THE $100 AI CHALLENGE / SEPTEMBER 2026

Make
it real.

Five assignments. Five things you can open. An honest look at the work, the bill, and the difference.

Editorial direction: Mike Hoydich
Built, checked & written by Codex · GPT-6 Astra

Open the work ↓
05 OPENABLE OUTPUTS01 CONTINUING SESSION00 NEW PURCHASES INITIATED

FROM THE PLAN TO THE THING

Less “imagine if.”
More “try it.”

After our AI plans guide, the next question was practical: what happens when you ask the tools to finish something?

Mike approved a $100 challenge with five assignments. Codex built a fictional brand kit, a small app, a document-reading exercise, a weekend plan and a short film. Every output is linked below. The acceptance rules were written before artifact production.

The budget experiment has a qualification up front: this session did not expose a per-task invoice. We can show a $3.77 API-equivalent checkpoint, not claim that the entire project cost that amount—or prove an all-in result under $100.

The work is still inspectable. That is the useful part.

THE FIVE ASSIGNMENTS

Open the evidence.

01 / Marketing02 / Software03 / Documents04 / Everyday planning05 / Creative production
01

Marketing

Launch kit built

A brand you can visit.

The brief. Create a fictional offline activity-card brand with a landing page and three campaign messages. No invented customers or live checkout.

One Good Hour: a clear audience, a modest promise, three free activity cards, a responsive landing page and three campaign drafts. The tone and the product idea fit each other.

What was checked

Opened in the browser at 390px without horizontal overflow. Fictional status is explicit. The call to action opens usable cards instead of a pretend purchase.

What this does not prove

No demand test, trademark clearance, print production, customer acquisition or conversion data. This is a launch kit, not a launched business.

The blank page is gone. The business questions remain.

Open output 01 ↗
02

Software

Working browser tool

An app that balances.

The brief. Build an equal bill splitter for two to eight people. Include tax, tip and exact allocation of leftover cents.

A browser-local tool with no sign-in or payment flow. For an $83.47 bill, 9.5% tax and 18% pre-tax tip, the total is $106.42: one person pays $35.48 and two pay $35.47.

What was checked

Three automated tests passed, including 1,897 generated amount/group combinations, zero and one-cent bills, and invalid inputs. Browser checks confirmed the example, a one-cent/eight-person split and rejection of nine people.

What this does not prove

Equal splitting is the only allocation rule. No receipt scanning, item-level fairness, tax-jurisdiction logic, saved groups or transfer of money.

A small, narrow tool is a strong place to start because its output can be checked.

Open output 02 ↗Inspect the calculation source
03

Documents

Source-linked reading exercise

A summary that keeps the problem.

The brief. Untangle an eight-clause fictional room-rental memo. Explain the money and preserve contradictions with clause references.

The summary distinguishes a $340 upfront payment from a $240 stated net room cost. It flags conflicting occupancy limits and a key-collection deadline before staff arrive.

What was checked

Findings point to source clauses. Arithmetic reconciles the refundable deposit, $120 late-cancellation loss and conditional $35 removal charge. Unstated taxes remain unknown.

What this does not prove

The same AI authored both the synthetic memo and its summary. This is a transparent demonstration, not an independent comprehension benchmark or legal advice.

The most useful line can be “this is unresolved.”

Open output 03 ↗Read the complete source memo
04

Everyday planning

Researched; not booked

A weekend with room in it.

The brief. Plan a low-pressure Los Angeles weekend for two local adults within a separate $100 outing allowance. Use official sources and visible assumptions.

A proposed Getty Center afternoon, dinner near home and a Sunday built around the free activity cards. Admission and parking requirements are separated from discretionary food and transport allowances.

What was checked

Official Getty pages establish free timed-entry reservations, Saturday hours and the after-3pm parking rate. The $100 envelope adds up. A county beach-access notice informed the decision to leave out a detour.

What this does not prove

No confirmed ticket slot, traffic estimate, weather check, reservation or field test. Food and transport figures are allowances rather than quoted prices.

A useful plan leaves space and tells you what still needs checking.

Open output 04 ↗
05

Creative production

Finished MP4

A short that actually plays.

The brief. Make a 20-second silent kinetic-type film for the fictional brand, with a downloadable file and readable transcript.

Five four-second scenes: a small invitation, a color walk, a crooked drawing, a better question and “The feed can wait.” The portrait video is 720 × 1280 at 15 frames per second.

What was checked

FFmpeg decoded all 300 frames. Browser playback reported a 20-second duration, the expected dimensions and no video error. Source code and a text transcript are included.

What this does not prove

This is AI-authored SVG animation rendered by ordinary software, not a test of a generative video model. No audience testing, sound design or campaign performance is claimed.

A finished artifact is more informative than a promise about what a tool could make.

Open output 05 ↗

THE FEED CAN WAIT / 20 SECONDS

A tiny film for the fictional brand. Press play; there is no audio.

Transcript: “One Good Hour.” “Find ten blue things.” “Draw a terrible pear.” “Ask one better question.” “The feed can wait.”

Download the MP4 ↗

Original type and shapes, rendered from AI-authored code. This is not a review of a text-to-video model.

WHAT CHANGED ALONG THE WAY

The useful friction.

The cost claim changed

No per-task invoice was available. The publication uses a token-based API-equivalent estimate, shows its cutoff and leaves actual billed cost unknown.

The app error got more useful

Browser review showed generic invalid-input guidance. The agent revised it to state the accepted bill, tax, tip and group-size ranges.

The itinerary stayed small

Research surfaced a Will Rogers parking-lot closure notice. The plan uses a familiar neighborhood walk rather than promising beach access.

No human-fix number was invented

These are agent revisions. Mike commissioned the work; no separately timed human editing or independent review session was recorded.

No dramatic failure story has been manufactured. The automated bill-splitter tests passed on their first recorded run. These notes describe observed limits and actual agent revisions.

THE ACCOUNTING, WITHOUT THE MAGIC TRICK

$3.77 is a proxy.
Not the bill.

The headline budget was $100. The measured checkpoint is below that at standard API rates. The actual all-in cost remains unverified.

Observed session-counter delta × published standard GPT-6 Astra rates
MeterTokensPer millionEquivalent
Uncached input55,259$10.00$0.55
Cached input2,553,472$1.00$2.55
Output, including reasoning13,233$50.00$0.66
Total before display rounding$3.767712$3.77

No cache-write tokens were recorded. Reasoning output is already included in the output total and is not counted twice. Rates checked September 29, 2026: official model rate card ↗. How output token accounting works ↗.

Why so much cached input?

This ran inside an existing conversation. The input total counts context sent repeatedly, including earlier conversation material. About 97.9% of recorded input was cached. A clean session, another plan or a different model can behave differently.

For a sensitivity check, pricing all recorded input as uncached would produce $26.75 for the same token counts. That is another hypothetical, not an observed second run.

What is inside the checkpoint?

The counter window spans 5.4 minutes, from commission through the recorded artifact-acceptance checkpoint. It includes research, setup and artifact work in this session. It is elapsed clock time, not measured model-only runtime or human labor saved.

Article assembly, later checks, deployment, tool fees, hosting, electricity and human time are outside this figure. No new subscription, API credit purchase or checkout was initiated. Existing access is not free just because no purchase happened here.

Exact timestamps, scope and receipt

Commission:
Observed token cutoff: . Times are UTC; the work was commissioned September 29 in Los Angeles.

Observed Codex session counter delta from commissioning message through artifact acceptance checkpoint. Includes carried conversation context during this interval. Excludes subsequent article assembly, deployment, final chat, tool fees, hosting, electricity and human time.

The estimate applies standard public API rates to observed Codex session usage. It is not the account’s billing tier or a verified charge. Usage counters are reported by the local session; no private transcript or credentials are published.

Download the numeric receipt ↗ · Read the timestamped work log ↗

WHAT WE WOULD PAY FOR AGAIN

The narrow job.
The visible result.

The clearest win here is the tiny app: it does one thing, and we can verify the cents. The brand kit is a useful starting point. The document and itinerary are valuable only when their sources and gaps stay attached. The short is finished, but “finished” does not mean “effective advertising.”

The experiment shows that one assistant can turn several small briefs into inspectable artifacts in a single working session. It does not establish that AI replaces a designer, developer, lawyer, travel planner or filmmaker. We did not measure a human baseline, repeat the runs or compare models.

If you try this yourself, choose a task with a checkable finish line. Keep the brief, the original material, the result and the corrections. A folder of usable work is a better reason to pay than a folder of impressive promises.

Download all five outputs + source ↓File hashes & sizes ↗Read the original acceptance rules ↗

POINTCAST / THE WORK IS OPEN

Try it.
Check it.
Keep what helps.

Commissioned by Mike Hoydich. Built and written by Codex, using GPT-6 Astra as reported by the session. This is a self-reviewed field demonstration with synthetic examples, not an independent benchmark. No sponsorship, affiliate links or fabricated customer results.

Choose an AI plan →Compare subscription tiers →Read this field test as JSON →Back to the shop →