Build Not yet confirmed elsewhere 1 publisher 3 min read Published Aug 30, 2026, 8:52 AM PDT

Per-PTU throughput spans 25x across three models in the same GPT-5.6 family

Microsoft Foundry's sizing constants turn provisioned throughput into a division problem you can finish before you buy anything, which matters because the quota you commit does not travel across regions or deployment types.

The Engineer · Build desk

How we use AI Send a correction

Illustration accompanying Per-PTU throughput spans 25x across three models in the same GPT-5.6 family — Generated illustration

What happened

Microsoft Foundry's sizing table lists Input TPM per PTU of 30,000 for GPT-5.6 Luna, 3,000 for GPT-5.6 Terra and 1,200 for GPT-5.6 Sol, a model-specific constant rather than a measurement of your traffic.
Prompt caching works on provisioned deployments and cached input tokens consume no PTU capacity, which takes the walkthrough's 1,000 RPM example from 80 PTUs down to 60.
Deployments must clear a 15-PTU minimum and land on 5-PTU increments, so a calculated 62.4 PTUs is rounded up to 65 before anything can be deployed.
The pay-as-you-go comparison case is costed at $635.04 per day, $19,051.20 over a 30-day month and $231,789.60 over a 365-day year, all extrapolated from one repeating daily profile.

Why it matters

decision Throughput per PTU is model- and version-specific while the quota pool is not, so pinning a new version silently invalidates an earlier capacity estimate and the sizing sum has to be run again before the deployment is updated.
exposure A commitment placed in the wrong region or deployment type cannot be redirected, which leaves whoever signed it holding idle capacity in one geography while requests arrive in another.
cost The 15-PTU floor sets a minimum bill for anyone whose arithmetic lands below it, so the smallest provisioned deployments pay for headroom their own sizing sum says they will not use.
constraint The rate tables that would decide PayGo against PTU were pulled on a single date and are absent from the text supplied, so reproducing the comparison means fetching current rates from the pricing page and the Retail Prices API yourself.

Next story · Invest · 2d ago

Microsoft expects Maia 300 to become its primary AI inference system

Confirmed29 publishers

Start with the division. A 1,000 RPM peak at 1,200 tokens per prompt is 1,200,000 input tokens per minute ^[6]. Luna's listed constant is 30,000 input TPM per PTU, so the input side of that workload is 40 PTUs ^[5]^[20]. Run the identical workload against Terra and it is 400 PTUs; against Sol, 1,000 ^[20]. Only the divisor moved, and the divisor is a property of the model version pinned in the deployment ^[3].

The dev.to walkthrough's own answer for that example is 80 PTUs before caching ^[1]. Input accounts for half of it ^[21]. The other 40 has to come from the output-side constants in Foundry's sizing table, and that table did not survive into the text supplied to me ^[21]. The caching delta is a decent check on the reading: cached input tokens consume no PTU capacity, and the drop from 80 to 60 is 20 PTUs, exactly half the input-side 40 ^[1]^[18]. Half the input tokens served from cache produces that number.

Then the rounding grammar. Estimates below the 15-PTU minimum go up to 15, and everything else lands on a 5-PTU increment, which is why 62.4 becomes 65, a 4.2 percent overshoot on the computed figure ^[8]^[22].

The PayGo case is checkable too. Six four-hour blocks summing to 7,000 RPM give 1,680,000 requests a day at 240 minutes per block, 2.016B input tokens a day, 60.48B over 30 days ^[9]^[23]. That reconciles with the published split of 24.192B regular input, 30.24B cached and 6.048B written, and output lands at 10.08B ^[11]^[23]. Against $19,051.20 for the month, the blend is $0.27 per million tokens across all 70.56B ^[12]^[24]. That looks cheap until you price the peak: 2,500 RPM against a 1,166 RPM daily mean, a factor of 2.14 ^[9]^[25]. Provisioned capacity is sized to the peak while PayGo is billed on the mean, and that gap is the spillover argument, one the rates could settle if they had reached me: the figures were pulled on 27 August 2026 from the Azure OpenAI pricing page and the Retail Prices API ^[13], and neither source reached the text I have.

One unit mismatch is worth catching before anyone signs. The hourly PTU monthly equivalent assumes 720 hours ^[14], while the PayGo annual figure assumes 365 days ^[12]. Those are not the same year. The annual total is 12.1667 of those 30-day months, and twelve 720-hour months is 8,640 hours against the 8,760 in a 365-day year, so multiplying a monthly PTU equivalent by twelve drops 120 hours, about 1.4 percent ^[26].

For any of this to transfer, your workload has to look like the example: an average prompt near 1,200 tokens whose first 1,024 tokens are identical across requests, which is what a cache hit requires ^[7]; roughly half your input tokens actually reading from cache ^[10]; and your region and deployment type matching the ones priced. The last condition is the one you cannot fix afterwards, because quota is granted per subscription, region and deployment type, and East US quota stays East US quota; it will not convert into West Europe quota ^[4]. The quota pool applies the same way no matter which model you point it at ^[3].

What to watch

Whether the Sweden Central PTU rates and Global Standard token rates pulled on 27 August 2026 still hold, since the whole crossover rests on them.
Whether Foundry's published output-token sizing constants for Luna reproduce the 80-PTU figure from the input side plus output side.
The spillover mechanics named in the article's own title, which are absent from the text supplied here.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence47
Adoption: Insufficient
Hype gap+12
Incentives24
Confidence55

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Prompt caching is available for Provisioned Throughput deployments and cached input tokens do not consume PTU capacity; in the example, caching reduces the calculated requirement from 80 PTUs to 60 PTUs, a reduction of 20 PTUs or 25 percent.
Reported Supported Source: dev.to2 sources— create a free account to open them View cited source
[2]
Provisioned Throughput provides a dedicated, fixed amount of processing capacity exclusively for a model deployment; unlike Standard/PayGo it provides a model-specific latency SLA and its capacity is not shared across tenants.
Reported Supported Source: dev.to walkthrough of GPT-5.6 Luna PTU sizing on Microsoft FoundryView cited source
[3]
PTU quota is model-independent, so the same quota pool can be allocated across supported models, while throughput per PTU remains model- and version-specific.
Reported Supported Source: dev.toView cited source
[4]
PTU quota is granted per subscription, region and deployment type: quota in East US does not carry over to West Europe, and Global Provisioned quota does not carry over to Data Zone Provisioned.
Reported Supported Source: dev.toView cited source
[5]
Input TPM per PTU is a model-specific sizing constant, distinct from workload-calculated Input TPM; listed values include 30,000 for GPT-5.6 Luna, 3,000 for GPT-5.6 Terra and 1,200 for GPT-5.6 Sol.
Reported Supported Source: dev.to, citing the Microsoft Foundry PTU sizing tableView cited source
[6]
The worked provisioned example assumes a peak rate of 1,000 RPM, an average prompt of 1,200 tokens and an average response of 200 tokens, on gpt-5.6-luna with a Global Provisioned deployment.
Reported Supported Source: dev.toView cited source
[7]
The 1,200-token representative prompt exceeds the 1,024-token minimum for prompt caching, and a cache hit also requires at least the first 1,024 tokens to be identical across requests.
Reported Supported Source: dev.toView cited source
[8]
Provisioned deployments carry a 15-PTU minimum and a 5-PTU scale increment; rounding is required if an estimate falls below the minimum or between increments, so 62.4 PTUs would be rounded up to 65 PTUs.
Reported Supported Source: dev.toView cited source
[9]
The pay-as-you-go example assumes traffic fluctuating between 0 and 2,500 RPM over 24 hours in six four-hour blocks (0, 500, 1,000, 2,500, 2,000 and 1,000 RPM), with 1,200-token average input and 200-token average output on a Global Standard deployment, and the daily profile repeating for 30 days.
Reported Supported Source: dev.toView cited source
[10]
The pay-as-you-go example assumes 50 percent of input tokens are cache reads, 10 percentage points are cache writes and the remaining 40 percentage points are regular input; prompt caching applies only to input tokens, and output tokens are always charged at the regular output rate.
Reported Supported Source: dev.toView cited source
[11]
The example's 30-day token totals are 24.19200B regular input, 30.24000B cached input, 6.04800B cache writes and 10.08000B output.
Reported Supported Source: dev.toView cited source
[12]
The estimated pay-as-you-go cost is $635.04 per day, $19,051.20 per 30-day month and $231,789.60 per 365-day year.
Reported Supported Source: dev.toView cited source
[13]
Standard/PayGo rates for gpt-5.6-luna Global Standard were taken from the Azure OpenAI pricing page as of 27 August 2026, and Sweden Central PTU rates were retrieved on 27 August 2026 using the Azure Retail Prices API.
Reported Supported Source: dev.toView cited source
[14]
The hourly PTU monthly equivalent in the article assumes 720 hours, described as 30 days.
Reported Supported Source: dev.toView cited source
[15]
The article carries a warning that hourly, non-reserved PTU is best suited to temporary use; the sentence is cut off in the supplied text.
Reported Supported Source: dev.toView cited source
[16]
The article states that the same methodology applies to other models once their model-specific throughput and pricing values are substituted.
Reported Supported Source: dev.toView cited source
[17]
The article positions PTU as a good fit for predictable, sustained traffic with consistent latency and high-throughput requirements.
Reported Supported Source: dev.toView cited source
[18]
The 20-PTU caching saving equals half of the 40 input-side PTUs, consistent with half of input tokens being served as cache reads.
Derived Supported2 sources— create a free account to open them
[19]
Per-PTU input throughput spans a factor of 25 between the highest and lowest listed GPT-5.6 constants.
Derived Supported1 source— create a free account to open them
[20]
The example workload's input side is 1,200,000 input TPM, which is 40 PTUs on Luna, 400 PTUs on Terra and 1,000 PTUs on Sol.
Derived Supported1 source— create a free account to open them
[21]
Input accounts for 40 of the 80 PTUs in the article's pre-caching figure, leaving 40 PTUs attributable to output-side constants that do not appear in the supplied text.
Derived Supported1 source— create a free account to open them
[22]
Rounding 62.4 PTUs up to 65 buys 4.2 percent more capacity than the calculation asked for.
Derived Supported1 source— create a free account to open them
[23]
The traffic profile implies 1,680,000 requests per day and 2.016B input tokens per day, or 60.48B input tokens over 30 days, which matches the published regular, cached and cache-write totals; output of 10.08B also reconciles.
Derived Supported1 source— create a free account to open them
[24]
The pay-as-you-go month covers 70.56B tokens in total, a blended $0.27 per million tokens.
Derived Supported1 source— create a free account to open them
[25]
The example's peak of 2,500 RPM is 2.14 times its 1,166.7 RPM daily mean.
Derived Supported1 source— create a free account to open them
[26]
The 365-day PayGo year is 12.1667 of the article's 30-day months, so annualising a 720-hour monthly PTU equivalent by twelve covers 8,640 hours against 8,760 in the year, a shortfall of 120 hours or about 1.4 percent.
Derived Supported1 source— create a free account to open them

Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.

Add to ChatGPT Add to Claude

Sources

1 independent publisher whose own reporting we read for this story.

dev.to
1 article · August 30, 2026
GPT-5.6 Luna on Foundry: PTU Sizing, PayGo vs. PTU + Spillover Pricing

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

This is the rare capacity piece you can execute rather than admire. The constants do the work: 30,000 Input TPM per PTU on Luna, 3,000 on Terra, 1,200 on Sol, which means the same 1,000 RPM at 1,200-token prompts is 40 input-side PTUs on Luna and 1,000 on Sol — the model you pick moves the capacity bill 25x before your traffic changes at all. The practical grit is there too: caching only helps past 1,024 tokens and only with an identical prefix, and the 5-PTU increment quietly sells you 4.2 percent more than you asked for when 62.4 rounds to 65. Where it stops short is the sizing table itself. The 80-PTU pre-caching figure has 40 PTUs of output-side math behind constants dev.to describes but never prints, so you can rebuild half this calculation and must trust the rest.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence47: 0100
Self-consistent where it shows its work
The verifiable parts survive recalculation: the traffic curve, the 50/10/40 cache split and the four 30-day token totals all agree with each other to the last decimal, and both rate sources are named with a retrieval date. The unverifiable parts are structural rather than incidental — the Foundry per-PTU table and the Sweden Central rate card are cited but not reproduced, so the 40 output-side PTUs inside the 80-PTU headline exist only on the author's authority. One practitioner, no second reading of the same tables.
AdoptionInsufficient: A method, not a rollout
Nobody in this reporting deploys anything. There is no team running 60 PTUs of Luna, no reserved-capacity purchase, no throttling incident, no vendor disclosure of who is buying provisioned throughput — only a worked example with invented traffic and a same-day price lookup. Rate cards tell you what capacity costs, never who took it.
Hype gap+12: −100 understated0 aligned+100 overstated
Cent-level precision on invented traffic
The prose overpromises nothing — there is no best-in-class language anywhere — but the numbers wear more confidence than their inputs earn. $231,789.60 a year is quoted to the penny from a six-block traffic curve someone made up and a 50 percent cache-hit rate assumed rather than measured. The 720-hour month adds a smaller distortion in the same direction: twelve of those months are 8,640 hours against a real 8,760, so any PTU-versus-PayGo total assembled by multiplying by twelve runs about 1.4 percent light.
Incentives24: 0100
Community post with no visible sponsor
An individual publishing on dev.to, and it reads that way. The details a vendor deck skips are exactly the ones foregrounded here: quota that expires at the region boundary, a 5-PTU increment that overshoots your estimate, a hybrid baseline that ends up costing slightly more per day than plain pay-as-you-go. There is no disclosure in either direction, and the whole exercise does end with the reader on Azure's meter, but nothing in the text is doing sales work.
Confidence55: 0100
Trust the arithmetic, verify the tables
Confidence here is uneven by construction and that is the honest reading. Where the post shows its inputs, the outputs replicate and the assessment is firm. Where it points at a Microsoft table, we are one unverified transcription away from every conclusion, and a single publisher offers no way to close that gap. Anyone acting on the sizing should pull the Foundry constants and the current rate card themselves; the method will survive the check even if a number does not.

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Inference Cost ModelingFollow
Prompt Caching Follow
Azure AI OperationsFollow
LLM Capacity PlanningFollow