Free to start

NVIDIA: Nemotron 3 Ultra vs Thinking Machines: Inkling

Try both on us. Starter credit covers your first requests, then one key routes Nemotron 3 Ultra and Inkling. Switch by changing the model name.

Compare routesProviders, price, context, and policy posture in one view.

Use auto when uptime mattersKeep a primary model and let fallback handle provider failures.

Same API shapeUse the OpenAI client and set the model you want.

Practical read

Monthly evidence

NVIDIA: Nemotron 3 Ultra has the lower published input-plus-output rate, more Credits provider routes and the lower measured p50 time to first token. Thinking Machines: Inkling has the larger context window.

$2.79575/1MNVIDIA: Nemotron 3 Ultra cheapest route

$5.32775/1MThinking Machines: Inkling cheapest route

883 msNVIDIA: Nemotron 3 Ultra fastest measured p50 TTFT via wandb

1441 msThinking Machines: Inkling fastest measured p50 TTFT via thinkingmachines

The price comparison adds each route's rate for one million input tokens and one million output tokens; your cost depends on your input/output mix. Prices and live route measurements can change. Use the monthly reports when you need a stable evidence window, then run a small task-specific eval before choosing a production default.

NVIDIA: Nemotron 3 UltraThinking Machines: Inkling

Model idnvidia/nemotron-3-ultra-550b-a55bthinkingmachines/inkling

PublisherNVIDIA Thinking Machines Lab

AI IQ — IQ 107#88

Context262,144 tokens524,288 tokens

Credits provider routes53

Cheapest route$2.79575/1M$5.32775/1M

Cached input$0.1055/1M to $0.1266/1M$0.17935/1M

Privacy posturehas ZDR routehas ZDR route

Fastest measured p50 TTFT883 ms via wandb1441 ms via thinkingmachines

Highest measured throughputnot enough datanot enough data

Recent route uptime range100.00%100.00%

Modes chat chat

NVIDIA: Nemotron 3 Ultra routes

Overview5 Credits routes
PricingInput, cached input, and output rates
ProvidersAll serving providers

Thinking Machines: Inkling routes

Overview3 Credits routes
PricingInput, cached input, and output rates
ProvidersAll serving providers

Related comparisons

Browse every comparison

NVIDIA: Nemotron 3 Ultra vs Z.ai: GLM 5.338 combined provider routes
NVIDIA: Nemotron 3 Ultra vs Z.ai: GLM 5.237 combined provider routes
NVIDIA: Nemotron 3 Ultra vs Z.ai: GLM 5.3 Flash37 combined provider routes
Thinking Machines: Inkling vs Z.ai: GLM 5.336 combined provider routes
Thinking Machines: Inkling vs Z.ai: GLM 5.235 combined provider routes
Thinking Machines: Inkling vs Z.ai: GLM 5.3 Flash35 combined provider routes
MoonshotAI: Kimi K3 vs NVIDIA: Nemotron 3 Ultra33 combined provider routes
DeepSeek: DeepSeek V4 Flash 0731 vs NVIDIA: Nemotron 3 Ultra31 combined provider routes

Production choice

Pick a default model. Keep fallback enabled.

TrustedRouter is useful when you know the model you want, but still need provider rollover, budget limits, usage records, and a prompt path you can verify.

OpenAI clientPython

client = OpenAI(
    base_url="https://api.trustedrouter.com/v1",
    api_key="sk-tr-v1-..."
)

response = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b-a55b",
    messages=messages,
)

Questions

Which should I use, NVIDIA: Nemotron 3 Ultra or Thinking Machines: Inkling?

Is NVIDIA: Nemotron 3 Ultra or Thinking Machines: Inkling cheaper?

The current cheapest TrustedRouter route is $2.79575/1M for NVIDIA: Nemotron 3 Ultra and $5.32775/1M for Thinking Machines: Inkling. The comparison uses current catalog prices and updates as provider pricing changes.

Is NVIDIA: Nemotron 3 Ultra or Thinking Machines: Inkling faster?

Current measured p50 time to first token is 883 ms for NVIDIA: Nemotron 3 Ultra and 1441 ms for Thinking Machines: Inkling. These are routed probe measurements, not vendor-advertised speeds, and update as new samples arrive.

Can I test NVIDIA: Nemotron 3 Ultra and Thinking Machines: Inkling with the same API?

Yes. Use the same OpenAI-compatible TrustedRouter base URL and API key, then change only the model id between nvidia/nemotron-3-ultra-550b-a55b and thinkingmachines/inkling. This makes side-by-side evals possible without maintaining two provider integrations.