Signal / Japan-hosted inference | Business software buyers / AI input and output costs
A cheap AI prompt can still leave a bigger bill for the answer
Fixstars now offers a version of GLM-5.3 to direct business customers. Its output rate is about 20 times its ordinary input rate. A team comparing AI costs can use a paper worksheet to count both and add the time people spend checking the answers. A low input price or a Japan server address does not show what an accepted answer costs.
- Use this if
You are authorised to compare business AI costs and have aggregate past counts or can work with made-up numbers on paper.
- No action if
No matched data or written rate for the route you would use? Keep the comparison open. Do not buy, run requests, upload files, use customer data or change services.
Asia/Singapore
Direct business route for the quoted launch. All quoted rates are Japanese yen per million tokens, not a converted rate or an invoice. Server location does not establish where data is stored or logged, approval, safety, service levels or equivalent quality. No account, inference request, data transfer or purchase was used.
First published
TechnologyOTR Emerging Data
Technology | What an accepted answer costs
Calculate the cost of the answer, not just the prompt
The new option puts one more price on the table. The mix of tokens and staff time determines what a usable result costs.
What the sources show
At 10 am Japan time on 20 November, 9 am in Singapore, Fixstars announced that direct customers could use its version of GLM-5.3. The company says it runs the model on servers in Japan. The launch lists ¥10.49 for ordinary input, ¥10.34 for cached input read back and ¥210 for output, each per one million tokens. Access through FastMetal was still planned for this model.
- 3 October 2026vLLM reports a separate GLM-5.3 validation
- 20 November, 09:00 SGTFixstars announces direct business availability
- Later, not datedFastMetal route planned for this model
The new connection
A low input price does not mean the whole task is cheap. At these rates, an output token costs about 20 times an ordinary input token. With one million ordinary input tokens and 100,000 output tokens, output costs ¥21 and input costs ¥10.49. That is a paper example, not usage or an invoice.
- Ordinary input10.49
- Cached input read back10.34
- Output210
- Fixstars: timed GLM-5.3 launch · 2026-11-20
- Fixstars statutory disclosure: rates and scope · 2026-11-20
- A low input price gets attention
- Multiply both rates by the actual mix
- Compare cost per accepted answer
A realistic next move
Use aggregate input and output counts already recorded for an approved workload, or use made-up counts. On paper, apply each route’s written rates, add any known contract fees and estimate staff time spent checking or correcting results. Keep tasks and acceptance rules the same. If the price for your route is not in writing or you do not have comparable evidence on results, leave the decision open. Do not sign up, purchase or run a test for this check.
Direct business route for the quoted launch. All quoted rates are Japanese yen per million tokens, not a converted rate or an invoice. Server location does not establish where data is stored or logged, approval, safety, service levels or equivalent quality. No account, inference request, data transfer or purchase was used.
The connection
A cheaper prompt can still leave output as the largest charge
A direct-contract route for GLM-5.3 began at Fixstars. The model already existed. FastMetal access for this model was not yet announced as live.
- Observed
At 10 am Japan time on 20 November, 9 am in Singapore, Fixstars announced that direct customers could use its version of GLM-5.3. The company says it runs the model on servers in Japan. The launch lists ¥10.49 for ordinary input, ¥10.34 for cached input read back and ¥210 for output, each per one million tokens. Access through FastMetal was still planned for this model.
- The connection
A low input price does not mean the whole task is cheap. At these rates, an output token costs about 20 times an ordinary input token. With one million ordinary input tokens and 100,000 output tokens, output costs ¥21 and input costs ¥10.49. That is a paper example, not usage or an invoice.
- How the effect travels
Price arithmetic uses three quantities: ordinary input, cached input read back and output. At the quoted rate, about 49,952 output tokens cost as much as one million ordinary input tokens. Adding retries, human review or other fees can change which route costs less per result.
The independent vLLM project records an October 3 validation using FP8 on AMD hardware. Fixstars describes a different numerical format, NVFP4. These are different implementations of the same named model, not proof they produce the same results or meet the same response time.
The decision path
Compare both routes on paper
- 1. Choose the same task
Use aggregate approved counts or made-up figures, with the same standard for an accepted answer.
- 2. Write each rate separately
Record currency and route. Multiply ordinary input, cached input if applicable, and output by the right rates.
- 3. Count accepted results
Add reviewer time, repeat work and confirmed fees; keep any unsupported quality difference open.
- 4. Keep or stop
If no comparable evidence or written scope is available, keep the existing decision. No live tests or purchases.
Japanese servers as described by the issuer; direct business pricing in yen. No generalisation to a FastMetal quote, tax, foreign exchange or data handling terms.
Optional research / Checks 1–10
What the sources support
- Money flow
Listed rates only; no money verified.
- Orders
No customer contracts or renewals observed.
- Bottleneck
Output can dominate despite a much larger input count.
- Beneficiaries
Possible use for authorized budget teams.
- Company finances
No margin or revenue inferred.
- Costs payback
Yen per million; calculate review time separately.
- Ownership
New route, no ownership change inferred.
- Hiring
No staffing effect verified.
- Government
No public award or residency certification claimed.
- Language
Japanese launch and filing; English prior technical documentation.
Optional research / Checks 11–20
What would change the reading
- Behaviour
No paid use, active user or demand evidence checked firsthand.
- Contradiction
A low input rate can coexist with a higher output share.
- Missing proof
No billing, quality or customer benchmark.
- Approval
Formal direct route announced; FastMetal planned.
- Alternatives
Retain existing route if it performs as well at lower full cost.
- Geography
Japan server claim is not proof of data storage, retention or eligibility elsewhere.
- History
October FP8 evidence precedes November NVFP4 offer; implementations differ.
- Awareness
No proven cross-currency savings or current tax and terms.
- Countercase
Short outputs or lower correction costs may favour the existing route.
- Action clock
Paper comparison only; stop if price, accepted outputs or comparable scope are missing.
Evidence & next review
Read the originals
Fixstars: timed GLM-5.3 launch
Published 2026-11-20T10:00:00+09:00
Issuer announced availability on the direct route at 09:00 Singapore time; the public clock is not first customer use. FastMetal was planned, not live for this model at announcement.
Source record
Publisher: Fixstars
Original language: Japanese
Source publication date 2026-11-20T10:00:00+09:00
Fixstars statutory disclosure: rates and scope
Published 2026-11-20
10.49 yen ordinary input, 210 yen output, 10.34 yen cache reads, each per million tokens. The latency benchmark is the issuer’s setup, not a customer result.
Source record
Publisher: Fixstars
Original language: Japanese
published date precision: day
Source publication date 2026-11-20
Explains direct corporate and gateway routes generally; its DeepSeek price table does not price GLM-5.3 or establish GLM-5.3 availability on the gateway.
Source record
Publisher: Fixstars
Original language: English
The recipe records FP8 validation dated 3 October 2026 on AMD MI350X hardware; not a matched Fixstars benchmark or an announcement date.
Source record
Publisher: vLLM project
Original language: English
Original developer identifies GLM-5.3 separately from GLM-5.3-Flash; docs about the model family are not evidence that two hosts give equivalent results.
Source record
Publisher: Z.ai
Original language: English
Direct business route for the quoted launch. All quoted rates are Japanese yen per million tokens, not a converted rate or an invoice. Server location does not establish where data is stored or logged, approval, safety, service levels or equivalent quality. No account, inference request, data transfer or purchase was used. Fixstars sources are one family. vLLM and Z.ai document earlier model use but do not verify launch outcomes. The headline vendor speed result was measured in a particular Tokyo setup and is not used as a forecast.
Consequences and economics
First-order consequence
A direct-business inference option is publicly offered.
Second-order consequence
A buyer can calculate how input mix changes the relative weight of output costs.
Third-order consequence
Longer outputs, retries and supervision can change what the same task costs.
Fourth-order consequence
Teams may compare how much they spend for each answer they can actually use, rather than the lowest posted rate.
Fifth-order consequence
Only observed costs and accepted results can support switching. No switching or savings were verified.
Observed costs and benefits
Illustration at listed rates: 1,000,000 ordinary input tokens cost ¥10.49 and 100,000 output tokens cost ¥21, for ¥31.49 in token charges before any tax treatment, fees or review. No cached input in this example. ¥210 divided by ¥10.49 is about 20.02; these are price ratios, not actual savings. No cross-currency comparison.
| Illustrative quantity | Charge in yen |
|---|---|
| 1 million ordinary input tokens | 10.49 |
| 100,000 output tokens | 21.00 |
| Token subtotal only | 31.49 |
- Fixstars statutory disclosure: rates and scope · 2026-11-20
Challenge and uncertainty
Disconfirming evidence
A workload with short outputs may not face the same price mix. The current service may also need fewer retries or less review. No matched data show which route costs less per usable answer.
What would confirm it
A matched comparison with actual approved aggregate counts and accepted outcomes shows a lower cost after accounting for errors, repeats, checking time and written rates.
What would disprove it
No matched accepted-output advantage remains after total costs, or the workloads, data scope or quoted terms are not comparable.
AI powered research