Frontier open weights, served on custom kernels for codegen.
| Model | Speed | Input | Cache Read | Output | Context | |
|---|---|---|---|---|---|---|
Kimi K3 2.8TFeatured morph-kimik3 | ~100 tok/sec | $2.80/1M | $0.29/1M | $14.00/1M | 1M | |
Kimi K3 2.8T, latency-tuned morph-kimik3-fast | fastest | $6.00/1M | $0.60/1M | $22.50/1M | 1M | |
GLM-5.2 744B morph-glm52-744b | ~80 tok/sec | $1.10/1M | $0.22/1M | $4.10/1M | 1M | |
MiniMax M3 morph-minimax3-428b | ~90 tok/sec | $0.30/1M | — | $1.20/1M | 256k | |
Qwen 3.5 397B morph-qwen35-397b | ~180 tok/sec | $0.50/1M | $0.30/1M | $3.50/1M | 262k | |
DeepSeek V4 Flash 0731 morph-dsv4flash | ~150 tok/sec | $0.14/1M | $0.07/1M | $0.28/1M | 1M | |
Gemma 4 31B morph-gemma4-31b | ~120 tok/sec | $0.14/1M | $0.08/1M | $0.40/1M | 175k | |
Qwen 3.6 27B morph-qwen36-27b | ~100 tok/sec | $0.29/1M | — | $2.40/1M | 131k |
Purpose-built APIs that offload the bottlenecks: code editing, search, context management, and per-turn classification.
| Model | Speed | Price | Context | |
|---|---|---|---|---|
Fast Apply Merges LLM code edits into files at 98% accuracy. | ||||
Fastest model morph-v3-fast | 10,500+ tok/sec | $0.80/1M in$1.20/1M out | 262k | |
Most diversePopular morph-v3-large | 5,000+ tok/sec | $0.90/1M in$1.90/1M out | 262k | |
Code Search Agentic repo search. #1 on SWE-Bench Pro. | ||||
Fast context for agentsNew morph-warp-grep-v2 | — | $0.80/100K | 100K (1M for Pro) | |
Compaction Verbatim context compression, shrinking context 50–70%. | ||||
Verbatim context compaction morph-compact | < 2s P99 | $0.20/1M in$0.50/1M out | 1M | |
Reflex Per-turn classifiers for jailbreaks, looping, and frustration. | ||||
Realtime per-turn classifiers morph-reflex-v1 | < 90ms | $0.001/event$0.0005 over 1M/mo | 64K | |
Batch (offline) morph-reflex-v1 | — | $0.0005/event$0.00025 over 1M/mo | 64K | |
| Model | Price | Context | |
|---|---|---|---|
Difficulty-based model routing morph-router | $0.005/request | — |
Pay per token, or go flat-rate with Scale.
Top up credits, use any model
For individuals who can't stop coding
Reserve B200 or B300 capacity by the GPU-hour, from $9.50 per reserved B200-hour, behind a 99.9% monthly availability SLA. Paying for GPU time instead of tokens runs up to 7× cheaper than DeepSeek API pricing.
$9.06/GPU-hr
up to 6.6x cheaper than token-based pricing
$8.84/GPU-hr
up to 6.8x cheaper than token-based pricing
$8.55/GPU-hr
up to 7.0x cheaper than token-based pricing
Adjust useful utilization and cache share to compare hourly capacity with the selected model's pay-per-token API rates.
Model the effective token price of hourly capacity.
Openrouter per-token reference is $0.14/M input, $0.0028/M cached input, and $0.28/M output.
| Reserved capacity | Tokens / hr | Hourly | API pricing | Savings |
|---|---|---|---|---|
2× B200 $20.14/endpoint-hr | 855.4M 760.4M in · 95M out | $20.14/hr $0.02354/M | $0.07628/M $65.26/hr | 3.2× cheaper |
4× B200 $39.31/endpoint-hr | 1.71B 1.52B in · 190.1M out | $39.31/hr $0.02298/M | $0.07628/M $130.51/hr | 3.3× cheaper |
8× B200 $76.04/endpoint-hr | 3.42B 3.04B in · 380.2M out | $76.04/hr $0.02222/M | $0.07628/M $261.03/hr | 3.4× cheaper |
Compare Morph Dedicated with OpenRouter's pay-as-you-go API.
Pay-as-you-go
Reserved capacity
Billing, activation, ownership, and retention details.
Charges accrue by reserved GPU-second while your endpoint is active, including idle time, and are invoiced monthly. Tokens are not a billing unit and there is no per-token overage charge.
No token allowance is consumed. You pay for reserved GPU time and can use the available capacity throughout that time. Calculator token figures are an illustrative economics model, not credits, measured benchmarks, or throughput guarantees.
The listed per-GPU-hour rate is the actual billing rate. The endpoint-hour rate multiplies it by the number of reserved B200-equivalents. Monthly reference costs use a standard 730-hour month; actual invoices use active reserved GPU-seconds.
No. The endpoint, credentials, and purchased service capacity are isolated for your use, but Morph may serve other traffic on idle hardware. Your requests retain the service entitlement you purchased, and successful overflow capacity counts toward availability.
Provisioning starts automatically as soon as payment is committed. An account with billing on file is charged directly; an account without one completes a bank-account checkout first. If Morph cannot activate the endpoint within two hours, the subscription is canceled and any collected payment is automatically voided or refunded.
The initial term is 90 days. After that, service continues month to month with 30 days’ notice. Cancellation takes effect on the first billing boundary at least 30 days after notice, without partial-month prorations.
No. Dedicated endpoints default to zero data retention: Morph does not persist prompts, responses, or model-generated content. Only bounded operational and billing metadata is retained.
Over 100 billion tokens per day served on dedicated capacity with custom speculators and caching, at large discounts over the public pricing above. Includes SSO, custom rate limits, and priority support.
