Follow the money behind intelligence.
AI looks like software. Underneath, it is a system of chips, power, models, labor and capital. Choose a thread and follow it all the way down.
Today on the desk.
Four readings, each dated and linked to the tracker it came from.
New AI models
18shipped in September 2026
8 labs. Newest: DeepSeek V4.1 Flash from DeepSeek.
IFM
Ant Group
As of Sep 13, 2026
32GB DDR5 kit at retail
$425 to $560$13.28 to $17.50 per GB
Plus DDR4, HBM and server RDIMM street prices.
As of Sep 11, 2026
Model prices and lifecycle
38models with verified rates
2 more announced or rumored. Every rate is source-linked.
As of Sep 12, 2026
AI benchmarks
116benchmarks, 12 categories
85 still active, the rest saturated or retired.
As of Sep 12, 2026
What the numbers look like.
Three pictures from the same datasets the trackers run on. Each one is drawn at build time from a dated snapshot, not a live feed.
Intelligence you get per dollar you spend.
| Model | Accuracy | Cost per solved task | On the cost-efficiency frontier |
|---|---|---|---|
| Claude Fable 5.1 | 53.4 | $20.00 | Yes |
| Claude Opus 5 | 50.7 | $10.00 | Yes |
| Gemini 3.8 Flash | 41.2 | $1.50 | No |
| GPT-6 Astra | 45.2 | $20.00 | No |
| GLM-5.3 | 44.9 | $2.20 | Yes |
| Grok 4.5 | 39.1 | $3.00 | No |
| Kimi K3 | 34.5 | $6.00 | No |
| Qwen3.8 Max | 40.3 | $3.00 | No |
| GLM-5.3-Flash | 41.9 | $0.20 | Yes |
| MiniMax M3 | 29.6 | $0.50 | No |
| Inkling | 25.5 | $2.60 | No |
| Nemotron 3 Ultra | 23.4 | $1.40 | No |
| Devstral 2 | 9.4 | $0.90 | No |
| Llama 4 Maverick | 9.3 | $0.50 | No |
| DeepSeek V4 Pro | 20.8 | $2.00 | No |
A 32GB DDR5 kit, since April 2025.
| Date | Typical listing | Cheapest tracked |
|---|---|---|
| 2025-04-04 | $125 | $59 |
| 2025-06-07 | $126 | $55 |
| 2025-07-25 | $129 | $54 |
| 2025-09-11 | $128 | $63 |
| 2025-10-29 | $192 | $78 |
| 2025-12-16 | $457 | $126 |
| 2026-02-02 | $595 | $195 |
| 2026-03-22 | $565 | $390 |
| 2026-05-09 | $564 | $201 |
| 2026-06-26 | $576 | $375 |
| 2026-08-13 | $592 | $390 |
| 2026-08-25 | $599 | $390 |
| 2026-08-28 | $595 | $390 |
| 2026-08-31 | $596 | $390 |
| 2026-09-03 | $597 | $420 |
| 2026-09-06 | $597 | $420 |
| 2026-09-09 | $599 | $420 |
| 2026-09-11 | $588 | $425 |
The price of intelligence, 300x down.
| Date | Cheapest GPT-4-class model | Frontier flagship |
|---|---|---|
| 2021-11-01 | $60 | |
| 2023-03-01 | $60 | $60 |
| 2024-05-01 | $15 | |
| 2024-07-01 | $0.60 | |
| 2024-12-01 | $60 | |
| 2026-07-01 | $0.20 | $30 |
| 2026-08-21 | $20 | |
| 2026-09-03 | $50 |
Pick a point of entry.
Each path starts with the big picture, then moves into the evidence, tradeoffs and tools behind it.
AI costs
Token prices are the sticker. Context windows, retries, cache misses, subscription floors and the work an agent abandons halfway decide the bill. A model that looks half the price per million tokens routinely costs more per finished task, and the gap only appears once the same job is priced end to end across providers.
Start hereWhy Is My AI API Bill So High? 8 Causes and Fixes- The Price Reversal Phenomenon: When Cheaper AI Costs More
- GPT-5.6 vs Claude Opus 4.8 vs Fable 5: Which to Pick
Enter this topic worldModel your cost per taskCoding agents
The model matters. The harness around it usually matters more: how an agent reads a repository, what it retries, how much context it wastes, and where the loop stops. Two teams running the same model see cost and success rates diverge by multiples once the scaffolding and workflow differ.
Start hereThe 2026 AI Coding Agent Landscape: Leaders, Costs, Harness- Best Claude Code Agentic Workflows: 2026 Playbook
- Build Your Own Agent Harness or Buy Claude Code?
Enter this topic worldMap the agent landscapeModels & benchmarks
Leaderboards compress messy reality into a single number, then labs quote whichever number flatters the release. Suites saturate, contamination inflates scores, and the same model under different scaffolding can move ten points. What matters is which benchmarks still separate frontier models, and what each one refuses to measure.
Start hereBest Open-Weight AI Models in 2026- Are AI Benchmarks Reliable? How the Scores Get Gamed
- Harbor-Index: The AI Benchmark Where Nothing Tops 30%
Enter this topic worldExplore the value leaderboardLocal AI & hardware
Memory capacity sets what a machine can load, bandwidth sets how fast it answers, and utilization decides whether owning hardware beats renting it. With DRAM in a shortage cycle the arithmetic keeps moving: a build that paid for itself last quarter may not this one, and the break-even turns on how many hours the box actually runs.
Start hereWhy Local LLMs Got Good in 2026: Capability & Cost- How Much RAM to Run a Local LLM? 2026 Sizing Guide
- Local LLM Tokenomics: Self-Hosted Cost Per Token (2026)
Enter this topic worldCheck what your hardware can runCompute infrastructure
Behind every model sits a physical stack measured in gigawatts, substations, GPUs and construction schedules, and a financial one measured in debt, leases and depreciation assumptions. Capacity announced today energizes years out, so the distance between what is promised and what is running is where the interesting numbers live.
Start hereWhat Is an AI Data Center? Cost, Power, and Scale- AI Data Center Financing 2026: Inside the $700B Buildout
- AI Training Costs 2026: GPT-5.6, Claude Fable 5, Gemini
Enter this topic worldCompare inference providersAI markets
Adoption, productivity, publishing and national strategy show where AI creates value and where it only moves it from one balance sheet to another. Measured gains rarely match the claims, search referral traffic is repricing entire publishing models, and countries subsidize compute for reasons that are strategic before they are economic.
Start hereAI Search Broke the Web's Business Model. What Replaces It.- World Models: Why AI's Biggest Names Bet Billions in 2026
- Ex-OpenAI and Ex-Anthropic Startups: The Money
Enter this topic worldFollow AI adoption by country
What AI really costs.
Every number checked, dated, and linked to its source.
The price you are shown is never the price you pay. Your flat monthly plan is still metering tokens behind the scenes. Your coding agent burns most of its budget on tasks it never finishes. The data center sold in gigawatts arrives three years late, stuck in line for electricity.
This site works out the real figure and shows you where it came from.
Six subjects, one question: what does this actually cost to run?
- AI costs
- What one task costs start to finish, not the sticker price per token.
- Coding agents
- Why the setup around a model drives your bill more than the model does.
- Models and benchmarks
- Whether a leaderboard score holds up on real work. Usually it does not.
- Local AI and hardware
- Buy the GPU or rent the API? The math, redone every time RAM prices move.
- Compute infrastructure
- The power, the buildings, and the debt behind every model you use.
- AI markets
- Who is making real money from AI, and who is only moving it around.
The free trackers handle the moving parts. Model prices, GPU and memory street prices, benchmark definitions, provider rates, cost per task: each one is a dated dataset, read again at the source whenever the market shifts. You will never be handed a screenshot from six months ago.
Nothing on this site is guessed. A benchmark score comes off the leaderboard, not out of a launch post. A price links to the page it was read from. No ranking is for sale and no link pays a commission. Anything that turns out to be wrong gets corrected in the open.
Fresh from the field.
Check the numbers yourself.
Every figure links to its source right where it appears, not in a footnote at the bottom. The data behind the trackers is published as JSON, so you can cite it directly instead of scraping the page. Each post says how many sources it stands on and when those numbers were last checked.
The label matters as much as the link. A preprint is called a preprint. Documentation written by a vendor is called documentation, not independent proof. Coverage from another outlet says whose reporting it is.
Prices and scores are stamped with the date they were confirmed. A rate checked in March proves nothing in August, and leaderboards reshuffle faster than any page can keep up. When a number is estimated rather than measured, the assumptions are laid out so you can redo the math, or argue with it.
- Published analyses
- 143
- Free tools
- 33
- Open datasets
- 20
- JSON endpoints
- 9
- Last published
- Sep 15, 2026
Questions about this site
- What does Capital & Compute cover?
- The economics of AI infrastructure: what coding agents and model APIs cost per finished task, whether running inference locally beats renting it, what a benchmark score means once you leave the test set, and the capital, power and memory supply sitting behind the models. Six topic areas, 143 published analyses, and a set of maintained price and benchmark trackers.
- How are the numbers sourced?
- Every factual or statistical claim links to its primary source at the point the claim is made, names who produced it, and states the year. A preprint is labeled a preprint. Vendor documentation is labeled documentation rather than an independent finding. Posts built on research close with a full source list, and each figure records the date it was checked.
- How often are the price trackers updated?
- Each tracker carries an as-of date and is refreshed against primary sources when the underlying market moves, not on a fixed schedule. Model pricing follows provider rate cards; memory, GPU and storage prices follow spot indices, contract forecasts and retail listings. A reading is a dated snapshot rather than a live feed, so confirm anything time-critical at the source before quoting it.
- Can I reuse the data?
- Yes. The datasets behind the trackers are served as JSON at stable URLs so a number can be cited or rebuilt instead of scraped, and an llms.txt describes the site for AI answer engines. Attribution with a link back is appreciated.
- Who writes it?
- Capital & Compute publishes under the brand rather than a personal byline. The work is meant to be judged on whether its sources hold up, which is why every claim is linked and dated, why methodology is stated alongside the number, and why corrections are published rather than edited in quietly.