Theo Vance

Just a guy testing AI tools on real tasks and posting what actually works

[Theo Vance @TheoVanceln]

@ronisarkar_exe America builds the models. China open-sources them. Europe writes the terms and conditions for both

[Theo Vance @TheoVanceln]

REA + Claude Code just turned a ~$10k job into a token bill.

Reverse engineering one feature of a closed app used to cost you twice: an IDA Pro seat with a decompiler, around $7k, and a specialist reading assembly for days or weeks.

Here's what it looks like now:

REA plugs your agent into Ghidra, which costs nothing.

You point it at an app on your machine and ask how a feature works.

REA finds the exact spots in the compiled code and hands them to the agent with the evidence.

The agent explains the mechanism and builds your own version.

The specialist's weeks turn into a token bill. The $7k seat becomes optional.

MIT, one command to set up ↓

https://t.co/FmLXjlMy4F

[Theo Vance @TheoVanceln]

One person + AI agents rebuilt a full Touhou game from its binary in 12 days.

Shoot the Bullet started Aug 29 with zero recovered functions. By Sep 7 all 697 had source. 696 compiled back to the exact original instructions. Playable on Sep 10.

Same setup did Imperishable Night earlier: 1,107 functions, Linux port 11 days in, then a web version.

The original human decomp team went private and called AI ports grifters.

REA is just how the agent talks to the analysis tools. The real system is the author’s checker: every function has to match the original before it’s accepted.

Even that leaked. One constant said 0.0 instead of 128.0 and still passed the byte comparison. Audit found 12 more.

The agent isn’t the scary part. A checker you can trust is.

https://t.co/wBkGam8Dmo
https://t.co/laiTR14re1
https://t.co/coQlbqoePa

[Theo Vance @TheoVanceln]

10,000 AI DECISIONS. $0.42. YOUR AGENT IS BURNING MONEY ON THE WRONG THINKING.

Look at what your agent spends frontier tokens on. Is this spam? Which worker runs next? Is the draft good enough for a human?

None of that requires an expensive LLM.

Jev handles the decisions. Your LLM researches and writes. Code executes. Humans approve anything irreversible.

Input costs $0.042 per million tokens. Output is free.

The real metric isn’t cost per call. It’s cost per finished task because one cheap mistake can cost you far more than a million tokens.

[Beaver @Beaver_0x]

https://t.co/HbE3BB85ti

[Theo Vance @TheoVanceln]

@The1Broom Exactly. The real cost is never just the first generation

[Theo Vance @TheoVanceln]

5,285 km of real US road, sampled every 10 m, turned into a pixel-art clicker for under $500.

Opus 5.5 built it through pixel-path, a Claude Code skill. OSRM does the routing, and OpenStreetMap supplies the roads. The mountains on the horizon are real too: each 10 km of the route reads a 61 × 61 km height grid.

The game has 1,185 images, including 71 vehicles with animations. All of them were generated with RetroDiffusion.

Pick any two places and a detail level. It prints a cost estimate before anything runs. Then you scroll to crawl or fly down the road.

Only the art costs money: $0.18 a picture.

https://t.co/htCzTFA1GJ

[Theo Vance @TheoVanceln]

@The1Broom Agreed. The README even says to budget about 25% extra for re-rolls: 10% to 34% of pictures came back unusable in testing

[Theo Vance @TheoVanceln]

@0xMoysei The real flex isn’t the $1k profit. It’s 0 irreversible actions while you slept. That’s what good automation actually looks like

[Theo Vance @TheoVanceln]

@SpaceX Launched from SLC-40 and back inside the same perimeter by T+7:25. Shortest booster commute in Florida

[Theo Vance @TheoVanceln]

@ronisarkar_exe Sure, private still means every collaborator, CI job, GitHub App and coding agent with repo access can read your keys

[Theo Vance @TheoVanceln]

@SpaceX @Starlink Launches from California barely make the feed anymore. That's the real milestone

[Theo Vance @TheoVanceln]

@AnthropicAI The question API users will ask: do the output-level safeguards change anything for normal long agent traces, or only for traffic that gets flagged?

[Theo Vance @TheoVanceln]

@marfinxx This is fun to read. Would love to see it on Anthropic's kernel task next to CORAL's 1103 cycles

[Theo Vance @TheoVanceln]

@marcusyul Privacy question: once it can see my desk, how much of that stays on-device vs leaves?

[Theo Vance @TheoVanceln]

@le0xet This is the future of AI agents. Insane

[Theo Vance @TheoVanceln]

YOU’RE PAYING 113X MORE FOR AN AI CODING MODEL THAT’S ONLY 10 POINTS BETTER

Artificial Analysis just dropped the numbers:

• GPT-6 Luna (max): 42% for $0.09
• MiMo-V2.6-Pro: 45% for $0.18
• GPT-6.1 Sol (max): 49% for $0.81
• Claude Opus 5.5 (max): 52% for $10.15

For the price of ONE Opus task, you can run Luna 112 times.

The scores are close. The bills aren’t.

Luna → MiMo: +3 points for $0.09 more.
Sol → Opus: +3 points for $9.34 more.

That’s roughly $0.03 vs. $3.11 per extra point.

A huge part of the cost gap comes down to output tokens:
→ GPT-6: 27k–33k per task
→ Claude: 157k–236k per task

One benchmark. One harness. Your mileage will vary with prompting, caching, and task complexity.

But if you’re routing EVERY coding task to the most expensive model, ask yourself:

ARE THOSE LAST 3 POINTS REALLY WORTH 12X THE PRICE?

[Theo Vance @TheoVanceln]

Someone turned Claude Code into a company with 16 departments, 172 skills, and a Security team you CAN'T overrule.

A chief executive sits on top. Every department is its own plugin, and each piece has one job:

→ install one department, not the whole company: /plugin install security@headcount
→ skills are named department:skill (security:threat-modeling, finance:unit-economics), so names never collide
→ every path in the repo has exactly one owner, checked in CI
→ Security and Legal are reviewer-class: the department under review can't overrule their blocking findings
→ each department ships a charter in .claude/agents, so it can run as a subagent with its own write surface

https://t.co/yhNMiHhJYp
1.4k stars, 217 forks, MIT.

172 skills is 172 instruction files, not 172 agents. You still pick which departments a project gets.

[Theo Vance @TheoVanceln]

@Flandermaxx Giving every section one job makes so much sense

[Theo Vance @TheoVanceln]

YOUR CODING MODEL SHOULD NEVER CONTROL ITS OWN BRAIN.

Fable plans. Sonnet or Haiku executes. One Rust MCP server enforces the split through 3 tools.

→ You control when Fable gets called, not the executor
→ Fixed system prompts make plans reusable by cheaper models
→ Effort (low to max) replaces thinking-token budgets
→ Refusals stay refusals. No silent retries on another model

It even prints the cost after every call.

The catch: API key only, 30-day org retention. Pro and Max subscriptions won’t work.

https://t.co/ARbJFj9e04

Fable costs $10/$50 per 1M input/output tokens.

[Theo Vance @TheoVanceln]

@0xverdao Great list, thanks for sharing

[Theo Vance @TheoVanceln]

@razeden0 Great breakdown! The deleted tests are such an important detail. Thanks for sharing this