What a fixed level of AI costs
Pick a model you know. The chart follows the cheapest model that matches it, from its launch to today. A model that was the best you could buy becomes something a small, cheap model can do, often within months.
Measured as: Price per million tokens for the cheapest model matching GPT-4's score on PhD-level science questions (GPQA Diamond).
Counting every token used per question, Epoch AI's figures give the same picture: from $30,720 per million questions (GPT-4, March 2023) to $65 (gpt-oss-20b, August 2025), about 470 times less. Nothing cheaper has appeared since, which may be a floor. Data →
Price of GPT-4-level performance
Cheapest non-reasoning model scoring at least GPT-4's level (≥33% on GPQA Diamond), USD per million tokens (3:1 input:output blend). Log scale.
What this measures: What it costs to use an AI at least as capable as the original GPT-4 (2023). A million tokens is roughly 750,000 words of input and output.
2023–2024 points are Epoch AI's. Points from late 2024 on are AI Tracker's extension using archived OpenRouter price lists and published GPQA Diamond scores (one provider route, not always the cheapest). The price of this capability level has flattened at roughly $0.05–0.13 since late 2024; the steep declines are now at higher capability levels (see the ARC-AGI cost chart). Per-token price is not cost per task.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 14 Mar 2023 | GPT-4 | $37.50 |
| 6 Nov 2023 | GPT-4 Turbo | $15 |
| 26 Feb 2024 | Mistral Large | $6 |
| 4 Mar 2024 | Claude 3 Haiku | $0.50 |
| 4 Nov 2024 | Gemini 1.5 Flash | $0.13 |
| 13 Dec 2024 | Phi 4 | $0.12 |
| 3 May 2025 | Gemma 3 12B | $0.063 |
| 1 Nov 2025 | Gemma 3 12B | $0.048 |
| 4 May 2026 | Mistral Small 3 | $0.058 |
| 1 Nov 2026 | Mistral Small 3 | $0.058 |
- LLM inference prices have fallen rapidly but unequally across tasks (Epoch AI)
- Lowest inference prices at fixed performance (CSV) (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)
Measured as: Not a list price: the full cost of answering a million PhD-level science questions on one test, GPQA Diamond, at o3's score (82%), counting the hidden thinking tokens. For o3 that was about 30 cents per question.
Chinese open-weight models were the cheapest way to reach this level for much of the period: DeepSeek-V3.2 (December 2025), then two Qwen 3.5 models (February 2026) and MiniMax-M3 (June 2026). On hard math (FrontierMath Tiers 1–3), matching o3 went from about $1.06 per problem to about $0.002 over the same time. Data →
Price of o3-level performance
Not a list price: the cost of answering a million PhD-level science questions (GPQA Diamond) at o3's score, using the cheapest model that reaches it. Counts thinking tokens. o3 cost about 30 cents per question. Log scale.
What this measures: What it costs to get answers as good as o3's (April 2025) on hard science questions. Unlike price per token, this counts every token a model actually uses, including the hidden 'thinking' tokens of reasoning models.
From Epoch AI's 'The plunging price of thought' (September 2026) and its data files; values are rounded to three figures. Data covers models released up to August 2026. Epoch's own chart dates the o3 run 31 January 2025 and shows an earlier Gemini 2.5 Pro point (25 March 2025, about $89,100). Here o3 is dated at its public release on 16 April 2025, the model Epoch ran (o3-2025-04-16), so that earlier point is left out. Epoch's token prices for GPT-5.6 Luna match OpenAI's batch rate, which is half its standard price, so the last point would be about twice as high at standard rates.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 16 Apr 2025 | o3 (high) | $299,000 |
| 5 Jun 2025 | Gemini 2.5 Pro (Jun 2025) | $12,900 |
| 1 Dec 2025 | DeepSeek-V3.2 | $8,820 |
| 13 Feb 2026 | Qwen3.5 397B-A17B | $7,100 |
| 25 Feb 2026 | Qwen 3.5 Flash (hosted 35B-A3B) | $2,180 |
| 1 Jun 2026 | MiniMax-M3 | $1,210 |
| 9 Jul 2026 | GPT-5.6 Luna (low) | $412 |
- The plunging price of thought (Epoch AI)
- inference-cost data and code (David Roodman / Epoch AI)
Measured as: List price per million tokens for the cheapest model at or above Claude Opus 4.6 on the Epoch Capabilities Index (156.3).
Chinese models caught up within months. On Epoch's science-question runs, the open-weight DeepSeek-V4-Pro (April 2026) matched Opus 4.6's score at about 15 times less per question, and MiniMax-M3 (June 2026) at about 21 times less. On the index, DeepSeek V4.1 Flash (October 2026) now scores above Opus 4.6 at $0.26 per million tokens, 38 times below Opus 4.6's launch price. Kimi K3, Qwen 3.8 Max, GLM-5.3 and Kimi K3.1 Pro match it too. Data →
Price of Claude Opus 4.6-level performance
Cheapest model scoring at least Claude Opus 4.6's 156.3 on the Epoch Capabilities Index, USD per million tokens (3:1 input:output blend). Log scale.
What this measures: What it costs to use an AI at least as capable as Claude Opus 4.6, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
AI Tracker's calculation: Epoch AI's index scores combined with each developer's own list price on 10 November 2026, blended 3 parts input to 1 part output. Each point is dated at the model's release. Index scores carry a margin of error of about 2 to 3 points, so a model just above the line and one just below it are not really different. Price per token is not cost per task: models use very different numbers of tokens.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 5 Feb 2026 | Claude Opus 4.6 | $10 |
| 5 Mar 2026 | GPT-5.4 | $5.63 |
| 30 Jun 2026 | Claude Sonnet 5 | $4 |
| 9 Jul 2026 | GPT-5.6 Luna | $0.45 |
| 22 Sept 2026 | GPT-6 Luna | $0.20 |
- Epoch Capabilities Index scores (CSV) (Epoch AI)
- API pricing (OpenAI)
- Pricing (Anthropic)
- Models API (list prices) (OpenRouter)
Measured as: List price per million tokens for the cheapest model at or above Claude Opus 4.8 on the Epoch Capabilities Index (159.9).
One Chinese model matches it on the index so far: Kimi K3.1 Pro (October 2026, 162.0) at $4 per million tokens, 2.5 times below Opus 4.8's launch price. DeepSeek V4.1 Flash sits just under the line (159.3 against 159.9, inside the margin of error) at $0.26. On hard math (FrontierMath Tiers 1–3), no Chinese model in Epoch's data reaches Opus 4.8's 80% yet. Data →
Price of Claude Opus 4.8-level performance
Cheapest model scoring at least Claude Opus 4.8's 159.9 on the Epoch Capabilities Index, USD per million tokens (3:1 input:output blend). Log scale.
What this measures: What it costs to use an AI at least as capable as Claude Opus 4.8, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
AI Tracker's calculation: Epoch AI's index scores combined with each developer's own list price on 10 November 2026, blended 3 parts input to 1 part output. Each point is dated at the model's release. Index scores carry a margin of error of about 2 to 3 points, so a model just above the line and one just below it are not really different. Price per token is not cost per task: models use very different numbers of tokens.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 28 May 2026 | Claude Opus 4.8 | $10 |
| 9 Jul 2026 | GPT-5.6 Terra | $4.50 |
| 22 Sept 2026 | GPT-6 Sol | $4 |
| 7 Oct 2026 | Claude Haiku 5.5 | $0.20 |
- Epoch Capabilities Index scores (CSV) (Epoch AI)
- API pricing (OpenAI)
- Pricing (Anthropic)
- Models API (list prices) (OpenRouter)
Measured as: List price per million tokens for the cheapest model at or above GPT-5.6 Sol on the Epoch Capabilities Index (163.0).
No Chinese model is rated at this level on the index yet. Data →
Price of GPT-5.6 Sol-level performance
Cheapest model scoring at least GPT-5.6 Sol's 163.0 on the Epoch Capabilities Index, USD per million tokens (3:1 input:output blend). Log scale.
What this measures: What it costs to use an AI at least as capable as GPT-5.6 Sol, judged by Epoch AI's index of more than 50 tests. A million tokens is roughly 750,000 words of input and output.
The launch point uses GPT-5.6 Sol's original price of $5 in and $30 out per million tokens. OpenAI lists a promotional $4 and $20 until 30 November 2026.
AI Tracker's calculation: Epoch AI's index scores combined with each developer's own list price on 10 November 2026, blended 3 parts input to 1 part output. Each point is dated at the model's release. Index scores carry a margin of error of about 2 to 3 points, so a model just above the line and one just below it are not really different. Price per token is not cost per task: models use very different numbers of tokens.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 9 Jul 2026 | GPT-5.6 Sol | $11.25 |
| 24 Jul 2026 | Claude Opus 5 | $10 |
| 22 Sept 2026 | GPT-6 Sol | $4 |
| 26 Oct 2026 | GPT-6.1 Terra | $2 |
- Epoch Capabilities Index scores (CSV) (Epoch AI)
- API pricing (OpenAI)
- Pricing (Anthropic)
- Models API (list prices) (OpenRouter)
Every level side by side
The same numbers in one table. Tap a model to see its chart.
| As capable as | Price at launch | Cheapest match now | Change |
|---|---|---|---|
| GPT-4released Mar 2023 | $37.50per million tokens | $0.048Gemma 3 12B, Nov 2025 | 789× cheaperin 2 years 7 months |
| o3released Apr 2025 | $299,000per million GPQA Diamond questions | $412GPT-5.6 Luna (low), Jul 2026 | 726× cheaperin 14 months |
| Claude Opus 4.6released Feb 2026 | $10per million tokens | $0.20GPT-6 Luna, Sept 2026 | 50× cheaperin 7 months |
| Claude Opus 4.8released May 2026 | $10per million tokens | $0.20Claude Haiku 5.5, Oct 2026 | 50× cheaperin 4 months |
| GPT-5.6 Solreleased Jul 2026 | $11.25per million tokens | $2GPT-6.1 Terra, Oct 2026 | 5.6× cheaperin 3 months |
How to read this
Two yardsticks. For GPT-4 and o3, "as capable" means the same score on PhD-level science questions (GPQA Diamond), from Epoch AI's price studies. That test is nearly maxed out by 2026 models, so for newer reference models "as capable" means the same score or better on the Epoch Capabilities Index, which combines more than 50 tests.
Two kinds of price. Most charts here show the list price per million tokens. The o3 chart is different: it shows what it costs to answer a million questions of one test, GPQA Diamond, which also counts the hidden "thinking" tokens. That is why its numbers look so large: $299,000 per million questions is about 30 cents per question. A cheap model that thinks for a long time can cost more per task than its token price suggests.
Close calls. Index scores carry a margin of error of about 2 to 3 points. A model just over the line and one just under it are not really different, so read "matches" as "about as capable on these tests".
What it leaves out. Benchmarks miss things people care about in daily use, such as taste in code, reliability over long sessions and how a model handles vague requests. A model can match another on the index and still feel worse to work with. The index-based charts are AI Tracker's own calculation from Epoch AI's scores and each developer's price list; Epoch AI has not published them in this form.
METR has not published a price series at a fixed level of capability. Its closest work,Expenditure Horizon(July 2026), compares what AI agents and people achieve per dollar on one task. See alsoEpoch AI's finding that a fixed level gets about 13× cheaper a year.