The best AI keeps getting smarter and stays expensive. Yesterday’s best gets cheap fast.
Since 2023, in three numbers and one chart.
The best model stays expensive. Its level gets cheap.
Top line: launch price of each best model, which swings between about $3 and $68 with no steady fall. Other lines: the cheapest model that matches an older one, from its launch to today. USD per million tokens, log scale.
- Best model at launch
- As capable as GPT-4
- As capable as Claude Opus 4.6
- As capable as Claude Opus 4.8
List prices, 3 parts input to 1 part output. “Matches” means the same test score, within its margin of error (AI Tracker’s calculation from Epoch AI’s scores). Price per token is not cost per task: top models also use more tokens, so the real gap is larger.
Show data table and sources
| Series | Date | Label | Value |
|---|---|---|---|
| Best model at launch | 14 Mar 2023 | GPT-4 | $37.50 |
| Best model at launch | 6 Nov 2023 | GPT-4 Turbo | $15 |
| Best model at launch | 4 Mar 2024 | Claude 3 Opus | $30 |
| Best model at launch | 13 May 2024 | GPT-4o | $7.50 |
| Best model at launch | 20 Jun 2024 | Claude 3.5 Sonnet | $6 |
| Best model at launch | 12 Sept 2024 | o1-mini | $5.25 |
| Best model at launch | 17 Dec 2024 | o1 | $26.25 |
| Best model at launch | 4 Apr 2025 | Gemini 2.5 Pro | $3.44 |
| Best model at launch | 16 Apr 2025 | o3 | $17.50 |
| Best model at launch | 10 Jun 2025 | o3-pro | $35 |
| Best model at launch | 7 Aug 2025 | GPT-5 | $3.44 |
| Best model at launch | 6 Oct 2025 | GPT-5 Pro | $41.25 |
| Best model at launch | 18 Nov 2025 | Gemini 3 Pro | $4.50 |
| Best model at launch | 11 Dec 2025 | GPT-5.2 Pro | $57.75 |
| Best model at launch | 5 Mar 2026 | GPT-5.4 Pro | $67.50 |
| Best model at launch | 24 Apr 2026 | GPT-5.5 Pro | $67.50 |
| Best model at launch | 9 Jun 2026 | Claude Fable 5 | $20 |
| Best model at launch | 24 Jul 2026 | Claude Opus 5 | $10 |
| Best model at launch | 1 Sept 2026 | Claude Fable 5.1 | $20 |
| Best model at launch | 3 Sept 2026 | GPT-6 Astra | $20 |
| Best model at launch | 9 Nov 2026 | Claude Fable 5.2 | $16 |
| As capable as GPT-4 | 14 Mar 2023 | GPT-4 | $37.50 |
| As capable as GPT-4 | 6 Nov 2023 | GPT-4 Turbo | $15 |
| As capable as GPT-4 | 26 Feb 2024 | Mistral Large | $6 |
| As capable as GPT-4 | 4 Mar 2024 | Claude 3 Haiku | $0.50 |
| As capable as GPT-4 | 4 Nov 2024 | Gemini 1.5 Flash | $0.13 |
| As capable as GPT-4 | 13 Dec 2024 | Phi 4 | $0.12 |
| As capable as GPT-4 | 3 May 2025 | Gemma 3 12B | $0.063 |
| As capable as GPT-4 | 1 Nov 2025 | Gemma 3 12B | $0.048 |
| As capable as GPT-4 | 4 May 2026 | Mistral Small 3 | $0.058 |
| As capable as GPT-4 | 1 Nov 2026 | Mistral Small 3 | $0.058 |
| As capable as Claude Opus 4.6 | 5 Feb 2026 | Claude Opus 4.6 | $10 |
| As capable as Claude Opus 4.6 | 5 Mar 2026 | GPT-5.4 | $5.63 |
| As capable as Claude Opus 4.6 | 30 Jun 2026 | Claude Sonnet 5 | $4 |
| As capable as Claude Opus 4.6 | 9 Jul 2026 | GPT-5.6 Luna | $0.45 |
| As capable as Claude Opus 4.6 | 22 Sept 2026 | GPT-6 Luna | $0.20 |
| As capable as Claude Opus 4.8 | 28 May 2026 | Claude Opus 4.8 | $10 |
| As capable as Claude Opus 4.8 | 9 Jul 2026 | GPT-5.6 Terra | $4.50 |
| As capable as Claude Opus 4.8 | 22 Sept 2026 | GPT-6 Sol | $4 |
| As capable as Claude Opus 4.8 | 7 Oct 2026 | Claude Haiku 5.5 | $0.20 |
- Epoch Capabilities Index (Epoch AI)
- API pricing (OpenAI)
- Pricing (Anthropic)
- Lowest inference prices at fixed performance (CSV) (Epoch AI)
- Introducing the next generation of Claude (Anthropic)
- Gemini 3 for developers (Google)
- o3-pro model page (OpenAI)
- LLM inference prices have fallen rapidly but unequally across tasks (Epoch AI)
- OpenRouter models API (archived 2025-11-01) (OpenRouter via Internet Archive)
- OpenRouter models API (archived 2026-05-04) (OpenRouter via Internet Archive)
- Gemma 3 Technical Report (GPQA Diamond scores) (Google DeepMind)
- Epoch Capabilities Index scores (CSV) (Epoch AI)
- Models API (list prices) (OpenRouter)
How fast: a fixed level gets about 13× cheaper a year on average: about 75× a year while it is new, about 4.7× a year two years on (Epoch AI). Pick a model and see its own chart →
Where the trends lead by 2031
Top: how capable the best model is. Bottom: what one fixed level of capability costs, not the best model. Dashed lines show where each trend leads by the end of 2031 if it slows, holds its pace or speeds up.
AI capability (Epoch Capabilities Index)
One score built from more than 50 tests, so models stay comparable even as single tests max out. GPT-4 (2023) scored 122, GPT-5 scored 150 and Claude Fable 5.2 scores 174. At 134 a model gets about halfway on PhD-level science questions; at 151, halfway on FrontierMath's hard maths.
Price of one fixed level of capability (o3's, April 2025)
Not the price of the best model: the cost of matching one fixed level. Here, answering a million PhD-level science questions (GPQA Diamond) at o3's score (82%), using the cheapest model that reaches it and counting reasoning tokens. Lighter line: GPT-4's level (36%), from 2023.
- Measured
- Extrapolation (slower, same pace, faster)
- Range between slower and faster
Even the slower lines add up by 2031
If capability grows at its pace from before reasoning models and prices fall at the slow end of Epoch AI’s range, by the end of 2031:
Index values above about 180 are harder than every test the index contains, so a score of 215 has no concrete meaning yet; read it as “well beyond today”.
Dashed lines extend past rates. They are not forecasts or probability bounds: trends can speed up, slow down or stop.
No line lets AI speed up its own progress, sometimes called recursive self-improvement. The faster line adds the 1.9× speed-up METR measured at Anthropic once, as a fixed boost. If that speed-up keeps growing as AI gets better, progress could outrun all three lines. How METR measured the speed-up →
Claude Fable 5.2's score is from its launch day and rests on 10 tests so far (range 170–179), so expect it to move. The projections start from it.
Above about 180 points the index is harder than every benchmark it currently contains, so projected values that high have no concrete meaning yet.
What each dashed line assumes
AI capability (Epoch Capabilities Index)
- Slower: Back to the pace before reasoning models: about 8 points a year (Epoch AI's estimate for GPT-4 to GPT-4.5).
- Same pace: The 2023–2026 trend continues: about 15.5 points a year (a straight-line fit to every record since GPT-4; Epoch AI estimates 16.6 since o1-mini).
- Faster: Acceleration spreads: the mid rate × 1.9, the speed-up METR measured at Anthropic in November 2026. This is an assumption; public data doesn't yet show an industry-wide speed-up.
Price of one fixed level of capability (o3's, April 2025)
- Slower drop: Declines slow down: about 4.7× cheaper a year, the rate Epoch AI sees two years after a capability level is first reached.
- Same pace: The average rate continues: about 13× cheaper a year (Epoch AI's 'price of thought' estimate).
- Faster drop: Fast end of Epoch AI's range: about 32× cheaper a year.
Projected values by year
| Measure | Scenario | End 2027 | End 2028 | End 2029 | End 2030 | End 2031 |
|---|---|---|---|---|---|---|
| Capability (points) | Slower | 183.3 | 191.3 | 199.3 | 207.3 | 215.3 |
| Capability (points) | Same pace | 191.9 | 207.4 | 222.9 | 238.4 | 253.9 |
| Capability (points) | Faster | 207.8 | 237.3 | 266.8 | 296.2 | 325.6 |
| Price (USD per million questions) | Slower drop | $45 | $9.7 | $2.1 | $0.44 | $0.095 |
| Price (USD per million questions) | Same pace | $11 | $0.86 | $0.068 | $0.0054 | $0.00042 |
| Price (USD per million questions) | Faster drop | $2.9 | $0.089 | $0.0028 | $0.000087 | $0.0000027 |
- Capability values are refitted by Epoch AI as new results arrive, typically by about a point.
- Price data covers models released up to July 2026. The GPT-4-level price has had no new record since August 2025, which may be a floor; that fits the low scenario better than the others.
- Prices are per question including reasoning tokens, from Epoch AI's runs. Older models are dated by their own release dates.
How to read this page
Capability uses Epoch AI's Capabilities Index, which combines results from more than 50 tests into one score, so it keeps working as single tests max out. The best model's score rose from 122 (GPT-4, March 2023) to 174 (Claude Fable 5.2, November 2026).
Price is not the price of the best model. The best model's launch price has gone up as often as down since 2023, and the top "Pro" tiers cost the most. What falls is the price of matching a model that is already out: smaller, cheaper models catch up with it.
Two ways to count. The first chart uses list prices per million tokens. The second figure's price panel counts the full cost of answering a million PhD-level science questions at o3's score, including the hidden "thinking" tokens: about $299,000 in April 2025 and about $412 in July 2026 (30 cents per question down to 0.04 cents), 725 times less.
Definitions: capabilities index,extrapolation,log scale, tokens.
What could change the trends
Faster: AI that helps build the next AI. METR measured progress at Anthropic running about 1.9× faster than before (November 2026).
Slower: limits on chips, electricity or data; regulation; or simply diminishing returns.
Price floors: the price of GPT-4-level answers hasn't set a new record since August 2025, which may be close to a floor.