March 2026 · Data

By the Numbers

Activity heatmaps, weekly token volume, model distribution, and cost comparisons — the raw data behind 78 days of building with AI.

Total Tokens

27.4M

processed

Sessions

1,200

over 78 days

Active Days

68/78

87% coverage

Current Streak

42 days

42-day longest

Activity Heatmap

Jan 28 – Mar 30 · Each cell = one day · Intensity = session volume

Jan 28

Feb 4

Feb 11

Feb 18

Feb 25

Mar 4

Mar 11

Mar 18

Mar 25

Less

Weekly Token Volume

Millions of tokens by model · Jan–Mar 2026

Model Mix

Share of total tokens

Sonnet 4.6 (79%)

Opus 4.6 (11.4%)

Opus 4.5 (7.1%)

Sonnet 4.5 (2.5%)

The ROI

My inputs vs. equivalent senior developer cost

Hours Invested

My hours~270 hrs

Dev equivalent~1,775 hrs

6–8×hour multiplier

Cost Comparison

My cost (Claude Max × 2mo)~$200

Dev equivalent (@$175/hr)$230K–$400K

~1,500×cost compression

37×

more tokens than War and Peace — processed in 78 days

Multi-Model Footprint

Claude Code + Codex · Deliberate routing, not redundancy

Claude Code

1,200 sessions

Days active68 / 78

Tokens27.4M

Median sessionMulti-turn

RoleDaily driver

Codex

136 sessions

Days active31 / 40

Tokens~1.1M (est.)

Median session1 message

RoleBug escalation

The median Codex session is one message — a surgical escalation when Claude produces a wrong diagnosis or gets the same bug wrong twice. The 42 web searches and ~2,900 tool calls in those sessions are mostly codebase exploration: reading files, tracing call stacks, isolating root cause before any fix is proposed.

Codex token estimates are derived from local session logs (visible text only), not billed API totals. Actual billed tokens would be higher due to system prompts and context packing.

What Was Shipped

5 production systems · All deployed · All serving real data

285

tests

R9 Bull Sales

Hedonic pricing model · 26K auction lots · Docker + Caddy + CI/CD

453

tests

R9 Herd

Prisma v7 · Breeding cycle AI logging · 8-gen COI · PWA

273

tests

R9 Lineage

192K+ animals · Maternal line scoring · Semen scrapers

tests

R9 Beef + Portal

Order management · Stripe · 4-app aggregation dashboard

745

tests

CSL Capital ML

BigQuery ML · $40M loan portfolio · 840 calls → 6

1,826total tests across all projects

Work with Nektar

This is what Nektar does for clients — embed in your domain, build the analytical layer, ship software that would have cost 100× more the traditional way.

Book a free data audit

How the data was collected

The charts on this page are pulled from a single corpus of work between February and June 2026. Every session ran through Claude Code or the Anthropic SDK with structured logging on, so the activity heatmap is literal session count per day, the token chart is the actual API usage ledger, and the model mix reflects which model handled each turn rather than a post-hoc estimate.

Two numbers tend to surprise people. First, the token-volume cumulative line crosses 100M about eight weeks in — heavy use of subagent fan-out, not a single agent running long. Second, the ROI bars compare equivalent contract-engineer hours at SLC rates, not list-price API cost; that's the comparison that matters when you're deciding whether to invest in an AI workflow versus hire.

For the narrative around what changed week to week — workflow shifts, failure modes, validated tactics — see the companion piece on 40 Days of Daily AI Engineering. For the workflow patterns that produced these numbers, the FYBY essay and smoke-testing notes cover the operating model.