. 75 wakeups, 75 full rereads panel: 2.5M tokens logs: 30.3M two lines in settings.json give them a full hour do you run agents overnight?
someone left 7 claude code agents running overnight. the panel said 2.5M tokens the logs said 30.3M. 12 times more, and the panel never said a f*cking word two lines in settings fix most of it do you leave agents running overnight?
5
13
282
accenture gave staff $29 a month for claude the docs say the average dev spends $13 per active day so the whole month lasts about 2 workdays and 10% spend over $30 a day. they would not make it through one how many days would yours last?
1
99
sonnet costs half of opus per token. so a swarm of sonnets saves money, right? not even close math on the example session from the claude code docs: > the edits themselves on sonnet: $0.25. on opus: $0.49 > but each worker boots up empty and reads the whole context again > 5 boots tack on $0.44–0.73 sonnet swarm: $0.74–1.03 plain opus: $0.55 haiku is a different story, one boot there costs a few cents and the swarm finally comes out cheaper than one big model. it takes one line in the subagent file: model: haiku how many workers do you run at once?
2
12
266
one of the biggest token burners at Accenture turned out to be f*cking PDF to slides in an internal note from June, reported by Startup Fortune, client group lead Stuart Henderson called it one of the company's main token chewers. Justice Kwak, who runs agentic AI there, confirmed it: reformatting documents was the heaviest load, development spent less running that through Opus is like spinning up a pulsar to move slides around: > on the example session from the Claude Code docs, Opus 5.5 costs about 30x more than Haiku 5.5 > the docs themselves suggest Haiku for simple subtasks > and the $29 monthly cap now cuts the same for the person converting decks and the developer who actually needs Opus full breakdown in the article which model runs your simple tasks?
11
284
claude code checks 4 spots top to bottom for a subagent's model and takes the first filled one leave them empty and the swarm rides your opus at $0.27 a worker. one model: haiku line and it's $0.009 got yours set?
2
8
259
in the spring, Claude Code's docs said an average developer at a company spends about $6 a day. after April 16 they quietly f*cking changed it to $13 Ed Zitron spotted it first. around the same time Anthropic's head of growth Amol Avasare wrote that engagement per subscriber was way up and the plans weren't built for this put that next to the Accenture story: > Anthropic's own estimate more than doubled in half a year > in October Accenture capped everyone at $29 a month > at the new number that's two workdays, at the old one it would have been almost five so people started living in agents all day, and companies saw it on the bill first full breakdown in the article has your own spend gone up this year?
14
466
there's a trick going around X: Opus writes the plan, a pack of cheap subagents does the edits in parallel vorty priced it on the example session from the Claude Code docs: > Opus alone: $0.55 > Opus plan + 5 Haiku workers: $0.10 > Opus plan + 5 Sonnet workers: $0.74 to $1.03 > Opus plan + 5 workers with no model line: $1.42 to $2.00 anyone on Opus who never wrote model: haiku lands in that last row. by default a subagent runs on your main chat's model, so the cheap swarm turns into five more copies of Opus the video shows it per worker: a Haiku firefly costs $0.009, an Opus owl about $0.27 every worker starts with an empty head and writes its own cache from scratch. that's why even a Sonnet swarm ends up pricier than one Opus full breakdown in the article does your subagent file have a model line?
3
19
558
in the video a brain swallows the whole terminal, a tree grows out of it, and thousands of sessions crawl out underneath. that's roughly what agent work looks like: one task, a pile of requests under it Claude Code will tell you what it cost. the /usage command shows the current session's price at Anthropic's list rates. if your company has a contract, an admin can plug in its own rates vorty broke down the example session from the docs on Opus 5.5: > input 1,200 tokens, $0.005 > output 5,300, $0.11 > cache write 50,000, $0.25 > cache read 940,000, $0.19 about $0.55 total, and almost 80% of it is cache. the agent resends the whole conversation on every step, and that's mostly what you pay for full breakdown in the article have you ever actually run /usage?
2
7
204
the goat in the video chews PDFs and spits out slides. that's roughly what one of Accenture's most expensive jobs looked like in a June internal recording reported by Startup Fortune, client group lead Stuart Henderson called PDF to slide conversion one of the company's biggest token chewers. Justice Kwak, who leads agentic AI strategy, confirmed it: reformatting documents used the most, engineering used less and the work itself is mechanical. move text and tables from one file into another vorty ran the example session from Claude Code's docs through Anthropic's price list: > on Opus 5.5 about $0.55 > on Haiku 5.5 with short context about $0.02, roughly 30x cheaper the docs themselves point you to Haiku for simple tasks full breakdown in the article which model makes your slides?
6
136
the beaver in the video is your agent, and the tree is the $29 a month Accenture now gives each employee for Claude he chews at $13 a day. that's what an average developer at a company spends per active day, per Claude Code's own docs the tree falls on the third workday. after that, without a project code, access just gets cut, and the month still has 19 workdays left vorty counted it in sessions too: > Opus 5.5 about 53 a month > Sonnet 5.5 about 106 > Haiku 5.5 anywhere from 315 to 1,574, it gets 5x pricier once a request passes 100k tokens and per Accenture's internal audio, one of the biggest token eaters there was converting PDFs into slides full breakdown in the article how many days would $29 last you?
1
1
10
403
a pelican swallows whatever you throw at him. vorty threw him all 13 free OpenCode models, one at a time he spat out exactly one, Muse Spark. it's the only one whose docs say outright that your requests train future Meta models the rest he swallowed and dropped into three bins based on what their privacy section says > green, keeps nothing: DeepSeek V4.1 Flash, LongCat 2.5, Step 5, Space Bunny > yellow: both Nemotrons keep logs, Cat Jam holds requests for 30 days, Jev says nothing about privacy > red: Big Pickle, Exo and both MiMo models may train on your code. Muse Spark ended up there too which bin is your model in?
1
13
630
"who even needs my code" is what almost everyone on OpenCode's free models thinks but along with your question, the agent sends the model everything it opened along the way: project files, plus command output with the logs. if it read your .env on the way, your keys went to the provider too vorty went through the privacy section for all 13 free models. four of them say that during the free period your data may be used to improve the model. that's Big Pickle with Exo, and both MiMo models. Muse Spark straight up trains future Meta models on it, and NVIDIA asks you not to send anything confidential to its Nemotron models that leaves four clean ones: DeepSeek V4.1 Flash, LongCat, Step 5 and Space Bunny does your .env sit in the same folder the agent works in?
1
14
406
in OpenCode the full name of Muse Spark 1.3 is Muse Spark 1.3 Contributor Free. the word Contributor is there for a reason the price table says Free. further down the same page, in the privacy section, it says something else: heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models so the contributor here is you. every request you send it goes into training vorty went through all 13 free models. four more say they may train on your data, Nemotron keeps logs. but only this one says straight out that it trains Meta's models would you trade your prompts for a discount?
1
9
383
two cars on the same highway. one kilometer is one session with the agent the blue one runs on paid DeepSeek V4.1 Flash, which keeps nothing. every kilometer it pays 7 cents, that's what an average session costs per OpenCode the grey one runs on a free model that trains on your data. the ride costs it nothing, but the whole way your files, logs and .env are flying out the window vorty did the math to the end: 100 km in the blue one is $7. the grey one is $0 plus 100 sessions of your code in someone else's training full breakdown in the article which one are you getting in?
1
13
412
on October 15 OpenCode announced DeepSeek V4.1 Flash is now free on their platform. the post blew up, around 800K views the same day, the old page of their own docs was still listing that model at $0.30 and $1.20 per million tokens. the old price is just lying there sunbathing, nobody touched it vorty opened both pages side by side. the new one says Free with a tiny note: free for a limited time. until when, it doesn't say most likely the docs just didn't catch up with the promo. still, their Free is worth reading together with whatever sits under the table full breakdown in the article do you ever actually read the fine print?
1
10
405
OpenCode has 13 free models. 8 of them come with a note about your data vorty scrolled their docs all the way down to the Privacy section. every model gets its own line there, and almost nobody reads it > 4 models keep nothing. the hyped DeepSeek is one of them > 4 may train on your data while they're free > NVIDIA logs everything and asks you not to send anything personal > Meta says straight up it trains its future models on your prompts one more keeps requests for 30 days, and the last one doesn't say anything at all full breakdown in the article have you ever opened the Privacy section?
1
9
351