“AI is calamitously expensive and, despite years of promises of it getting cheaper for both those running AI services and its customers, costs have only ever increased.”
Our reading:
AI is extremely expensive, and its costs for both providers and customers have only ever risen despite promises that they would fall.
Assessment: Contradicted
What happened
Was it true for the period described?
Outcome: Contradicted
The evidence goes against what they said.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
He says that despite promises AI would get cheaper, 'costs have only ever increased' for both providers and customers. This is an absolute claim, so one documented price cut contradicts it. On 2025-06-10, OpenAI cut the price of the same o3 model by 80%, from $10/$40 to $2/$8 per million tokens (e1). In November 2025, Anthropic priced Opus 4.5 at $5/$25 per million tokens (e2). That is two-thirds below the $15/$75 that e1 lists for Opus 4. This is the same tier but a newer model. A survey covering the field to February 2026 reports benchmark cost per task falling 390x in a year (e3). Its authors say this largely reflects reduced test-time parallelism. Anthropic's total compute cost rose from $3.4 billion in Q1 2026 to $5.6 billion in Q2 2026 (e4). That fits a total-spending reading but does not show unit costs rising. On the contract's unit-price reading, the o3 cut alone contradicts the claim for customers.
Caveats
- If 'costs' means total spending rather than unit prices, the result could differ.
- The evidence does not show providers' unit costs, such as cost per token served.
- The Opus comparison is between two different models in the same tier. It uses Anthropic's announcement and an earlier third-party record of the old price.
- The per-task cost fall in e3 compares different models and largely reflects less parallel compute.
- 'Calamitously expensive' is a judgment and is not scored.
Evidence (4 items)
Goes against the claim: Simon Willison's Weblog,
OpenAI just dropped the price of their o3 model by 80% - from $10/million input tokens and $40/million output tokens to just $2/million and $8/million for the very same model.
Why it matters: Customers' price for the very same model fell 80%. That directly contradicts 'costs have only ever increased' for customers.
- Where
- Paragraph 1
- Publisher
- Simon Willison's Weblog
- Kind of source
- Secondary summary
- About events on
- Published
Goes against the claim: Anthropic,
Pricing is now $5/$25 per million tokens—making Opus-level capabilities accessible to even more users, teams, and enterprises.
Why it matters: Anthropic cut its Opus-tier API price to $5/$25 per million tokens, from the $15/$75 for Opus 4 recorded in e1's source. The price of top-tier Claude use fell by two-thirds.
- Where
- Paragraph 4
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Math
- (15-5)/15 = 0.667; (75-25)/75 = 0.667 (inputs: Opus 4 price $15 input / $75 output per million tokens (simonwillison.net, 2025-06-10); Opus 4.5 price $5 input / $25 output per million tokens (anthropic.com); rounded to nearest percent: 67% reduction)
- About events on
- Published
Goes against the claim: arXiv (Vahdati et al.),
Cost fell 390x in one year (o3's $4,500/task to GPT-5.2's $12/task), although this largely reflects reduced test-time parallelism.
Why it matters: Measured per-task cost on a benchmark fell sharply within a year. The authors note that the fall partly reflects less parallel compute.
- Where
- Abstract
- Publisher
- arXiv (Vahdati et al.)
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Yahoo Finance,
The cost of the computing power that runs its Claude models rose far less, to $5.6 billion from $3.4 billion.
Why it matters: A provider's total compute cost rose from Q1 to Q2 2026. That fits the total-spending reading of his claim, though it does not show that unit costs rose.
- Publisher
- Yahoo Finance
- Kind of source
- News report
- Measures
- Anthropic, compute cost (cost of computing power for Claude models), USD billions, company reported, 2026-04-01 to 2026-06-30
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- 'Calamitously expensive' is a judgment and is not scored.
- The checkable part is the absolute statement that costs 'have only ever increased', set against promises of AI 'getting cheaper'. Plain reading for customers: the price paid for comparable AI use (per token for the same or equivalent model, or per task) never fell. A documented price cut for the same model or capability tier contradicts 'only ever increased'.
- 'Years of promises' points to roughly the period since ChatGPT's launch. I use 2023-01-01 to the post date.
- Total spending, as opposed to unit price, is a possible alternative reading. It is noted in the flags.
Right if: Evidence that, from 2023 to the post date, the prices customers paid for comparable AI use and providers' unit costs only rose, with no meaningful cuts.
Wrong if: A dated, documented decrease in price for the same or comparable model or task within the period.
We wrote this test after the outcome was known.
We found 28 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, AI Is Slowing Down, paragraph 87.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2026-06-08.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, AI Is Slowing Down
AI is calamitously expensive and, despite years of promises of it getting cheaper for both those running AI services and its customers, costs have only ever increased.
- Where
- paragraph 87
- Published
- 2026-06-08
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Contradicted. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.