Reasoning models
296 models that reason step by step before they answer, for math, code and multi-step problems.
/
| Grok 4.20 Beta 0309 ReasoningxAIPrompt cachingJSON+3 | $6/1Mvia AIHubMix | 1M | $2/1M | $6/1M | — | — | |
| Grok 4.20 ReasoningxAIPrompt cachingJSON+3 | $2.50/1Mvia Vercel AI Gateway | 1M | $1.25/1M | $2.50/1M | — | — | |
| Grok-4-FastxAIPrompt cachingJSON+2 | $0.50/1Mvia Azure | 2M | $0.20/1M | $0.50/1M | 1407 | — | |
| Hermes 4 405BNous ResearchLlama 3JSON+1 | $3/1Mvia OpenRouter | 131K | $1/1M | $3/1M | — | — | |
| Hermes-4-70BNous ResearchLlama 3JSON+2 | $0.40/1Mvia Requesty | 128K | $0.13/1M | $0.40/1M | — | — | |
 Hy3TencentApache 2.0Prompt caching+3 | $0.33/1Mvia OpenRouter | 262K | $0.083/1M | $0.33/1M | 1506 | — | |
| Hy3 PreviewTencentOpen weightsPrompt caching+2 | $0.60/1Mvia OpenRouter | 262K | $0.18/1M | $0.60/1M | 1401 | — | |
| Hy4 previewTencentApache 2.0Prompt caching+3 | $2.25/1Mvia OpenRouter | 1.05M | $0.751/1M | $2.25/1M | 1631 | — | |
 Inkling (256K)Thinking MachinesPrompt cachingReasoning+2 | $1.50/1Mvia Venice AI | 256K | $0.563/1M | $1.50/1M | 1442 | — | |
| Inkling SmallThinking MachinesPrompt cachingReasoning+2 | $1.20/1Mvia Deep Infra | 1.05M | $0.45/1M | $1.20/1M | 1413 | — | |
| Kimi K2 ThinkingMoonshot AIOpen weights · with termsPrompt caching+3 | $1.68/1Mvia APIMart | 262K | $0.40/1M | $1.68/1M | — | — | |
| Kimi K2.7 Code HighspeedMoonshot AIPrompt cachingJSON+3 | $6.17/1Mvia APIMart | 262K | $1.49/1M | $6.17/1M | — | — | |
| Kimi K3.1 Pro HighspeedMoonshot AIPrompt cachingJSON+3 | $6.13/1Mvia Kenari | 1.05M | $1.81/1M | $9.44/1M | — | — | |
 Laguna XS 2.1PoolsideReasoningTool calling | $0.12/1Mvia OpenRouter | 262K | $0.06/1M | $0.12/1M | — | — | |
 Ling 3.0 Flash FinInclusionaiMITPrompt caching+1 | $0.18/1Mvia Hugging Face | 256K | $0.06/1M | $0.18/1M | — | — | |
| Ling 3.0 Flash SanteInclusionaiMITPrompt caching+2 | $0.22/1Mvia Novita AI | 262K | $0.075/1M | $0.22/1M | — | — | |
| Ling 3.0 Flash VLInclusionaiMITPrompt caching+4 | $0.18/1Mvia Deep Infra | 131K | $0.06/1M | $0.18/1M | — | — | |
 Llama-3.2-90B-Vision-InstructMetaLlama 3.2JSON+1 | $0.40/1Mvia Requesty | 131K | $0.35/1M | $0.40/1M | 1000 | — | |
 Macaron V1 TallMindlabPrompt cachingReasoning | $1.43/1Mvia Novita AI | 256K | $0.247/1M | $1.43/1M | — | — | |
| Macaron V1 VentiMindlabPrompt cachingReasoning | $2.48/1Mvia Novita AI | 1M | $0.825/1M | $2.48/1M | — | — | |
| Mercury 2.5 PreviewInceptionPrompt cachingReasoning+1 | $0.75/1Mvia AIHubMix | 260K | $0.20/1M | $0.75/1M | — | — | |
 MiMo-V2.6-Pro-UltraSpeedXiaomiPrompt cachingJSON+3 | $6.96/1Mvia CometAPI | 1.05M | $3.48/1M | $6.96/1M | — | — | |
 MiniMax M1MiniMaxReasoningTool calling | $2.20/1Mvia OpenRouter | 1M | $0.55/1M | $2.20/1M | 1343 | — | |
| MiniMax M2MiniMaxOpen weights · with termsPrompt caching+3 | $1/1Mvia DevPass (LLM Gateway) | 205K | $0.20/1M | $1/1M | 1340 | — | |
| MiniMax M2.1 Lightning (MiniMax)MiniMaxPrompt cachingJSON+2 | $0.48/1Mvia DevPass (LLM Gateway) | 205K | $0.12/1M | $0.48/1M | — | — | |
| MiniMax M2.5 highspeedMiniMaxPrompt cachingJSON+2 | $1.15/1Mvia AIHubMix | 205K | $0.288/1M | $1.15/1M | — | — | |
| MiniMax-M2.7MiniMaxOpen weightsPrompt caching+3 | $0.60/1Mvia EmpirioLabs AI | 205K | $0.15/1M | $0.60/1M | 1405 | — | |
 Mistral Medium (latest)MistralJSONReasoning+1 | $2.20/1Mvia Requesty | 131K | $0.44/1M | $2.20/1M | — | — | |
| Mistral Small 3.1 24BMistralApache 2.0JSON+2 | $0.33/1Mvia Requesty | 33K | $0.11/1M | $0.33/1M | — | — | |
| Muse Glimmer 30BMetaPrompt cachingJSON+3 | $0.80/1Mvia EmpirioLabs AI | 131K | $0.20/1M | $0.80/1M | — | — | |
| Muse Spark 1.1MetaPrompt cachingJSON+3 | $4.25/1Mvia EmpirioLabs AI | 1.05M | $1.25/1M | $4.25/1M | 1543 | — | |
| Muse Spark 1.2MetaPrompt cachingJSON+3 | $4.25/1Mvia EmpirioLabs AI | 1.05M | $1.25/1M | $4.25/1M | 1534 | — | |
 Nemotron 3.5 Super VL 120B A12BNVIDIAJSONReasoning+2 | $0.50/1Mvia Baseten | 1M | Free/1M | Free/1M | — | — | |
| Nemotron-nano 12b v2-vlNVIDIAReasoningTool calling+1 | $0.60/1Mvia Vercel AI Gateway | 131K | $0.20/1M | $0.60/1M | — | — | |
 Nex-N2.5-MiniNex AgiPrompt cachingJSON+2 | $0.10/1Mvia OpenRouter | 262K | $0.025/1M | $0.10/1M | — | — | |
| Nex-N2.5-ProNex AgiPrompt cachingJSON+3 | $0.25/1Mvia OpenRouter | 262K | $0.075/1M | $0.25/1M | — | — | |
 North Mini CodeCohereApache 2.0Reasoning+1 | Free/1Mvia OpenRouter | 256K | Free/1M | Free/1M | — | — | |
 Nova 2 LiteAmazonReasoningTool calling+1 | $2.50/1Mvia OpenRouter | 1M | $0.30/1M | $2.50/1M | — | — | |
| Nova 2 Lite (EU)AmazonPrompt cachingJSON+3 | $2.50/1Mvia Amazon Bedrock | 1M | $0.30/1M | $2.50/1M | — | — | |
| NVIDIA Nemotron Nano 9B v2NVIDIAOpen weights · with termsReasoning+1 | $0.23/1Mvia Vercel AI Gateway | 131K | $0.06/1M | $0.23/1M | — | — | |
| O1 PreviewOpenAIPrompt cachingReasoning+1 | $48/1Mvia APIMart | — | $12/1M | $48/1M | — | — | |
| O1 ProOpenAIJSONReasoning+1 | $480/1Mvia APIMart | 200K | $150/1M | $600/1M | — | — | |
| O3 ProOpenAIJSONReasoning+2 | $64/1Mvia APIMart | 200K | $20/1M | $80/1M | — | — | |
| O4 MiniOpenAIPrompt cachingJSON+3 | $3.52/1Mvia APIMart | 200K | $1.10/1M | $4.40/1M | 1353 | — | |
| OpenAI GPT-5 MiniOpenAIPrompt cachingJSON+3 | $1.60/1Mvia APIMart | 400K | $0.25/1M | $2/1M | 1373 | — | |
| OpenAI GPT-5 NanoOpenAIPrompt cachingJSON+3 | $0.32/1Mvia APIMart | 400K | $0.05/1M | $0.40/1M | 1320 | — | |
| OpenAI GPT-5.2OpenAIPrompt cachingJSON+3 | $11.20/1Mvia APIMart | 400K | $1.75/1M | $14/1M | 1438 | — | |
| OpenAI GPT-5.2 ProOpenAIJSONReasoning+2 | $134.40/1Mvia APIMart | 400K | $21/1M | $168/1M | — | — | |
| OpenAI GPT-5.3 CodexOpenAIPrompt cachingJSON+3 | $11.20/1Mvia APIMart | 400K | $1.40/1M | $11.20/1M | 1371 | — | |
| OpenAI GPT-5.4OpenAIPrompt cachingJSON+3 | $12/1Mvia APIMart | 1.05M | $2.50/1M | $15/1M | 1469 | — |
Frequently asked questions
A model trained to work through a problem in intermediate steps, often called thinking, before it answers. It does better on math, coding and planning, at the cost of more tokens and slower replies.
The thinking steps are billed as output tokens, so one answer can use many times the tokens of a direct reply. Compare output prices, and check whether the model lets you limit or switch off its reasoning.
For short lookups, classification, extraction or chat, a regular model is faster and cheaper with little loss in quality. Keep reasoning for problems that really need several steps.
Can't find what you're looking for? Contact us