Model Comparison
OpenAI's GPT-4.1 mini costs less per intelligence point, even though Anthropic's Claude 4 Sonnet (Non-reasoning) scores higher.
Data last updated March 4, 2026
GPT-4.1 mini delivers more intelligence per dollar, while Claude 4 Sonnet (Non-reasoning) leads on raw benchmark scores. Claude 4 Sonnet (Non-reasoning) costs $0.03 per request vs $0.0036 for GPT-4.1 mini (at 5K input / 1K output tokens). GPT-4.1 mini scores proportionally higher on mathematical reasoning (AIME: 0.43), while Claude 4 Sonnet (Non-reasoning)'s scores skew toward general knowledge (MMLU-Pro: 0.84). The question is whether Claude 4 Sonnet (Non-reasoning)'s higher scores justify the 8x price premium.
| Metric | Claude 4 Sonnet (Non-reasoning) | GPT-4.1 mini |
|---|---|---|
| Intelligence IndexComposite score from MMLU-Pro, GPQA, and AIME. Higher is better. | 33.0 | 22.9 |
| MMLU-ProGeneral knowledge and reasoning. Higher is better. | 0.8 | 0.8 |
| GPQAGraduate-level science questions. Higher is better. | 0.7 | 0.7 |
| AIMEMathematical problem solving. Higher is better. | 0.4 | 0.4 |
| Output tokens/secTokens generated per second. Higher means faster responses. | 48.5 | 70.6 |
| Time to first tokenSeconds until first token. Lower is better. | 1.03s | 0.48s |
| Context windowMax tokens per request. Larger handles more text. | 200,000 | 1,047,576 |
List prices as published by the provider. Not adjusted for token efficiency.
| Metric | Claude 4 Sonnet (Non-reasoning) | GPT-4.1 mini |
|---|---|---|
| Input price / 1M tokens | $3.00 | $0.40 |
| Output price / 1M tokens | $15.00 | $1.60 |
| Cache hit price / 1M tokens | $0.30 | $0.10 |
Cost per IQ point based on a typical request of 5,000 input and 1,000 output tokens.
Cheaper (list price)
GPT-4.1 mini
Higher Benchmarks
Claude 4 Sonnet (Non-reasoning)
Better Value ($/IQ point)
GPT-4.1 mini
Claude 4 Sonnet (Non-reasoning)
$0.0009 / IQ point
GPT-4.1 mini
$0.0002 / IQ point
GPT-4.1 mini is dramatically cheaper — 8x less per request than Claude 4 Sonnet (Non-reasoning). GPT-4.1 mini is cheaper on both input ($0.4/M vs $3.0/M) and output ($1.6/M vs $15.0/M). At a fraction of the cost, GPT-4.1 mini saves significantly in production workloads. This comparison assumes a typical request of 5,000 input and 1,000 output tokens (5:1 ratio). Actual ratios vary by workload — chat and completion tasks typically run 2:1, code review around 3:1, document analysis and summarization 10:1 to 50:1, and embedding workloads are pure input with no output tokens.
Claude 4 Sonnet (Non-reasoning) scores higher overall (33.0 vs 22.9). Claude 4 Sonnet (Non-reasoning) leads on MMLU-Pro (0.84 vs 0.78), with both within 5% on GPQA and AIME. GPT-4.1 mini scores proportionally higher on AIME (mathematical reasoning) relative to its MMLU-Pro, while Claude 4 Sonnet (Non-reasoning)'s scores are more weighted toward general knowledge. Claude 4 Sonnet (Non-reasoning)'s higher MMLU-Pro score suggests better performance on general-purpose tasks.
GPT-4.1 mini is 46% faster at 70.6 tokens per second compared to Claude 4 Sonnet (Non-reasoning) at 48.5 tokens per second. GPT-4.1 mini also starts generating sooner at 0.48s vs 1.03s time to first token. The speed difference matters for chatbots but is less relevant in batch processing.
GPT-4.1 mini has a much larger context window — 1,047,576 tokens vs Claude 4 Sonnet (Non-reasoning) at 200,000 tokens. That's roughly 1,396 vs 266 pages of text. GPT-4.1 mini's window can handle entire codebases or book-length documents; Claude 4 Sonnet (Non-reasoning) works better for shorter inputs.
GPT-4.1 mini offers dramatically better value — $0.0002 per intelligence point vs Claude 4 Sonnet (Non-reasoning) at $0.0009. GPT-4.1 mini is cheaper, which offsets Claude 4 Sonnet (Non-reasoning)'s higher benchmark scores to deliver more value per dollar. If raw benchmark scores matter less than cost for your use case, GPT-4.1 mini is the efficient choice.
With prompt caching, GPT-4.1 mini is dramatically cheaper — 8x less per request than Claude 4 Sonnet (Non-reasoning). Caching saves 45% on Claude 4 Sonnet (Non-reasoning) and 42% on GPT-4.1 mini compared to standard input prices. Both models benefit from caching at similar rates, so the uncached price comparison holds.
Pricing verified against official vendor documentation. Updated daily. See our methodology.
Related Comparisons
Create an account, install the SDK, and see your first margin data in minutes.
See My Margin DataNo credit card required