NGen-4.5 Vayu
Our new flagship frontier model. NGen-4.3 Pro intelligence, delivered at 400 tokens per second.
Reasoning
Highest
Speed
400 tokens/s
Intelligence
NGen-4.3 Pro
Same model weights
Price
$3.65 / $11.35
Input / output per 1M tokens
Built for speed
Frontier intelligence at the speed of wind.
NGen-4.5 Vayu is our new flagship frontier model. It matches NGen-4.3 Pro intelligence while generating up to 400 tokens per second, and is served from TNSA regions India South, India Central, and US East-1.
- Model ID
- ngen-4.5-vayu
- Output speed
- Up to 400 tokens/s
- Context window
- 131,072 tokens, extendable to 1,048,576
- Max output
- 131,072 tokens
- Input
- Text
- image
- audio
- video
- Output
- Text
- audio
- Reasoning levels
- Low, Medium, High, Extended
- Rate limits
- 1,000 requests/min · 250,000 tokens/min
Availability
Served from TNSA regions.
Served from every TNSA region.
India South
in-south
India Central
in-central
US East-1
us-east-1
Pricing
Simple, metered pricing.
Developer API pricing in USD. Input and output are metered and billed separately; taxes are added where applicable.
Input
$3.65
per 1M tokens
Output
$11.35
per 1M tokens
Cached input
—
Not included in standard pricing
Capabilities
Everything in the API, out of the box.
Endpoints
- Chat Completionsv1/chat/completions
- Responsesv1/responses
- Batchv1/batch
- Realtimev1/realtime
- Embeddingsv1/embeddings
- Image generationv1/images/generations
- Speech generationv1/audio/speech
- Transcriptionv1/audio/transcriptions
Features
- Streaming
- Structured outputs
- Function calling
- JSON mode
- Fine-tuning
- Distillation
Tools
- Web search
- File search
- Code interpreter
- Apply patch
- Hosted shell
- MCP
- Computer use
- Tool search
Model evaluations
NGen-4.3 Pro intelligence, unchanged.
NGen-4.5 Vayu runs the same model weights as NGen-4.3 Pro on a faster inference stack, so its evaluation scores are identical. Scores are percentages; higher is better.
| Benchmark | NGen-4.5 Vayu | Kimi K3 | GPT-5.6 Sol | GPT-5.6 Terra | Claude Fable 5 | Claude Opus 4.8 | Claude Opus 5 | GLM-5.2 | Gemini 3.7 Flash |
|---|---|---|---|---|---|---|---|---|---|
| Terminal Bench 2.1 | 80.2 | 88.3 | 88.8 | 87.4 | 83.1 | 78.9 | 89.1 | 81.0 | 78.4 |
| SWE-bench Pro | 65.3 | 58.5 | 64.6 | 63.4 | 80.0 | 69.2 | 74.8 | 62.1 | 52.6 |
| NL2Repo-Bench | 48.1 | 49.2 | 54.8 | 52.6 | 68.2 | 69.7 | 71.5 | 48.9 | 44.0 |
| DeepSWE 1.1 | 50.0 | 67.5 | 72.7 | 69.6 | 69.7 | 59.0 | 68.8 | 46.2 | 45.1 |
| AgentCoWorkBench | 75.1 | 71.5 | 72.3 | 69.8 | 74.8 | 68.2 | 75.9 | 67.1 | 64.6 |
| JobBench | 35.7 | 52.9 | 48.4 | 44.1 | 53.4 | 45.4 | 57.4 | 38.6 | 31.3 |
| Agents' Last Exam — Pass@1 | 21.8 | 15.2 | 24.6 | 21.8 | 23.1 | 18.7 | 27.8 | 14.5 | 10.6 |
| Agents' Last Exam — Score | 47.3 | 38.4 | 52.7 | 50.4 | 40.5 | 45.2 | 56.8 | 37.0 | 32.1 |
Coding benchmarks
NGen-4.5 Vayu against leading models. Scores are percentages; higher is better.
- NGen-4.5 Vayu
- Kimi K3
- GPT-5.6 Sol
- GPT-5.6 Terra
- Claude Fable 5
- Claude Opus 4.8
- Claude Opus 5
- GLM-5.2
- Gemini 3.7 Flash
CI denotes code interpreter. Scores are those measured for NGen-4.3 Pro.
The team