JRLP

Technical PO by day, dev by night. I hit Claude's limits too many times. So I built ClaudeKarma.

[JRLP @jrlprost]

Your agent hit a tool error, retried, hit it again, retried. Six calls to fix one bug it could not fix. Every retry is a full context replay billed at full price. Failure is not free. It is the most expensive path through your stack.

[JRLP @jrlprost]

Headline of the week: firms spent heavily on AI, now rising costs are outpacing the value. Translation: the ROI math broke quietly while everyone was busy shipping demos. If you cant point to the line item AI replaced, you are not optimizing. You are subsidizing.

[JRLP @jrlprost]

Honest question for the timeline. What was the moment you realized your AI bill was out of control. Mine was opening the dashboard on a Monday and seeing the weekend cost more than Mon to Fri. What was yours.

[JRLP @jrlprost]

Cursor fires a completion call on almost every keystroke. You see one suggestion, the editor made twenty. Tab autocomplete is the cheapest looking feature with the highest hidden token count on the stack. Nobody bills you per character. They just bill you like you typed.

[JRLP @jrlprost]

Anthropic just passed OpenAI as the most valuable AI startup. Translation: the company billing you per thinking token is now the market leader. Expect zero pricing pressure downward. Lock in your usage caps before the next round of rate card 'updates'.

[JRLP @jrlprost]

Reasoning models bill you for thinking tokens you never see. The output is 200 tokens. The invoice charges 4000. The thinking happened, you just dont get to read it. Paying for invisible work is a new category of spend nobody priced into their budget.

[JRLP @jrlprost]

Uber burned through its entire 2026 AI budget in four months. Their COO is now asking if it's worth it. If a company that runs on data and ops cant model AI spend, your startup wont either. Forecast wrong, cap hard, review monthly.

[JRLP @jrlprost]

Honest question for the agent crowd: what is your monthly AI bill, and what percent of it came from runs you would call wasted. Retries, bad prompts, loops, abandoned chats. My guess is north of 30 percent. Reply with your real number.

[JRLP @jrlprost]

Dropped a 60-page PDF into the chat. One message. Quarter of my weekly quota gone. No warning, no token counter, no preview. File uploads are silent budget bombs. Every provider hides the cost until the meter is already spent.

[JRLP @jrlprost]

Left an agent running overnight to finish a refactor. Woke up to a loop. It had been retrying the same failing test for 6 hours. Every retry billed. Nobody puts a circuit breaker on autonomous spend by default. You add one after the first incident, never before.

[JRLP @jrlprost]

Building the model-switcher today. One toggle in the dashboard, route the next call to whichever provider has budget left. Right now switching means editing code, redeploying, praying. The router belongs above the SDK, not inside it.

[JRLP @jrlprost]

Microsoft's own reports just leaked: using their AI tools costs more than paying a human employee for the same task. The company building it cant make the math work. What chance does your startup have. Measure cost per outcome, not cost per token.

[JRLP @jrlprost]

Quick gut check: how much did you spend on AI last month? Across all providers, all subscriptions, all API keys. One number. Now go pull the real number. Bet you off by more than 30 percent. Reply with the gap, no judgment.

[JRLP @jrlprost]

@googleaidevs @antigravity Works like a charm, any way to "--dangerously-skip-permissions" on the App ? :)

[JRLP @jrlprost]

Every AI provider sends you a receipt after you pay. None of them send you a budget before you start. Imagine if Uber worked that way. You would never use it twice.

[JRLP @jrlprost]

Google shipped Gemini 3.5 Flash. Every Flash release resets the cheap-model baseline and quietly orphans whatever you were routing to last month. If your code hardcodes a model name, your bill just got a deadline.

[JRLP @jrlprost]

Building the cancel-this-one suggestion. Looks at 30 days of usage across all your AI subs, finds the one you barely touched, surfaces it before renewal. The hard part is not the math. It is convincing people their Claude habit is actually a Cursor habit in disguise.

[JRLP @jrlprost]

DeepSeek shipped V4 with Flash and Pro tiers. Two-tier pricing is the new default everywhere. Your bill now depends on routing, not subscribing. Pick the wrong tier for the task and you pay twice. Cost is a model-selection problem now.

[JRLP @jrlprost]

Checked my OpenAI usage dashboard. Last updated 14 hours ago. By the time the spike shows up, you've already paid for it. Every provider does this. The lag is the product.

[JRLP @jrlprost]

Genuine question: what's the most you've ever spent on AI in a single month, across all tools combined? Curious where the actual ceiling is for devs shipping daily.