theo vs Sentry CEO: How Much Cheaper Has AI Intelligence Really Gotten

On October 8, developer theo and Sentry founder David Cramer (@zeeg) engaged in a multi-round public debate over whether AI model intelligence got orders of magnitude cheaper within a year, each offering their own math and data in a quintessential clash over AI cost narratives.

Confirmed

• The debate started when Cramer questioned whether current models offer an order-of-magnitude price advantage over a year ago, arguing that comparisons to models from years past are meaningless; theo fired back that workflows a year ago were still "select text in the editor, cmd+k, then manually read the code," which has completely changed, with costs down more than 10x
• theo's specific math: 3 years ago there were no reasoning models or tool calling, so no comparison is possible; he cited the o1 (2024) era as an example, claiming same-tier tasks fell 55x in cost within a year
• theo also noted: last September GPT-5 was the strongest model, but today the same performance tier is available via DeepSeek V4 Flash, with input price dropping from $1.25 to $0.06—21x cheaper (he called it nearly 20x)
• Cramer's rebuttal: pick tasks that a three-year-old model could already get right, recalculate at today's prices, then extrapolate across the full task spectrum, and "orders-of-magnitude cheaper intelligence" doesn't hold; older models are being retired, so actual running costs haven't dropped significantly
• Cramer said he hasn't hand-written code since July of last year, and doesn't deny models have improved, but argues you should compare the cost today of tasks a model from 12 months ago "could already do"
• Cramer said he has about six months of billing data: in absolute dollar terms, costs for tasks at the same intelligence level haven't dropped notably; falling prices buy "higher-tier intelligence," not "the same intelligence significantly cheaper"
• Cramer also griped that Google deprecated a model they used happily in production, replacing it with a pricier new one they didn't need—it didn't actually save money
• Cramer pointed out that a 10% benchmark improvement is barely noticeable to ordinary users, and cited Luna as a positive example—models that are genuinely better on both price and intelligence are the only fair comparison; he added that even aligning benchmarks to real use cases, you can't find a pricing gap anywhere near 10x, and benchmark scores get conflated with real-world value

Unconfirmed

• Both sides' cost figures (theo's 55x/21x drops, Cramer's six months of billing data) are personal claims with no public breakdowns or third-party verification

Why It Matters

• The debate hits a core divide in AI cost narratives: falling token prices don't equal falling real-world production spend—model retirements, workflow upgrades, and rising task complexity all eat into theoretical savings. For enterprises relying on LLM APIs, how they define "same intelligence" and calculate true costs directly shapes procurement and budget decisions

2026-10-08 ~ 2026-10-08 · 9 related posts

• theo's math: same intelligence now ~20x cheaper via DeepSeek vs GPT-5 a year ago — theo · 2026-10-08
• [source] Sentry CEO pushes back: benchmarks aside, real-world task costs haven't dropped orders of magnitude — zeeg · 2026-10-08
• Zeeg challenges the "cheap intelligence" narrative: 10% benchmark gains are imperceptible — zeeg · 2026-10-08
• theo counters Zeeg: same task 55x cheaper in a year with Gemini 3 Flash vs o1 — theo · 2026-10-08
• Zed founder disputes 'intelligence getting 10x cheaper a year' claim — zeeg · 2026-10-08
• [source] theo challenges Sentry CEO: same AI coding tasks now over 10x cheaper than a year ago — theo · 2026-10-08
• [source] Sentry CEO: Six Months of Data Shows Same-Level AI Intelligence Isn't Actually Cheaper — zeeg · 2026-10-08
• Sentry cofounder blasts Google for deprecating a key model, says LLM pricing hasn't gotten cheaper in 6 months of data — zeeg · 2026-10-08

1 near-duplicate retellings: zeeg