REAL DATA · MARCH 2026

We counted every token.
Here's what it cost.

In March 2026, we were debugging Agilint's own SaaS hub. Three bugs. One session. We tracked every single token — the input, the re-transmissions, the growing context window. This is the exact bill.

📅 March 2026 🐛 3 bugs fixed 💬 13 conversation turns ⏱ 2.8 human hours 🤖 Claude Sonnet 4.6
The Setup

A normal debugging session. Abnormal token bill.

Every time you send a message in Claude or ChatGPT, the model receives your entire conversation history as input — not just your new message. By turn 5, you're paying to re-transmit 40,000 tokens the model already processed in turns 1–4. The cost compounds silently. Most developers never see it.

"The danger isn't the LLM pricing. It's that you're buying the same context over and over — and calling it progress."

We decided to count it. Exactly. What follows is the turn-by-turn breakdown of a real debugging session, with real token counts, real costs, and a direct comparison to what Agilint would have charged for the same work.

Bug 001 · 4 turns

The 400 that shouldn't have existed

The launcher pushed project knowledge to the hub and got a 400 Unknown topic back. Simple-sounding. The root cause was an Express route ordering bug — a specific /knowledge handler registered after a generic /:topic catch-all. Four turns to find and fix two stacked issues.

Knowledge Push 400 Bug
4 turns api-server.ts
Turn 1
8,500
Turn 2re-sends turn 1
12,000
Turn 3re-sends 1–2
15,500
Turn 4re-sends 1–3
19,200
Input tokens
55,200
Output tokens
5,500
Cost at Claude Sonnet 4.6
$0.249
Human time
45 min
Bug 002 · 5 turns · 3 stacked root causes

The bug that kept coming back

The agent kept using claude-code despite Gemini Pro being configured globally. Five rounds of debugging. Three separate root causes stacked on top of each other — a wrong merge in the run payload, a priority inversion in config.ts, and a silent env var filter in terminal-helper.ts that dropped every LLM_* variable before it reached the orchestrator.

Each failed fix prompted another turn. Each turn re-sent everything. By round 5 the context window carried 38,000 tokens of conversation — for a bug whose fix was three lines across three files.

LLM Provider Not Applied — 3 Stacked Root Causes
5 turns api-server.ts config.ts terminal-helper.ts
Turn 1
23,000
Turn 2re-sends turn 1
27,500
Turn 3re-sends 1–2
31,000
Turn 4re-sends 1–3
35,000
Turn 5re-sends 1–4
38,000
Input tokens
154,500
Output tokens
6,000
Cost at Claude Sonnet 4.6
$0.554
Human time
90 min

"The fix was three lines. The conversation to find it was 154,500 tokens."

Bug 003 · 4 turns

The update that never arrived

The agent completed two features. The hub dashboard still showed them as pending. The orchestrator was pushing task updates to an in-memory store. The dashboard was reading from a JSON file on disk. Two data systems, never connected. Four turns to design and implement the fix across three files.

Task Status Not Propagating to Hub Dashboard
4 turns api-server.ts orchestrator.ts task-queue.ts
Turn 1
42,000
Turn 2re-sends turn 1
46,000
Turn 3re-sends 1–2
50,000
Turn 4re-sends 1–3
55,000
Input tokens
193,000
Output tokens
7,200
Cost at Claude Sonnet 4.6
$0.687
Human time
35 min
The Bill

Three bugs. One session. Here's the total.

Total input tokens (chat session)
402,700
55,200 + 154,500 + 193,000 — of which ~60% was redundant re-transmission
Total cost at Claude Sonnet 4.6
$1.49
+ 2.8 hours of engineer time

At even $60/hr, those 2.8 engineer hours cost $168 — making the LLM bill a rounding error. The real cost is attention: a senior engineer context-switching into debugging mode, losing flow, and re-explaining the same codebase to an AI that forgot it between tabs.

The Alternative

What Agilint would have sent instead

No conversation. No re-transmission. Three self-contained task prompts — one per bug — each with only the context the agent needed to fix it. The agent reads the files autonomously, makes the changes, commits, and opens the PR. No human in the loop.

TASK · bug-001 · P1 CRITICAL
~3,170 tokens total
Fix 400 error on POST /api/hub/agents/:projectId/knowledge Root cause: generic /:topic handler registered before /knowledge. Fix route ordering so /knowledge is first. Also fix save path — knowledge is saved to project.dataDir but hub reads from hubDataDirOf(project). Update save location accordingly. Files: api-server.ts
3,170 tokens
TASK · bug-002 · P1 CRITICAL
~4,805 tokens total
Fix LLM provider not applied when hub runs agent (3 issues) 1. api-server.ts: replace manual merge with resolveEffectiveLLM() 2. config.ts: read process.env.LLM_PROVIDER before savedLlm.provider 3. terminal-helper.ts: add k.startsWith('LLM_') to all 3 platform env filters so LLM_PROVIDER/MODEL/API_KEY reach the bat file Files: api-server.ts, config.ts, terminal-helper.ts
4,805 tokens
TASK · bug-003 · P2 HIGH
~5,545 tokens total
Connect task status pushes to hub task-queue.json 1. api-server.ts: add POST /tasks route (before /:topic) that calls TaskQueue.markCompleted/markFailed/markInProgress on hub-side JSON 2. api-server.ts: add POST /task-statuses bulk reconciliation endpoint 3. orchestrator.ts: push bulk status reconciliation at end of main() Files: api-server.ts, orchestrator.ts, task-queue.ts
5,545 tokens
Head to Head

Same bugs. Different economics.

✗ TRADITIONAL CHAT
Conversation turns13
Total input tokens402,700
Total output tokens18,700
Redundant re-transmission~60%
LLM cost$1.49
Engineer time2.8 hrs
Runs at 3amNo
Parallel on other bugsNo
✓ AGILINT AGENT
Task prompts sent3
Total input tokens13,520
Total output tokens6,000
Redundant re-transmission0%
LLM cost (Claude Sonnet 4.6)$0.131
Engineer time0 hrs
Runs at 3amYes
Parallel on other bugsYes
Same Model. Different Approach.

Claude Sonnet 4.6 — engineer vs Agilint

The model is identical. The pricing is identical. The only variable is how the prompt is structured and who drives the work. This isolates the value of Agilint's approach from any model pricing advantage.

INPUT TOKENS
$3.00 per 1M tokens
OUTPUT TOKENS
$15.00 per 1M tokens
TOTAL COST
ENGINEER TIME
RUNS OVERNIGHT
vs AGILINT
COST — SAME MODEL, BOTH APPROACHES
Engineer · Claude Sonnet
$1.49
Agilint · Claude Sonnet ✓
$0.131
Every Dimension

It's not just cost. It's everything.

Same model. The difference is structural — how work is framed, who drives it, and what happens when something fails.

DIMENSION
ENGINEER + CLAUDE
AGILINT + CLAUDE
💰LLM cost (3 bugs)
$1.49
$0.131 11× cheaper
🔢Input tokens consumed
402,700
13,520 97% fewer
🔁Redundant re-transmission
~60% of every prompt
0% — fresh prompt per task
⏱Engineer time required
2.8 hours active
0 hours fully autonomous
🌙Availability
9–5, one time zone
24 / 7 / 365
⚡Parallel execution
1 bug at a time
Multiple agents simultaneously
🎯Context precision
Noisy history + irrelevant turns
Only the files that matter
🔄Retry on failure
Engineer must re-engage
Auto-retry up to 3× per task
📋Audit trail
Chat log (not code-linked)
Git commits linked to roadmap
🧠Context window pollution
Grows every turn — model degrades
Reset each task — model stays sharp
💸Engineer hourly cost (est. $60/hr)
$168 in human time
$0 fully unattended
TOTAL COST OF THE SESSION
$169.49 Engineer + Claude ($1.49 LLM + $168 human time)
$0.131 Agilint + Claude (LLM only — zero human time)
11×
cheaper. Same model. Different approach.
$1.49 with chat-based debugging. $0.131 with Agilint.
Claude Sonnet 4.6 in both cases. The difference is 97% fewer tokens and zero engineer hours.

Your roadmap deserves better economics.

Stop re-paying for context your AI already read. Define your tasks once — Agilint handles the rest.

Request Enterprise Demo → See all features

Methodology: token counts measured from live API responses. Costs calculated at published Claude Sonnet 4.6 ($3/$15 per MTok) and Gemini 2.0 Flash ($0.075/$0.30 per MTok) rates.