elsegundo.xyz · University of El Segundo

What does an hour of AI actually get you?

A plain guide for neighbors, Main Street shops and anyone within the 25 miles who is about to use AI at work or at home. Real numbers, read this week, with the guesswork labeled.

Figures read Nov 8, 2026.

The hour, engine by engine

Each bar is one hour. Every slice is one hard test task the AI finished inside that hour. Many thin slices means fast. One wide slice means it is still working on the first problem.

Score is the Artificial Analysis Intelligence Index (higher is better). Meta's Muse Wonder and OpenAI's open gpt-oss models showed no time or cost figures on the board, so they are left out.

The effort dial matters as much as the brand

The same model at a lower effort setting finishes far more tasks per hour for a few points of score. For everyday work, medium or high effort is usually the better buy.

What $20 buys at API prices

Hours of nonstop, flat-out work that twenty dollars pays for at each engine's list price. A $20 monthly plan is not sold this way, but this shows why plans meter you in 5-hour windows and weekly caps.

Measure it against your own hour

Everyday units

The benchmark tasks above are hard ones. Most daily jobs are small. Rows marked estimate are worked out from published speeds and prices, not measured.

A 200-word emailSonnet 5.5 writes about 154 tokens a second
~2 sec, under 1¢estimate
A 1,500-word report draftAbout 2,000 output tokens at $10 per million
~13 sec, ~2¢estimate
If it only wrote for an hour154 tokens a second for 3,600 seconds, about 0.75 words per token; reasoning tokens count too
~400,000 words, ~$5.50estimate
One image, Gemini 3.1 Flash Image at 1KGoogle list price
$0.067list
Images in an hour, gpt-image-2 mediumOne at a time, about 53 seconds each in a third-party test
~68 images, ~$4.50estimate
A 60-second video, Veo 3.1 Fast at 1080p$0.12 per second, from Google's pricing page
$7.20list
One Grok Imagine imagexAI says "starting at"
from $0.02list

Raw writing speed, output tokens per second

What the monthly plans say

Claude, ChatGPT and Google meter use in rolling 5-hour windows with a weekly ceiling, and Grok sells usage as multiples of its $8 Plus plan. None publishes an hour count, so the usage bar on your screen is the real unit.

Claude

Free $0
Limited use every 5 hours and every week
Pro $20/mo
More than Free; includes Claude Code
Max $100/mo
5x Pro every 5 hours and week; $100/mo API credit
Ultra $200/mo
10x Pro each week; no 5-hour limit when verified; $200/mo API credit

ChatGPT

Free $0
Basic access
Go $8/mo
More than Free
Plus $20/mo
Per 5 hours: about 5 to 45 Astra or 350 to 3,000 Luna messages
Pro from $100/mo
$100, $200 and $500 tiers; no 5-hour limit right now

Google AI

Plus $4.99/mo
2x the Free limits
Pro $19.99/mo
4x the Free limits
Ultra $99.99/mo
5x Pro
Ultra $199.99/mo
20x Pro

Grok

Free $0
Limited chat and agent requests
Plus $8/mo
Basic usage; Chat, Voice, Imagine, Build and API access
Premium $30/mo
4x Plus usage; latest models, video, Grok Bot
Super $100/mo
15x Plus usage; priority at peak times
Ultra $200/mo
50x Plus usage; earliest access to new models

API list prices, dollars per million tokens

ModelInputOutputSource
Claude Fable 5.1$10$50Anthropic pricing page
Claude Opus 5.5$4$20Anthropic pricing page
Claude Sonnet 5.5$2$10Anthropic pricing page
Claude Haiku 5.5$0.1$0.5Anthropic pricing page (prompts up to 100K tokens)
GPT-6 Astra$10$50OpenAI API pricing page (standard tier)
GPT-6.1 Sol$2$10OpenAI API pricing page (standard tier)
GPT-6.1 Terra$1$5OpenAI API pricing page (standard tier)
GPT-6 Luna$0.1$0.5OpenAI API pricing page (standard tier)
GPT-5.3 Codex$1.75$14OpenAI API pricing page (standard tier)
Gemini 3.1 Pro (preview)$2$12Google Gemini API pricing (prompts up to 200K)
Gemini 3.5 Flash$1.5$9Google Gemini API pricing
Gemini 3.1 Flash-Lite$0.25$1.5Google Gemini API pricing
Grok 4.7$2$6xAI models page

Google's newest Gemini 4 models were not priced on the Google page read on Nov 8, 2026; that page was dated June 2, 2026.

If you are new to this

Most of us will be handed an AI tool in the next two years, at the front desk, in a spreadsheet or on a phone. Here is how to start well and how to keep your footing.

Try these first

  1. Ask it to explain something from your own job, then check one fact it gives you.
  2. Paste a long email you need to answer and ask for a shorter reply in your voice.
  3. Give it a plan with real limits: a week of dinners under $80, or a Saturday with two kids and a 4 p.m. game.
  4. Upload a lease, a bill or a contract and ask what questions you should ask before signing.
  5. Build one tiny tool: a spreadsheet formula, a checklist, a one-page website.

Good at, today

  • First drafts, summaries and rewrites
  • Explaining a topic at your level
  • Writing and fixing code
  • Brainstorming options and naming things
  • Translating and changing tone

Still weak at

  • Recent facts unless it can search
  • Knowing your situation without being told
  • Admitting it is unsure; it can be wrong in a confident voice
  • Long jobs without checkpoints

Privacy basics

  • Never paste passwords, card numbers or ID numbers.
  • Check the setting that decides whether your chats are used for training.
  • For work data, follow your employer's rule before your own habit.
  • Memory features remember you across chats; you can review and delete what they keep.
  • Assume what you type could be stored.

Five measures worth keeping

  1. Minutes saved per task. Time a job by hand once, then with AI, including the checking.
  2. Cost per hour of model time. The chart above: about $0.70 to $35 depending on engine and effort.
  3. Errors you caught. Keep a rough count for a week. That number tells you how much to trust it.
  4. Your weekly limit. On a plan, the weekly bar is your budget. Spend the top model on the hard jobs.
  5. Who is accountable. A person signs off on anything that goes out. That does not change.

The next two years, and after

Looking ahead. This is a reading of trends, labeled as outlook. It is not a promise.

The same capability keeps getting cheaper.Epoch AI measured the price of reaching a fixed level of performance falling about 40 times per year, with a range of 9 to 900 times, and noted the fastest drops may not last. Expect today's top-tier hour to cost a fraction as much within two years.
Jobs get longer, not just faster.METR's last published reading had an early Claude Mythos preview finishing, half the time, tasks that take people at least 16 hours, with wide error bars. METR stopped updating that page in September 2026 because its test set ran out of long tasks.
Limits become the unit.Plans already sell 5-hour windows and weekly caps instead of message counts. Expect to budget AI like a phone plan or a utility bill.
Checking becomes the skill.When an hour of model time produces days of draft output, the scarce part is a person who can tell good work from plausible work.
The labs say discovery is speeding up.OpenAI's Greg Brockman posted that "frontier models, with lots of test time compute, are becoming the engine of scientific discovery," and Google DeepMind's Pushmeet Kohli that his teams are "witnessing a genuine phase transition in the pace of scientific discovery." Both are the labs' own words, from November 2026, not independent measurements.
Beyond two years.Compute behaves more like electricity: metered, everywhere, mostly invisible. The durable advantages are the ones a model cannot buy: judgment, taste, trust with people, and knowing your place and your customers well.

How this was measured

Sources

Questions or a correction? This is a University of El Segundo open lesson, and it will be updated when the numbers move.

Not an official City of El Segundo page. Drafted and checked by Claude, an AI model by Anthropic, from the public sources listed above.