All topicsSTATE OF AI REPORT.

Is AI inference getting cheaper?

By Nathan Benaich (Air Street Capital)

Across five benchmarks, Epoch AI estimates that achieving a fixed score has become 47% cheaper each quarter since 2023, roughly 13-fold cheaper each year. That makes more work economical to delegate. It does not mean every company’s total AI bill falls: agents can make many model calls as they work.

Source: Epoch AI: The plunging price of thought.

Evidence you can use

Quarterly decline in cost at a fixed score

2023 onward. Analysis published 2026-09-22

Quarterly decline in cost at a fixed score
BenchmarkCost decline
AIME (OTIS Mock)Source47.1%
Chess puzzlesSource43.0%
FrontierMath, tiers 1-3Source53.1%
GPQA DiamondSource47.0%
Mystery game puzzlesSource44.0%

These are the model-free estimates in Epoch AI’s Table 1. The average quarterly decline is 47.0%. The comparison holds benchmark performance fixed. A forecast of total spending also needs usage, task difficulty, and the number of model calls.

Sources: Epoch AI: The plunging price of thought.

Data & methodology →Download CSVJSON

Cheaper intelligence creates more uses for it

Agents repeatedly call trained models as they work, creating recurring demand across APIs, subscriptions, and infrastructure. Many companies are rushing with their empty buckets to capture water from this inference (revenue) waterfall. In fact, I would argue that the industry is in a state of “either you die trying to get to the frontier, or you live long enough to serve inference.”

Frequently asked questions

Answers drawn from the report and the sources below.

Sources and dates

2026 report snapshot. Preview revised 2026-10-07. Individual data periods and source checks are listed below. This is not a claim that every source was updated on that date.

  1. Epoch AI: The plunging price of thought2026-09-22. Primary source checked 2026-10-07.
  2. State of AI Report 2026, slide 95: The top 1% of firms spend about 580x the median per employee on AI2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  3. State of AI Report 2026, slide 101: AI performance still varies widely across financial work2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  4. State of AI Report 2026, slide 103: Heavy AI spenders hire faster…except for scientists2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  5. State of AI Report 2026, slide 104: Early AI labor studies point to risks for junior workers2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  6. State of AI Report 2026, slide 106: The AI build-out is adding jobs even as some office roles shrink2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  7. State of AI Report 2026, slide 110: Reasoning makes token price a poor proxy for the cost of an answer2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  8. State of AI Report 2026, slide 113: Vertical AI companies post-train open models past the frontier in their own domain2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  9. State of AI Report 2026, slide 114: Production feedback guides improvements across the AI stack2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  10. State of AI Report 2026, slide 115: Agents now build and fix customer service agents, and the customer's staff approve2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  11. State of AI Report 2026, slide 116: What training data is valuable? Execution traces and in-domain records2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  12. State of AI Report 2026, slide 117: Teaching AI is now generating billions of dollars in revenue2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  13. State of AI Report 2026, slide 152: Private AI valuations have risen fast, very fast2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  14. State of AI Report 2026, slide 156: Mega rounds continue to eat the lion’s share of private AI company raises2026 report snapshot. Read against the report PDF on 2026-10-07. Study-specific limits retained.
  15. State of AI Report 20262026 report snapshot. Report PDF. See individual slide references for the expanded analysis.

Cite this page

Benaich, Nathan. “Is AI inference getting cheaper?.” State of AI Report 2026. Published 2026-10-08.