IBM「開放權重模型九成效果、一成成本」説法是否成立？作者以公開標價檢視

TL;DR

On IBM's Q3 2026 earnings call, CEO Arvind Krishna said open-weight models now deliver about "90 percent as effective" results at "anywhere from 25 percent to 10 percent of the token cost of the frontier models", and that most businesses can build useful AI infrastructure in "two or three racks" (The Register, 2026-10-22). I checked the cost half of that claim against the public price lists on 2026-10-23. What I found:

1. The price claim is true, and then some. On an illustrative 8,000-input, 500-output-token job, DeepSeek V4.1-Flash at off-peak list prices costs 2.9% of Claude Opus 5 and 7.1% of GPT-6 Sol. V4-Pro costs 11.9% and 29.9%.
2. "Frontier" is doing a lot of work in that sentence. The closed vendors' own small tiers (GPT-6 Luna, Claude Haiku 5.5) cost 1.05 dollars per 1,000 of those requests, which is 30% less than DeepSeek's cheapest. Open weights are not uniquely cheap. Small models are.
3. On-prem is a utilization bet, not a price bet. Under an assumed 30 dollars per 8-GPU node-hour and 12,000 output tokens per second per node, owning hardware beats GPT-6 Sol's output price once the cluster is busy about 7% of the time, but never beats DeepSeek Flash's hosted price at any utilization.
4. The honest case for racks is data control and predictable spend, not "10 percent of the cost".

All hardware numbers below are illustrative assumptions, labeled as such, and the formulas are included so you can replace them with your own.

What exactly did IBM claim?

According to The Register's report of the call, Krishna described a shrinking quality gap: in 2024 "the top model might score a 70, and the open model might score a 20 or a 30", by late 2025 it was "70 to 50", and now open weights are "90 percent as effective". The cost line was "anywhere from 25 percent to 10 percent of the token cost of the frontier models", framed as "Would you like something that's 10 percent of the cost if it's 90 percent as good as the best?" He named tasks like summarizing an email, examining a contract, answering a customer call and composing an email.

Two things are worth noting before any math. First, the article names no benchmark behind the 70, 50 and 90 figures, so the quality half cannot be checked from the public record. I did not try. Second, the same call reported Q3 revenue of 17.1 billion dollars (up 5%), infrastructure revenue of 3.7 billion dollars and Red Hat growing 11%. IBM sells hardware, software and consulting around on-prem AI, so the framing is a business position as well as an observation. That does not make it wrong. It means the cost claim deserves a calculation.

What does one workload cost on nine price lists?

I picked a job that matches IBM's examples: ticket triage or email summarization with 8,000 input tokens (instructions, a thread, some context) and 500 output tokens, run 1,000 times, with no caching. List prices, USD per million tokens, were read on 2026-10-23 from DeepSeek, Anthropic and press coverage of OpenAI's launch pricing (The Next Web). OpenAI's own pricing page returned HTTP 403 to my fetcher, so treat the GPT-6 numbers as secondary-sourced.

Reading the table against IBM's sentence: if "frontier" means Opus 5, V4-Pro sits inside the 10 to 25% band (11.9%) and Flash sits far below it. If "frontier" means the mid-priced flagship tier (Sonnet 5 or GPT-6 Sol), V4-Pro lands at about 30%, just above the band, and Flash lands under it. IBM's range is, if anything, conservative on hosted list prices.

Two billing details move these numbers. DeepSeek's peak window is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, at double the off-peak rate. And caching changes input-heavy jobs a lot. If 6,000 of the 8,000 input tokens are a shared prefix that hits the cache, the same 1,000 requests cost 25.50 dollars on Opus 5 (cache read 0.50), 10.20 on GPT-6 Sol (0.20), 0.618 on DeepSeek Flash off-peak (0.003) and 0.51 on both Luna and Haiku 5.5 (0.01). Cached prefixes shrink the absolute gaps but not the ordering.

Is the cheap part really "open"?

Look at the bottom two rows. GPT-6 Luna and Claude Haiku 5.5 are closed models, and for this job they cost 1.05 dollars per 1,000 requests, about 30% below DeepSeek Flash's off-peak 1.50. The closed vendors have collapsed their own small-tier prices to the same neighborhood as hosted open weights.

That matters for how to read "open weight is 10% of the cost". The saving comes from picking a small, sparse model for a simple task. It is a model-size and routing effect, and it is available inside a closed API too. What open weights add is not a lower token price at the API but a different set of options: you can host them yourself, fine-tune them, pin a version forever, and keep data inside your network. Those are real benefits. They are just not the same claim as price.

When does owning hardware beat the API?

This is where "two or three racks" comes in. Self-hosted cost per token is a fixed hourly cost divided by the tokens you actually produce, so utilization decides everything:

# cost per 1M output tokens when you own the hardware
def cost_per_m_out(node_usd_per_hour, node_out_tok_per_s, utilization):
    return node_usd_per_hour / (node_out_tok_per_s * 3600 * utilization) * 1e6

# break-even utilization against an API output price
def breakeven_util(node_usd_per_hour, node_out_tok_per_s, api_usd_per_m_out):
    return cost_per_m_out(node_usd_per_hour, node_out_tok_per_s, 1.0) / api_usd_per_m_out

print(cost_per_m_out(30, 12000, 1.0))   # 0.694 USD per 1M tokens at full load
print(breakeven_util(30, 12000, 10.0))  # 0.069, i.e. 6.9% vs GPT-6 Sol output price

My assumptions, to be replaced with yours: 30 dollars per 8-GPU node-hour all-in (hardware depreciation, power, space, support) and 12,000 output tokens per second per node at full batch for a large sparse model. Neither number is measured. Real throughput depends on the model, context length, batch shape and serving stack, and your hardware price depends on what you negotiated.

Under those assumptions:

The curve reads like this. At 10% average utilization the owned cluster costs about 6.94 dollars per million output tokens: cheaper than Sol's 10, far cheaper than Opus's 25, and 3.5 times more than V4-Pro. At 50% it costs 1.39 dollars, which beats V4-Pro but still loses to Flash at 0.60. Even at 100% it is 0.69 dollars.

Sensitivity matters here, so I varied both inputs against Sol's 10 dollar output price:

Against flagship-priced APIs, owning hardware pays off at modest utilization across a wide range of assumptions. Against hosted small or sparse models, it does not pay on price in any cell of this grid.

What would two or three racks look like?

Take "three racks" as 24 eight-GPU nodes. At 30 dollars per node-hour that is 6.31 million dollars a year. The same money buys 631 billion output tokens a year at GPT-6 Sol's price, which is 1.73 billion a day, or an average of 20,000 tokens per second. Peak capacity of the 24 nodes under my assumptions is 288,000 tokens per second, so you need roughly 6.9% average utilization to break even against Sol, the same number as above. At Opus pricing the break-even average is 8,000 tokens per second.

Running the same arithmetic against Flash's off-peak output price gives 10.5 trillion tokens a year, an average of 333,000 tokens per second, which is more than the cluster can produce even at full load. In other words, on these assumptions, a company that only wanted the cheapest token should not buy racks at all.

What does this model leave out?

The calculation is a deliberately simple floor and ceiling, and several things push it both ways:

• Prefill. My capacity figure counts output tokens only. Input-heavy jobs such as the 8,000-in, 500-out workload above spend most compute on prefill, so real capacity in "requests per second" is lower than the output-token rate suggests.
• Idle headroom. Latency targets and peak-hour traffic force you to run below full batch. A cluster that averages 30% utilization may be fine for users and still be wasteful on paper.
• People and refresh. Operations staff, monitoring, security reviews and hardware refresh cycles are not in a 30 dollar hourly figure unless you put them there.
• Data egress and compliance. For contracts, patient notes or call recordings, the alternative to racks may not be "cheaper API" at all. It may be "no AI on that data". That changes the comparison from price to permission.
• Quality. Everything above assumes the cheaper model is good enough for the task. IBM's "90 percent as effective" is the load-bearing claim, and it comes without a named benchmark.

How should a team decide?

A practical order of operations, based on the numbers above:

1. Measure your own task quality on a small model first. If Luna, Haiku 5.5 or Flash passes your acceptance test, you have already captured most of the saving without any hardware.
2. Add caching and routing before racks. A shared prefix cut Sol's cost on this job from 21.00 to 10.20 dollars per 1,000 requests, without changing anything about the model.
3. Estimate your average, not peak, token rate. Convert monthly spend to tokens per second. If it is far below what a cluster can produce, the cluster will sit idle, and the break-even utilization above tells you how idle is too idle.
4. Buy racks for reasons that price does not capture: data that cannot leave, fixed versions for regulated workflows, or spend that must be predictable.

Reproduce it

The model is the two functions above plus the price table: workload tokens times list price for the API side, and node cost divided by tokens actually produced for the hardware side. Replace the workload size, node cost and node throughput with your own measurements. Prices change often, so re-read the three pricing pages before trusting any figure here.

FAQ

Is IBM's claim that open-weight models cost 10 to 25% of frontier models accurate?

On hosted list prices, yes, and it is conservative for the cheapest options. For an 8,000-input, 500-output job, DeepSeek V4.1-Flash costs 2.9% of Claude Opus 5 and 7.1% of GPT-6 Sol, and V4-Pro costs 11.9% and 29.9%. The quality claim of "90 percent as effective" cites no benchmark in the published report, so it cannot be verified from public sources.

Are open-weight models cheaper than closed models?

Not by nature. Closed small tiers such as GPT-6 Luna and Claude Haiku 5.5 list at 0.10 dollars per million input tokens and 0.50 per million output tokens, which makes the example job cheaper (1.05 dollars per 1,000 requests) than DeepSeek V4.1-Flash off-peak (1.50). Price tracks model size and sparsity more than license.

At what utilization does running your own GPUs beat paying an API?

It depends on three numbers: node cost per hour, output tokens per second per node, and the API price you compare against. With 30 dollars per node-hour and 12,000 output tokens per second, break-even is 6.9% against GPT-6 Sol's 10 dollar output price and 35.1% against DeepSeek V4-Pro's 1.98 dollars. Against DeepSeek Flash's 0.60 dollars it does not break even even at 100%. These hardware inputs are illustrative, not measured.

Does prompt caching change the open versus closed comparison?

It lowers every bill and keeps the ordering. With 6,000 of 8,000 input tokens cached, 1,000 requests cost 25.50 dollars on Opus 5, 10.20 on GPT-6 Sol, 0.618 on DeepSeek Flash off-peak and 0.51 on Luna or Haiku 5.5.

What is the real reason to run open-weight models on-prem?

Control rather than price: keeping sensitive data inside your network, pinning a model version, fine-tuning on private data and getting a fixed cost line. Those are legitimate reasons, and they hold even when the API is cheaper per token.

Further reading

• Shorter outputs, same price: what DeepSeek V4.1-Flash-1017 changes in an agent's bill
• The Same Model Can Cost 14x More Depending on Who Serves It