Predictable capacity
Capacity reserved for your endpoint.
Reserved B200 and B300
Dedicated inference. Reserved capacity.
Isolated endpoints. Predictable performance. GPU time pricing.
- GPU time
- No token charges
- 99.9%
- Monthly availability SLA
- 7 models
- Current catalog
Why dedicated inference
Own your API capacity
Reserve capacity. Morph runs it.
Production model serving
Open models on B200 or B300.
Zero data retention
No stored prompts or responses.
OpenAI compatible
Keep your OpenAI client.
Built for coding agents
Deploy from Claude Code or Codex
Your agent can choose, buy, and monitor an endpoint.
- 1Compare models and plans.
- 2Approve price and terms.
- 3Track setup and connect.
- 4Verify request metadata.
agent terminal
# List models
morph dedicated models --json
# Buy after approval
morph dedicated create deepseek-v4-flash --balanced --yes
# Track setup
morph dedicated status <endpoint-id> --json
# Tail metadata
morph dedicated logs <endpoint-id> --followWorkload fit
When dedicated fits
Reserve for steady demand. Use serverless while demand is uncertain.
Sustained production traffic
Steady traffic favors reserved capacity.
A model you already trust
You already chose the model.
Capacity as an operating input
Shared limits now hurt production.
Common workloads
Coding agents and tool use
Long agent loops with tools and code.
High volume chat
Steady production assistants.
Reasoning and extraction
Predictable document processing.
Multi step workflows
Chained calls without shared limits.
Models and capacity
Choose a model and capacity
Monthly reservations are invoiced upfront.
ModelBest forContext
DeepSeek V4 Flash 0731
Coding, reasoning, tools
1M context
Qwen 3.6 27B
Fast chat, extraction, multimodal
131K context
Gemma 4 31B
Reasoning, tools, vision
175K context
MiniMax M3 428B
Agents, software, reasoning
256K context
GLM 5.3 Flash
Multimodal coding and tools
1M context
GLM 5.3 744B
Complex coding and tools
1M context
Kimi K3 2.8T
Frontier agents, vision, long runs
1M context
Test with your own traces before buying.
2× B200
$8.64/GPU-hr
up to 6.6x cheaper than token-based pricing
- 2× B200-equivalent reserved capacity
- Dedicated endpoint with zero data retention
- 99.9% availability SLA
- Migration services
- Slack support
- Dedicated customer success
- 24/7 Incident monitoring
4× B200
$8.43/GPU-hr
up to 6.8x cheaper than token-based pricing
- Everything in 2× B200
- 2× the reserved capacity
- Lower GPU-hour rate
- ~2× modeled token capacity
8× B200
$8.16/GPU-hr
up to 7.0x cheaper than token-based pricing
- Everything in 4× B200
- 2× the reserved capacity
- Lowest GPU-hour rate
- ~2× modeled token capacity
- Full-node-equivalent capacity envelope
Make the right tradeoff
Dedicated or serverless?
Dedicated buys control. Serverless avoids idle cost.
- Dedicated for steady traffic.
- Serverless for uncertain traffic.
- Model your own workload below.
| Decision | Dedicated | Serverless |
|---|---|---|
| Traffic | Steady production | Variable or early |
| Billing unit | Reserved GPU time | Input and output tokens |
| Capacity | Reserved | Shared |
| Limits | Purchased capacity | Account limits |
| Best for | Control at scale | No idle cost |
Planning tools
Turn traffic into a capacity plan
Compare monthly serverless cost with reserved capacity using your model, token volume, cache reuse, and required user speed.
LLM cost and capacity calculator
Enter monthly input and output tokens, cache hit rate, and required generation speed. Unverified capacity routes to a sizing review.
Open calculatorLLM inference benchmarks, explained
Translate InferenceX latency, speed, and concurrency into buying guidance.
Read benchmarksAgent economics
Optimize cost per completed task
Token price is only one input. Completion quality, generation speed, retries, and context reuse determine what a useful task costs.
Coding agents repeatedly send system instructions, repository context, tool output, and conversation history. Tiered caching keeps the hottest prefixes in GPU memory, warm prefixes in CPU memory, and longer lived context on NVMe.
Choose a latency target first. Then measure throughput inside that target using a representative trace. An attractive batch result is not a capacity promise for an interactive product.
Compare attributed inference measurementsOpenAI compatible
Change one endpoint
Keep the OpenAI SDK. Swap the base URL.
python
from openai import OpenAI
client = OpenAI(
api_key=os.environ["MORPH_API_KEY"],
base_url="https://your-endpoint.morphllm.com/v1",
)
response = client.chat.completions.create(
model="your-dedicated-model",
messages=[{"role": "user", "content": "Hello"}],
)Managed operations
Morph runs it
You own the app. Morph owns operations.
- Provisioning and serving
- Monitoring and incidents
- Model performance tuning
- Migration and support
Operational visibility
See traffic, tokens, cache, latency, status, and capacity.
Security and privacy
Isolated credentials. Details in the Privacy Policy.
Availability and support
99.9 percent SLA, monitoring, Slack, and customer success.
Endpoint health
Illustrative dashboard
Live signals
Request volume
Requests over time
Latency
p50 and p95
Cache efficiency
Hit rate and cached tokens
Capacity demand
Reserved service usage
Request status
Success and error codes
Token usage
Input, output, and cached
CLI request logs
Metadata only live tail
Compare DeepSeek V4 Flash 0731
Compare reserved capacity with token pricing.
Workload assumptions
Set utilization and cache share.
Capacity utilization100%
Capacity vs. API
API reference: $0.14/M input, $0.0028/M cached, $0.28/M output.
| Reserved capacity | Tokens / hr | Hourly | API pricing | Savings |
|---|---|---|---|---|
2× B200 $19.20/endpoint-hr | 815.6M 725M in · 90.6M out | $19.20/hr $0.02354/M | $0.07628/M $62.22/hr | 3.2× cheaper |
4× B200 $37.48/endpoint-hr | 1.63B 1.45B in · 181.2M out | $37.48/hr $0.02298/M | $0.07628/M $124.44/hr | 3.3× cheaper |
8× B200 $72.50/endpoint-hr | 3.26B 2.9B in · 362.5M out | $72.50/hr $0.02222/M | $0.07628/M $248.88/hr | 3.4× cheaper |
Join the teams behind 400+ production agents.
Read customer stories
Inference at work today
Private tokens / dayProduction agentsFast Apply throughputDedicated availability
Private tokens / dayProduction agentsFast Apply throughputDedicated availability
Capacity or tokens?
Morph Dedicated vs. OpenRouter.
Public Gateway
Pay-as-you-go
Morph Dedicated
Reserved capacity
Economics
Billing unit
Reserved GPU-hour
Cost at sustained utilization
Falls with utilization
Full-use economics
Up to 7× cheaper
Idle capacity
Idle time billed
Credit purchase fee
None
Initial commitment
90 days
Best fit
Steady traffic
Platform
Model catalog
7 dedicated models
Serving path
Morph operated
OpenAI-compatible API
Spend controls
Reliability and privacy
Zero data retention
Default
Prompt and response logging
Off by default
Included with Morph Dedicated
Reserved B200 or B300 capacity
Isolated endpoint
No credit purchase fee
99.9% availability SLA
Migration services
Slack support
Dedicated customer success
24/7 endpoint monitoring
Details
Dedicated inference FAQ
Billing and operations.
How does hourly billing work?+
GPU hours set the monthly price. Each month is invoiced upfront, including idle capacity. Tokens are not billed.
Are tokens included or capped?+
No. You buy GPU time, not tokens. Calculator figures are estimates, not guarantees.
How is the GPU-hour price calculated?+
GPU rate times reserved GPUs times 730 hours. The month is billed upfront.
Is the physical hardware exclusive to me?+
No. Your endpoint, credentials, and capacity are isolated. Idle hardware may serve other traffic.
How quickly is an endpoint activated?+
Provisioning starts after payment. If activation exceeds two hours, Morph cancels and refunds the order.
What is the term and cancellation policy?+
The initial term is 90 days, then monthly with 30 days’ notice. No partial month refunds.
Does Morph retain prompts or responses?+
No. Morph stores only operational and billing metadata.
Can a coding agent deploy and manage an endpoint?+
Yes. Agents can choose, buy, track, list, and inspect endpoints through the Morph CLI. Purchases require user approval.
What does Morph manage?+
Morph handles provisioning, serving, monitoring, incidents, tuning, and migration. You own the app.
How can I monitor the endpoint?+
Use the dashboard for metrics and the CLI for status and metadata logs.
Is the API compatible with OpenAI clients?+
Yes. Change the base URL, key, and model name.
What does the availability SLA cover?+
It includes 99.9 percent monthly availability. Your agreement defines measurement and remedies.