Open-weight inference pricing: how to pay less for DeepSeek and Kimi

What does open-weight inference cost per million tokens?

Provider choice changes the bill a lot. Anyone with GPUs can host the public weights for DeepSeek V4 Pro or Kimi K3, and for one model in the table below, the highest input price we found was seven times the lowest.

We should flag a conflict here. Wafer is an ernd partner, which gives us a reason to favor it. Each price comes from the provider’s own page, linked below the table, so you can check it.

We checked each provider’s serverless pricing page on November 5, 2026. These are list prices in US dollars per million tokens.

Model	Provider	Input	Cached input	Output
DeepSeek V4 Pro	Wafer	$0.28	$0.044	$3.96
DeepSeek V4 Pro	DeepInfra	$1.30	$0.10	$2.60
DeepSeek V4 Pro 0813	Together	$1.32	$0.13	$3.96
Kimi K3	Wafer	$0.60	$0.20	$15.00
Kimi K3	DeepInfra	$2.85	$0.285	$14.25
Kimi K3	Together	$3.00	$0.30	$15.00
Kimi K3	Fireworks	$3.00	$0.30	$15.00
GLM-5.3	Wafer	$0.20	$0.18	$1.80
GLM-5.3	DeepInfra	$0.5625	$0.125	$2.50
GLM-5.3	Together	$1.40	$0.26	$4.40
GLM-5.3	Fireworks	$1.40	$0.26	$4.40

Wafer’s docs name its public model catalog as the source for prices, which change as the catalog evolves. Other prices came from Together, Fireworks and DeepInfra. DeepInfra displayed GLM-5.3 at 38% off list prices of $0.90 input and $4.00 output, a discount that could change. Fireworks didn’t list DeepSeek V4 Pro.

There’s more to compare than dollars. Providers use different precision. OpenRouter’s providers page for DeepSeek V4 Pro 0813 shows mxfp4 for Wafer endpoints and fp8 for DeepInfra. When output quality matters, test the provider you’ll actually use with your own evals, and recheck its price before committing.

What does a month of agent traffic cost?

Input adds up fast with agents. Typical loops resend the full conversation at each step, tool definitions included, so input usually grows much faster than output.

Say your DeepSeek V4 Pro month uses 2 billion input tokens and 200 million output tokens, without caching.

Provider	Input cost	Output cost	Month
Wafer	2,000 × $0.28 = $560	200 × $3.96 = $792	$1,352
DeepInfra	2,000 × $1.30 = $2,600	200 × $2.60 = $520	$3,120
Together	2,000 × $1.32 = $2,640	200 × $3.96 = $792	$3,432

That’s $1,352 at Wafer against $3,120 at DeepInfra, less than half, and all of the gap is input pricing. Wafer’s output actually costs more. Reverse the token ratio and DeepInfra’s $2.60 output starts winning, so work out the bill with your own input and output counts before you move.

How do open-weight providers bill?

Most of them want prepayment. You load credits and usage spends them, and a request gets turned away if your balance won’t cover it.

Provider	How you pay	Auto-reload and limits	When the balance runs out
Wafer	Serverless is prepaid. You add credits in the dashboard	The docs don’t describe auto top-up	HTTP 402, insufficient_credits
Together	Fully prepaid. $15 minimum first purchase, $5 after. Scale and Enterprise contracts bill on contract terms	Auto-recharge only when the default payment method is a credit or debit card	API access suspended until you add credits
Fireworks	Serverless draws down prepaid credits. Contracted customers can move to post-paid invoicing	Monthly spend limits and Auto Reload	Not covered on the pricing page
DeepInfra	Add a card or prepay. Invoiced accounts are billed at the start of the following month	Spending limit	Not covered on the pricing page
OpenRouter	Prepaid credits, with a 5.5% fee on Standard and 8% on Business. Invoicing on Enterprise	Auto top-up	Not covered on the pricing page

Wafer’s usage docs gave us something worth checking when building agents. It compares your balance with a request’s maximum possible cost before running it. Set max_tokens too high with little credit left and you can hit a 402, even for an answer that would have been short. Keep max_tokens to what you need. Rejection before generation costs nothing. An interrupted stream can still incur charges for tokens already generated.

When does a dedicated endpoint pay off?

Only when you can keep the GPUs busy. Serverless bills by the token. With a dedicated endpoint you pay for every reserved GPU hour, including the hours with no traffic.

Provider	H100	H200	B200	Billing notes
DeepInfra	$2.90	$3.49	$4.89	Billed by the minute, invoiced weekly
Together Dedicated Inference	$5.49	Contact Together	$8.99	On-demand, pay as you go
Fireworks	$8.00	$8.00	$13.00	On Demand bills per GPU-second. Idle deployments scale to zero after one hour by default. Region-restricted deployments cost 1.5x
Wafer Dedicated	Not published	Not published	Not published	Reserved GPUs billed by the hour. Wafer sizes the setup with you

Hourly prices per GPU come from DeepInfra, Together and Fireworks. Wafer doesn’t publish its dedicated rates.

A 730-hour month with one H100 costs $2,117 at DeepInfra, $4,008 at Together or $5,840 at Fireworks. Big open models take multiple GPUs. One model in Wafer’s catalog lists 391B parameters. At 4 bits each, the weights alone take roughly 196 GB. Two 80 GB H100s can’t hold that, even before making space for the cache.

We’d start by doing the math. Add up a full month’s rental for every GPU your model needs, then compare that with serverless for the same model. Stick with serverless if it costs less. Dedicated becomes worth considering for steady traffic with latency requirements shared endpoints won’t guarantee, or when you need to serve your own fine-tuned weights.

How do you cut the bill without switching models?

Start with caching. Agents send the same system prompt, tools and history repeatedly, and cached input is much cheaper. DeepInfra’s DeepSeek V4 Pro input costs $0.10 per million tokens cached versus $1.30 uncached. Wafer’s usage docs say caching happens automatically, with cached prompt tokens billed at the cache-read rate. Fireworks charges half price or less for cached input, without a write fee. The prefix must match something the cache has already seen. Keep the beginning of your prompts unchanged between calls.

Next, try smaller models for less demanding steps. Fireworks prices DeepSeek V4.1 Flash at $0.30 input and $1.20 output, versus Together’s V4 Pro 0813 at $1.32 and $3.96. Wafer lists DeepSeek-V4.1-Flash at $0.19 input and $0.65 output. Your own evals have to settle whether Flash is good enough.

If a job can wait, try a cheaper tier. DeepInfra Flex is 0.8x base pricing, with slower replies and occasional unavailability, and its Priority tier costs 1.5x. Fireworks offers Standard, Priority, Fast, Batch and Spot serverless modes. Priority there is steep. Kimi K3 input costs $3.75 and output $18.75, against $3.00 and $15.00 on Standard.

Routing through OpenRouter adds a cost to watch. Standard credit purchases carry a 5.5% fee, according to its pricing page.

Pick a spending cap ahead of time. Fireworks supports monthly limits, and DeepInfra has a spending limit setting. We’d prefer stopping at a number we chose over discovering an empty balance in the overnight error logs.

How does the ernd card fit?

We make the ernd card, so keep that in mind here. It’s a credit card for US companies. AI vendor spend earns 4% back and everything else earns 1.5%. You get points redeemable as cash back.

Our partner Wafer gives ernd cardholders 10% back on Wafer inference. Its serverless endpoints host DeepSeek, Kimi and other open-weight models, and it offers dedicated endpoints too. On the $1,352 DeepSeek month above, that’s $135.20 back.

You shouldn’t have to take a partner’s word for how it performs, so here are two numbers from OpenRouter’s public data. From October 29 to November 4, 2026, Wafer served 12.4 trillion tokens on OpenRouter, more than any other independent open-weight inference provider. In OpenRouter’s one-week latency numbers as of November 5, Wafer had the lowest latency for DeepSeek V4 Pro 0813, at 0.32 seconds to first token. You can see both on OpenRouter’s Wafer provider page and its DeepSeek V4 Pro 0813 providers page.

Hopper, another partner, gives cardholders 8% back on its API credits plus a free first month, and pays ernd for its listing. The rest of our partners are on ernd.ai/partners. Request the card at app.ernd.ai.

Using closed models? Our how to cut LLM API costs guide covers ways to save with OpenAI and Anthropic.