sibidharan

What you think about attaching the chat transcript that lead to the final output when marking AI generated? it may be interesting to see how people will be feel a shame at how they generate slop vs genuinely using AI to come up with something. I see lot of VC applications follow this. But not sure how it will be applicable for HN.

Just need to bust low effort outputs!

I heard claude models are of trillions of parameters !!! 284B, 1M... I wouldn't trust on long running autonomous agents! But for the API costs, this is justified if a better hardware with bigger model is used and will be a claude equivalent for comparable quality at long context retrieval on long running autonomous tasks. At least for me that is important.

So being imperfect is fine... mmmm!!!

Its just me steering all tokens... Multiple projects for my bootstrapped Academy, running parallel researches on things I wouldn't do myself, and deriving the right patterns for the future. Using different Max on same machine. I have similar ~$22K on another machine, where I am working in parallel with the same 2xMax subscriptions!

So if the prices were to go high, I honestly have no idea at the moment. I might need a rehabilitation centre!!!

But I am mentally prepared that someday this will be gone... So make hay while the sun shines! I am doing the most complex of all works I want to do... And pushing the limits. Knowing how much it costs in actual tokens, I give me some kind of seriousness, like an invisible funding that I must use to grow!! Considering this like a launchpad... If any of the projects hit and scale and I get some funding in future, then I might not need to worry when prices go high. So either this or rehab lol !

It's been like this all along! We can write complex applications with HTML/CSS/Vanilla JS... But then.. React happened! UI rendering JS is bad - thats just my opinion. If server can render HTML and serve, whats better than that? Its ok for apps where browser is the source of truth like Figma, Google Sheets.. But using React for everything ... that is overkill. Most apps has server as the source of truth, data lives in server, and it can generate HTML right there.. No need for JSON+Hydrate, which is unnecessary- because server can generate HTML, the browser can understand HTML... Virtual DOM makes React Fast, not our apps fast..

AI didn't eliminate the need for frameworks.. It just picks the right one for the Job.. when plain HTML is ok, it chooses plain HTML.

I am running a self hosted Wireguard for my academy with ~15k student devices connected... Even during high disk IO, it's all RAM only and it never goes down. Once our NVME got bad blocks and was working degraded - but ... Wire guard is in-memory and bloody stable! Its the backbone of our academy for several years for students to access their cloud labs and connected devices.

1 Static IP, A linux machine onprem with necessary redundancy. Working like a clockwork.

It's a lottery however!

Social-networks are anti-AI! Here for soul-soul communication, not kinking for machines!

Which models are we talking about? Is there any degradation in quality, long context retrieval?

TL;DR = ~$22,720 total compute @ Opus 4.7 if no caching = $113,418 (5.9x more) - this is just one month on one server... I have 3 more servers like this where I work all time!

// Generated wit Claude

Ran this on my own ~/.claude/projects/ (933 sessions, 93,842 model calls, mix of main thread + spawned subagents). Numbers came out very close to yours in shape, different in scale.

cost.py (Opus 4.7 list rates, main thread only):

  cache reads (re-reading context)   21.69B tok   $10,843   56%
  cache writes (1h)                     678M tok   $6,781   35%
  output (incl. reasoning)             63.5M tok   $1,589    8%
  fresh uncached input                  1.6M tok       $8    0%
  TOTAL                              22.43B tok   $19,221

  if no caching: $113,418 (5.9x more)
  input:output ratio: 353:1
  cache hit rate: 97.0%

token_time_breakdown.py (179M unique tokens, 166h wall clock):

  reasoning (hidden thinking)   29% of tokens,  102h (61%) of time
  bash                           1.4%           23h (14%)
  writing tool calls             4%             14h  (8%)
  summaries                      2.5%            9h  (5%)
  reading/searching/web          1.6%            7h  (4%)
  subagents                      0.2%            6h  (4%)
  editing                        0.1%            5h  (3%)
  pasted attachments             25.3%           -
  typed prompt                   34.4%           -
  system+tools                   1.4%            -

reread_breakdown.py (per-activity share of billed input):

  reasoning           59.5%   (~$11.4k of the bill is re-sending old
                               hidden thinking back to the model)
  attachments         22.6%
  tool calls           7.8%
  bash                 3.0%
  reading/web          2.4%
  my prompt            1.6%
  summaries            1.5%
  system+tools         0.8%
  subagents            0.4%

main_vs_sidecar.py:

                          main         sidecar      combined
  sessions/agents          449         484           933
  assistant calls       63,820      30,022        93,842
  cache hit             97.0%        94.4%         96.6%
  turns/agent             142 (median 20, max 11,058 in one session)
                                       62 (median 44)
  reasoning % of out    82%          51%           77%
  cost @ Opus 4.7    $19,225       $3,495       $22,720

  sidecar = 32% of calls but only 15% of cost. Subagents are doing
  their job (cheap, focused, short context).

Same shape as yours: re-read dominates, reasoning is the biggest re-read line, caching is the only thing keeping it sane. The one that surprised me was a single 11,058-turn main session - some autonomous loop I forgot to kill. Going to grep for that.

Repo: github.com/Coral-Bricks-AI/coral-ai/tree/main/claude-code-token-xray