The cost of intelligence
The same capability keeps getting cheaper. Each line is the cheapest first-party API price for a model at or above a GPQA Diamond score.
- GPQA ≥ 35%GPT-4 level286×cheaper than when first reachedHalves every 4.4 months
$37.50 on 2023-03-14 → $0.13 now (Gemini 1.5 Flash 001) - GPQA ≥ 50%GPT-4o · Claude 3.5 Sonnet level44×cheaper than when first reachedHalves every 4.9 months
$6.00 on 2024-06-20 → $0.14 now (GPT-5 Nano 2025-08-07) - GPQA ≥ 75%o1 level135×cheaper than when first reachedHalves every 5.9 months
$26.25 on 2024-12-17 → $0.19 now (GLM-5.3 Flash) - GPQA ≥ 90%GPT-5.2 · Gemini 3 Pro level23×cheaper than when first reachedHalves every 1.8 months
$4.50 on 2025-11-18 → $0.19 now (GLM-5.3 Flash)
USD per million tokens, blended 3:1 input to output
Hollow dots are launch prices still awaiting a check against an archived vendor page.
Which model brought each step
GPQA ≥ 35%: 7 steps
| From | Model | GPQA | Blended / M tokens | List price |
|---|---|---|---|---|
| 2023-03-14 | GPT-4 0314 · OpenAI | 36% | $37.50 | USD 30 / 60 |
| 2023-11-06 | GPT-4 1106 Preview · OpenAI | 42% | $15.00 | USD 10 / 30 |
| 2024-02-26 | Mistral Large 2402 · Mistral AI | 39% | $12.00 | USD 8 / 24 |
| 2024-03-04 | Claude 3 Sonnet 20240229 · Anthropic | 41% | $6.00 | USD 3 / 15 |
| 2024-03-13 | Claude 3 Haiku 20240307 · Anthropic | 36% | $0.50 | USD 0.25 / 1.25 |
| 2024-07-18 | GPT-4o Mini 2024-07-18 · OpenAI | 38% | $0.26 | USD 0.15 / 0.6 |
| 2024-08-12 | Gemini 1.5 Flash 001 · Google | 40% | $0.13 | USD 0.075 / 0.3 |
GPQA ≥ 50%: 6 steps
| From | Model | GPQA | Blended / M tokens | List price |
|---|---|---|---|---|
| 2024-06-20 | Claude 3.5 Sonnet 20240620 · Anthropic | 54% | $6.00 | USD 3 / 15 |
| 2024-09-12 | o1 mini-2024-09-12 · OpenAI | 62% | $5.25 | USD 3 / 12 |
| 2024-10-01 | Gemini 1.5 Pro 002 · Google | 57% | $2.19 | USD 1.25 / 5 |
| 2025-01-20 | DeepSeek R1 · DeepSeek | 69% | $0.96 | USD 0.55 / 2.19 |
| 2025-02-05 | Gemini 2.0 Flash 001 · Google | 64% | $0.18 | USD 0.1 / 0.4 |
| 2025-08-07 | GPT-5 Nano 2025-08-07 · OpenAI | 69% | $0.14 | USD 0.05 / 0.4 |
GPQA ≥ 75%: 5 steps
| From | Model | GPQA | Blended / M tokens | List price |
|---|---|---|---|---|
| 2024-12-17 | o1 2024-12-17 · OpenAI | 77% | $26.25 | USD 15 / 60 |
| 2025-01-31 | o3 mini-2025-01-31 · OpenAI | 77% | $1.93 | USD 1.1 / 4.4 |
| 2025-04-09 | Grok 3 mini Beta · xAI | 76% | $0.35 | USD 0.3 / 0.5 |
| 2025-12-01 | DeepSeek Reasoner · DeepSeek | 83% | $0.32 | USD 0.28 / 0.42 |
| 2026-08-26 | GLM-5.3 Flash · Z.ai | 90% | $0.19 | CNY 0.8 / 2.8 |
GPQA ≥ 90%: 4 steps
| From | Model | GPQA | Blended / M tokens | List price |
|---|---|---|---|---|
| 2025-11-18 | Gemini 3 Pro Preview · Google | 93% | $4.50 | USD 2 / 12 |
| 2026-04-24 | DeepSeek V4 Pro Preview · DeepSeek | 91% | $0.54 | USD 0.435 / 0.87 |
| 2026-07-09 | GPT-5.6 Luna · OpenAI | 92% | $0.45 | USD 0.2 / 1.2 |
| 2026-08-26 | GLM-5.3 Flash · Z.ai | 90% | $0.19 | CNY 0.8 / 2.8 |
How it is computed
Capability is GPQA Diamond accuracy as measured by Epoch AI, taking each model's best configuration. Price is the first-party API list price per million tokens: standard tier, cache-miss input, shortest context, no launch promotions or batch discounts. Input and output are blended 3:1.
On every date a model launched or a price changed, each line takes the cheapest model at or above its threshold. The halving time is a least-squares fit of log price against months, with the line read on the first of every month.
What it does not say
- Prices are per token, not per task: reasoning models spend far more tokens on the same job.
- Each model takes the score of its best configuration, which may also be the one that spends the most tokens.
- Retired models stay in: the line answers what could be bought at the time.
- Only first-party list prices count; open models served by third parties (Llama, for example) are left out.
- GPQA is saturating: above 90% it separates models less well, and a harder benchmark will take over the top tier.