Microsoft Foundry Pricing: Luna $0.10 vs Astra $10 [2026]

Azure AI Foundry pricing now means Microsoft Foundry pricing, and the price sheet moved fast this fall. Microsoft put GPT-6 Astra into general availability on September 3, 2026, added GPT-6 Sol and GPT-6 Luna on September 22, and listed GPT-6.1 Sol on September 29. The gap between the cheapest and the priciest model is 100x on both input and output tokens: Luna costs $0.10 per million input tokens, Astra costs $10. OpenAI’s own GPT-6 model guide lists Astra at $10 input, $50 output and $1 cached input per million tokens, and Microsoft’s Global Standard rate for Astra matches those figures.

That 100x spread is the headline, but it is not the whole bill. Microsoft sells the same models through Global Standard, Data Zone Standard, Provisioned Throughput, Batch and a Flex preview, and each route changes the math. If you want the three-cloud view first, our Bedrock, Foundry and Vertex AI side-by-side covers the platforms. This piece stays inside Foundry and prices every route.

All figures below are USD per one million tokens, taken from Microsoft’s Foundry announcements dated September 3 and September 22, 2026 (the second updated October 20) and retrieved on October 26, 2026. Treat them as list prices. Enterprise agreements, currency and region can move what lands on your invoice, and I flag every figure that rests on a single source.

·

Google · Preferred Sources

Don't miss new tech stories on Google

Add Tech Insider once in the Google app and our stories appear in your news suggestions.

Add Now

Azure AI Foundry Pricing: What Changed and What It Is Called Now

Microsoft’s current name for the platform is Microsoft Foundry. Older pages, forum threads and plenty of blog posts still say Azure AI Foundry, and some documentation paths keep the old wording or the Foundry classic label. The same goes for the OpenAI models: Microsoft’s current documentation reportedly describes them as Azure OpenAI in Microsoft Foundry, and no retirement of the Azure OpenAI capability has been announced. If you search for Azure AI Foundry pricing today, you land on the Microsoft Foundry price sheet.

Microsoft says Foundry is used by developers at more than 80,000 enterprises and digital natives, including 80% of the Fortune 500. That is a vendor claim, not an audited count, but it explains why the pricing page gets so much attention. A model catalog that big needs a pricing model that finance teams can read, and the fall 2026 changes made it both richer and harder to read.

Here is the timeline that matters for budgeting:

• July 9, 2026: GPT-5.6 (Sol, Terra and Luna) becomes generally available in Foundry. GPT-5.6 Sol lists at $5 input and $30 output in Global Standard for short context.
• September 1, 2026: Microsoft starts a promotional price for GPT-5.6 Sol of $4 input and $20 output, running through at least November 30, 2026.
• September 3, 2026: GPT-6 Astra reaches general availability with Standard and Provisioned Throughput deployments in Global and US Data Zone options.
• September 22, 2026: GPT-6 Sol and GPT-6 Luna arrive, and Microsoft publishes token prices for Global and Data Zone Standard.
• September 29, 2026: GPT-6.1 Sol appears in the catalog with a 2026-09-29 version date.
• October 16 and 17, 2026: Microsoft documents normalized-token accounting for sizing Provisioned Throughput on GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol.
• October 20, 2026: the GPT-6 pricing announcement is updated, and Microsoft says Foundry Local becomes generally available at Ignite (November 17 to 20, 2026).

One caution on naming. The catalog now holds both gpt-6-sol (dated September 22) and gpt-6.1-sol (dated September 29). Microsoft’s documentation reportedly treats GPT-6.1 Sol as the current generally available Sol release, so check which identifier your deployment points to before you compare invoices. The price table I could retrieve covers gpt-6-sol. OpenAI’s direct model guide lists GPT-6.1 Sol at the same $2 input and $10 output, which suggests parity, but I could not confirm a separate Foundry price for 6.1.

The Models on the Foundry Price Sheet: Astra, Sol, Luna and GPT-5.6

Microsoft positions the three GPT-6 models by job. Astra is the option for the most demanding reasoning. GPT-6.1 Sol is the default for production agents and complex workflows. Luna is the pick for high-volume tasks. GPT-5.6 stays in the catalog as the previous generation, and Microsoft has not announced a retirement date for it.

All three GPT-6 models share a 1,050,000-token context window with a 128,000-token maximum output. They support the Responses API, Chat Completions, streaming, structured outputs, tools, parallel tool calling, reasoning controls and text-and-image input with text output. Multi-agent orchestration is listed as a preview feature. Luna additionally lists the Decisions API.

Spec	GPT-6 Astra	GPT-6 Sol	GPT-6 Luna
Foundry release date	September 3, 2026	September 22, 2026 (6.1 Sol: September 29)	September 22, 2026
Model ID	gpt-6-astra	gpt-6-sol and gpt-6.1-sol	gpt-6-luna
Microsoft’s positioning	Most demanding reasoning	Production agents, complex workflows	High-volume tasks
Context window	1,050,000 tokens	1,050,000 tokens	1,050,000 tokens
Maximum output	128,000 tokens	128,000 tokens	128,000 tokens
Global input, short context	$10.00	$2.00	$0.10
Global output, short context	$50.00	$10.00	$0.50
Global input, long context	$20.00	$4.00	$0.20
Global output, long context	$75.00	$15.00	$0.75
Cached input, short context	$1.00	Sources differ	$0.01
Cached write, short context	$12.50	$2.50	$0.125
Data Zone US input, short context	$11.00	$2.20	$0.11
Input modalities	Text and image	Text and image	Text and image
Multi-agent orchestration	Preview	Preview	Preview
Deployment types confirmed	Standard and Provisioned Throughput	Global and Data Zone Standard	Global and Data Zone Standard

The Sources differ cell is deliberate. Two Microsoft-linked pages give different cached-input figures for Sol, $0.10 and $0.20 per million tokens, and I took the conservative route of leaving it out rather than picking one. Check the live price sheet before you model a Sol deployment that leans on caching.

For readers coming from the previous generation, our GPT-5.6 Sol, Terra and Luna price comparison shows the tiers GPT-6 replaces. The short version: GPT-6 Sol lists at $2 input and $10 output, against $5 and $30 for GPT-5.6 Sol at launch. That is 60% lower on input and about 67% lower on output. Against the $4 and $20 promotional price, the cut is a flat 50% on both sides.

Global Standard Pricing: Short Context vs Long Context

Global Standard is the default pay-per-token route. Microsoft may route your request across Azure regions, which is why it carries the lowest list price. The price category depends on the number of input tokens in the request: below Microsoft’s cutoff you pay the short-context rate, above it you pay the long-context rate. I did not find the exact cutoff in the sources I retrieved, so read it off the live table.

Model	Short input	Short output	Long input	Long output
GPT-6 Astra	$10.00	$50.00	$20.00	$75.00
GPT-6 Sol	$2.00	$10.00	$4.00	$15.00
GPT-6 Luna	$0.10	$0.50	$0.20	$0.75
GPT-5.6 Sol (July 9 list)	$5.00	$30.00	Not retrieved	Not retrieved
GPT-5.6 Sol (promo to at least Nov. 30)	$4.00	$20.00	Not retrieved	Not retrieved

Three patterns fall out of that table, and they hold for every GPT-6 model:

• Output costs 5x input in short context. Astra is $10 to $50, Sol is $2 to $10, Luna is $0.10 to $0.50.
• Crossing into long context doubles the input price and raises the output price by 50%. Astra’s input goes from $10 to $20 and its output from $50 to $75.
• The ladder between tiers is clean. Astra is 5x Sol and Sol is 20x Luna on both input and output, which makes Astra 100x Luna.

Those patterns matter because they let you estimate without a spreadsheet. A workload that sits in short context and writes as much as it reads is dominated by output cost, since each output token costs five input tokens. A retrieval-heavy workload that stuffs big documents into every request pays mostly for input, and it can tip into the long-context bracket without anyone noticing. Log the input token count per request for a week before you commit to a model.

The Luna-versus-Sol decision is the one most teams face. A single Luna request that reads 6,000 tokens and writes 800 is billed at $0.10 per million input tokens and $0.50 per million output tokens on Global Standard short context. The same request on Sol is billed at $2.00 per million input tokens and $10.00 per million output tokens. If Luna handles the task at acceptable quality, the savings are large. Our look at GPT-6 Luna against MiniMax and DeepSeek Flash pricing puts that number in a wider market context.

Data Zone Standard Pricing: What the 10% Residency Premium Costs

Data Zone Standard keeps processing inside a Microsoft-defined geography, with the United States and the European Union as the documented zones. Microsoft states that Data Zone Standard is priced at a 10% premium over Global Standard. The same 10% premium applies to Provisioned Throughput and Priority Processing, so the rule is easy to apply across deployment types.

Model (US Data Zone)	Short input	Short output	Long input	Long output
GPT-6 Astra	$11.00	$55.00	$22.00	$82.50
GPT-6 Sol	$2.20	$11.00	$4.40	$16.50
GPT-6 Luna	$0.11	$0.55	$0.22	$0.825

The Astra and Luna rows come straight from Microsoft’s table. The Sol row follows from the verified 10% rule and matches a separate Microsoft listing. EU Data Zone prices run higher than US Data Zone. One source reports Astra at $12 input and $60 output for short context in the EU zone, and Sol at $2.40 input and $12 output. Treat those EU figures as reported rather than confirmed, since I found them in a single listing.

Is the premium worth paying? The question is really about your legal exposure, not your token math. Data Zone Standard costs more than Global Standard, so the surcharge scales with your token bill. A privacy review that blocks a launch for a quarter costs far more than that. For teams subject to data residency rules, the premium is cheap insurance. For teams with no residency requirement, it is a tax on every token, and Global Standard is the better default.

One more point on regions. Microsoft reportedly lists GPT-6 Astra and GPT-6 Luna for Global Standard in all Azure regions where Global Standard is available, so the global route should not narrow your region choice. Data Zone availability is documented separately, and it was the later addition: Astra launched in September with Global and US options, and the EU rows showed up in the later pricing table.

Prompt Caching: The 10% Read Price and the 125% Write Price

Caching is where Foundry’s price sheet rewards careful engineering. Look at Astra’s short-context row: $10 for fresh input, $1 for cached input, $12.50 for cached writes, $50 for output. Cached reads cost 10% of fresh input. Cached writes cost 125% of fresh input. Luna follows the same shape at $0.10, $0.01 and $0.125, and GPT-5.6 Sol’s launch table did too, at $5.00, $0.50 and $6.25.

That structure sets a simple break-even. You pay a 25% surcharge the first time a prefix enters the cache, then 90% off every time it is read. The cache pays off by the second send: a write plus one read costs 1.35 units against 2.0 for two fresh sends. A prefix sent three times costs 1.45 units cached against 3.0 fresh, and one reused ten times costs about 2.15 against 10.

Real production traffic shows how much this matters. OpenAI reports that GitHub saw the share of prompt tokens requiring fresh processing fall by more than 50% across billions of requests to OpenAI models over several months, which helped Copilot respond faster. That is OpenAI’s account of a customer result, not a Microsoft benchmark, but the mechanism is the one Foundry prices.

Here is what caching does to a real bill. Take the support-assistant scenario I use later: 300 million input tokens and 40 million output tokens a month on Astra. Astra’s Global Standard short-context rates are $10.00 per million input tokens, $1.00 per million cached input tokens and $50.00 per million output tokens. If 80% of the input tokens are cached reads, that share is billed at the cached rate instead of the full input rate, so the input line shrinks sharply. That ignores cache-write fees, which add a bit back. The output side is untouched.

Three design rules follow from the numbers:

• Put stable content first. System prompts, tool definitions, policy text and few-shot examples belong at the front of the request, because caching works on repeated prefixes.
• Keep dynamic content last. A timestamp or a user name at the top of the prompt breaks the prefix and forces a fresh write on every call.
• Watch output tokens. Caching does nothing for the 5x output multiplier, so cap response length and use structured outputs where you can.

Deployment Types Compared: Global, Data Zone, Provisioned, Batch and Flex

Microsoft’s documentation describes three broad deployment categories in Foundry: Standard, Provisioned and Batch. Flex processing sits alongside them as a preview option. The table below compares them on the things that change a budget.

Deployment type	Billing basis	Price relative to Global Standard	Best for	Status
Global Standard	Pay per token, global routing	Baseline	General workloads	Generally available
Data Zone Standard	Pay per token, routed inside a US or EU zone	+10%	Residency requirements	Generally available
Global Provisioned Throughput	Billed per provisioned throughput unit (PTU)	Varies by deployment	Steady, latency-sensitive traffic	Generally available
Data Zone Provisioned Throughput	Billed per PTU, routed inside a zone	+10% over Global PTU	Steady traffic with residency rules	Generally available
Batch	Asynchronous processing	Discounted	Offline jobs	Documented, GPT-6 prices not confirmed
Flex processing	Pay per token, capacity not guaranteed	50% of Standard rates	Interruptible work	Preview

A few notes on that table. Microsoft documents Flex input and output tokens at half the corresponding Standard rates, and reportedly does not bill requests that get rejected because capacity is unavailable. The Flex documentation reportedly lists GPT-6 Sol, Astra and Luna under Global Standard. Because it is a preview, I would not put a customer-facing path on Flex, but a nightly enrichment job is a good fit.

Batch is where I hit a gap. Microsoft’s documentation reportedly describes Global Batch at a 50% discount against Global Standard with a 24-hour target, and Data Zone Batch as a similar discounted route. The pricing material I retrieved does not confirm a GPT-6 Astra batch price or availability date. Before you budget a batch run on any GPT-6 model, check the live table. If Batch is available, it halves the bill for anything that can wait a day.

Priority Processing shows up in the Microsoft footnotes with the same 10% Data Zone premium, but the pricing excerpts I retrieved did not confirm a Foundry Priority price for GPT-6 Astra. I would not plan around it until Microsoft publishes one. For readers who want to see how another cloud structures the same idea, the Amazon Bedrock pricing tiers for Flex, Priority and Batch follow a similar three-way split with different multipliers.

Provisioned Throughput vs Pay-As-You-Go: When PTUs Win

Provisioned Throughput reserves dedicated model-processing capacity, billed hourly per PTU regardless of how many tokens you send. Microsoft’s Provisioned Throughput documentation describes it as the choice for predictable throughput and low latency. GPT-6 Astra is confirmed for it. Microsoft says its pricing varies by deployment type and does not publish a universal per-token figure, so you need a quote or the live calculator to price it.

The decision comes down to utilization. Pay-as-you-go charges nothing when you are idle and charges list price per token when you are busy. A PTU deployment charges every hour whether you send one request or a million. The crossover is the utilization level at which your hourly PTU cost equals what the same tokens would cost on Standard. Above that level, PTUs are cheaper per token, and you also get the latency guarantee. Below it, you are paying for empty capacity.

How to size a PTU deployment

Microsoft documented a new sizing approach on October 16 and 17, 2026. For GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol, PTU sizing now uses normalized tokens, which Microsoft reportedly designed to align PTU capacity use with Global Standard pay-as-you-go pricing. In plain terms, a token that costs more on Standard should also draw down more PTU capacity. If that holds, your Standard price sheet doubles as a rough capacity model, with output tokens weighing more than input tokens and cached reads weighing less. Confirm the exact weights in Microsoft’s sizing page before you buy.

A practical sizing routine looks like this:

1. Export two weeks of request logs with input tokens, cached tokens and output tokens per call.
2. Weight them with the Standard price ratios as a rough proxy, using 1 for fresh input, 0.1 for cached input and 5 for output on a GPT-6 model.
3. Find the peak hour and the average hour, since PTUs are bought for the peak and billed for every hour.
4. Divide average load by peak load. If the ratio sits below roughly one half, pay-as-you-go will almost always win on cost.
5. Price the PTU quote against the Standard cost of the same hourly tokens and compare.

A hybrid is common in practice. Teams provision enough PTUs for the steady baseline, then let overflow requests spill to Global Standard. That caps the idle-capacity cost and keeps latency stable for the traffic that matters. If you also rent GPUs instead of buying tokens, our AWS, Azure and GCP GPU price gap piece covers that side of the build-versus-buy line.

Cost Scenarios: What a Month of Tokens Actually Costs

List prices only mean something against a workload. The table below prices four scenarios on Global Standard. I computed each figure from the short-context rates (the long-context scenario uses long-context rates). These are my own illustrations, not Microsoft figures, and they exclude cache-write fees, Provisioned Throughput and taxes.

Scenario	Monthly tokens (input / output)	GPT-6 Astra	GPT-6 Sol	GPT-6 Luna
Light internal tool	100M / 20M	Not listed	$400	$20
Customer support assistant	300M / 40M	$5,000	$1,000	$50
Long-context document review	10M / 1M	$275	$55	$2.75
Agent fleet, heavy coding	500M / 100M	$10,000	Not listed	$100

Add 10% to any cell for Data Zone Standard. For comparison, a Microsoft Tech Community article reportedly lists GPT-5.6 Sol at $4.00 per million input tokens and $20.00 per million output tokens on Global Standard, while GPT-6 Sol is listed at $2.00 and $10.00. The same 100M-in, 20M-out workload is therefore cheaper on GPT-6 Sol.

A cost estimator you can run

This short script reproduces the table and lets you plug in your own volumes. Prices are the Global Standard list prices above.

PRICES = {  # USD per 1M tokens: (short_in, short_out, long_in, long_out)
    "gpt-6-astra": (10.00, 50.00, 20.00, 75.00),
    "gpt-6-sol":   (2.00, 10.00, 4.00, 15.00),
    "gpt-6-luna":  (0.10, 0.50, 0.20, 0.75),
}
CACHED_READ = {"gpt-6-astra": 1.00, "gpt-6-luna": 0.01}
DATA_ZONE = 1.10

def monthly_cost(model, in_m, out_m, long_ctx=False, data_zone=False, cached_share=0.0):
    s_in, s_out, l_in, l_out = PRICES[model]
    p_in, p_out = (l_in, l_out) if long_ctx else (s_in, s_out)
    fresh = in_m * (1 - cached_share) * p_in
    cached = in_m * cached_share * CACHED_READ.get(model, p_in)
    total = fresh + cached + out_m * p_out
    return round(total * (DATA_ZONE if data_zone else 1.0), 2)

print(monthly_cost("gpt-6-astra", 300, 40))                      # 5000.0
print(monthly_cost("gpt-6-astra", 300, 40, cached_share=0.8))    # 2840.0
print(monthly_cost("gpt-6-luna", 100, 20, data_zone=True))       # 22.0

What the scenarios say

Luna’s price is so low that model cost stops being the main line item. At $0.10 per million input tokens and $0.50 per million output tokens on Global Standard short context, engineering time, monitoring and evaluation can cost more than inference for customer support traffic. At the other end, an Astra-powered agent fleet at $10,000 a month deserves a hard look. If a Sol or a Luna run can complete 80% of tasks, routing only the hard 20% to Astra cuts that bill sharply. The router pays for itself fast.

Benchmarks and Real-World Examples

Benchmarks for these models come mostly from the vendors, so read them with that in mind. I used three kinds of source here: OpenAI’s own published comparisons, a Microsoft engineering write-up on fine-tuning, and customer results reported by OpenAI and Microsoft.

On OpenAI’s agentic evaluation Agents’ Last Exam, which tests complex professional workflows, GPT-6 Sol at max effort scores 56.4%, and OpenAI says it beats Claude Opus 5’s highest score on that test at 60% lower cost per task. OpenAI also says GPT-6 Sol exceeds Claude Fable 5.1 at far lower cost and can match Fable 5.1 at xhigh effort on FrontierCode, a benchmark that grades coding agents on whether their changes are ready to merge. For Luna, OpenAI reports a max-effort score of 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task respectively. These are vendor-run, so treat them as directional. Our GPT-6.1 Sol, Sonnet 5.5 and Gemini 4 Argon comparison offers a cross-vendor read.

The second source is Microsoft’s own Foundry engineering blog. Its information-extraction study evaluated base and fine-tuned versions of two open-weight models on the same 20 documents. The gpt-oss-20b model rose from an F1 of 0.4379 to 0.8115, an 85.3% relative gain after 100 GRPO steps. The larger gpt-oss-120b rose from 0.6816 to 0.8573 and finished with the higher score. The post lists GPT-5.5, GPT-5.6 Sol and GPT-6 Sol as frontier-model references against Azure inference prices. The lesson for pricing: a small tuned model can close much of the gap on a narrow task at a fraction of frontier token costs.

The third source is customer and platform evidence. Here are the real-world examples I would weigh:

1. GitHub Copilot and caching. OpenAI reports that fresh-processing prompt tokens fell by more than 50% across billions of requests, which maps directly to the cached-read discount on Foundry’s sheet.
2. Microsoft’s extraction study. The 20-document evaluation shows tuned gpt-oss-20b climbing to 0.8115 F1, evidence that a cheap model can handle a bounded job.
3. Foundry’s enterprise footprint. Microsoft says more than 80,000 enterprises and digital natives, including 80% of the Fortune 500, use Foundry. It is a vendor figure, but it signals that these price tiers see heavy production use.
4. Voice workloads. Microsoft’s Voice Live API documentation reportedly lists GPT-5.6 Luna as a Standard-tier model with speech input and output, a sign that the cheap tier is being used for latency-sensitive voice agents.
5. Multi-cloud distribution. Gadgets 360 reported on October 22 that Google Cloud added GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to Model Garden in private preview, which shows how quickly these same models can appear on a competing cloud.
6. ChatGPT model allocation. OpenAI’s release notes say Free and Go users run on GPT-6 Luna, while Plus, Pro, Business and Enterprise users run on GPT-6 Sol. OpenAI itself splits its consumer traffic along the same cheap-versus-middle line that Foundry prices.

Taken together, the evidence says the same thing from three angles. The cheap tier is production-ready for high-volume work, the middle tier is the agent default, and the top tier earns its price only on the hardest tasks.

How Foundry Compares With Bedrock and Google Cloud

Foundry is not the only place to buy these models. Amazon Bedrock and Google Cloud both sell frontier models with tiered pricing, and the structures differ in ways that matter for a procurement decision. I keep this section qualitative because the figures I verified today concern Foundry.

Foundry’s pricing identity is the Global versus Data Zone split, plus a separate Provisioned Throughput track. Bedrock leans on named tiers such as Flex and Priority plus a Batch option. Google’s Gemini platform, which Google materials reportedly describe as the former Vertex AI, also splits pay-as-you-go into Standard, Flex and Priority tiers alongside Provisioned Throughput. Our deep dive on AgentCore, Foundry and Vertex AI Agent Builder compares the agent layers, which is where most of the long-run spend lands.

The distribution picture is also shifting. OpenAI’s September 3 release of GPT-6 Astra reportedly reached ChatGPT, the OpenAI API, Microsoft Azure and AWS Bedrock. On October 21, Google Cloud reportedly opened a private preview of Astra, GPT-6.1 Sol and Luna for selected enterprise customers, and a Yahoo Finance report on OpenAI models landing on Google Cloud framed it as frontier distribution going multi-cloud. For buyers, that means Microsoft no longer holds a pricing monopoly on OpenAI models, and commitments can be negotiated across clouds. Google’s own Gemini 4 Argon has an introductory price that changes on November 30, which our note on the Gemini Argon price doubling tracks.

What developers actually run on

Cloud choice follows developer habits, and the Stack Overflow 2025 Developer Survey is the best public measure of them. Its technology section reports cloud platform use among professional developers:

Stack Overflow, 2025 Developer Survey, Technology section

Azure’s edge for Foundry is that most enterprise Azure customers already have a Microsoft billing relationship, and Foundry’s models sold directly by Azure can bill through an Azure subscription, carry Azure service-level agreements and get Microsoft support. For a company with an Azure commitment, that can mean AI spend that draws down existing commitments instead of opening a new vendor contract. Treat that as a base case, not a verdict, and check your own agreement.

Best Foundry Deployment for Each Use Case

No single deployment wins everywhere. Match the route to the workload shape and the compliance exposure. The table gives the short answer, and the list below explains the logic.

Use case	Recommended model	Recommended deployment	Why
High-volume chat, tagging, routing	GPT-6 Luna	Global Standard	Lowest per-token cost at $0.10 in and $0.50 out
Production coding and workflow agents	GPT-6.1 Sol (or gpt-6-sol)	Global Standard, or Data Zone if regulated	Microsoft’s stated default for agents, 5x cheaper than Astra
Hardest reasoning, low volume	GPT-6 Astra	Global Standard with heavy caching	Top capability, caching cuts input by 90%
EU or US residency requirements	Any GPT-6 model	Data Zone Standard	10% premium buys in-zone processing
Overnight enrichment and backfills	Luna or Sol	Batch or Flex preview	50% off, if capacity and availability fit
Steady real-time traffic with latency targets	Astra, Sol	Provisioned Throughput with Standard overflow	Reserved capacity and predictable latency

Here is the reasoning in prose for the five cases that come up most.

1. Chat and classification at scale. Start with Luna and prove it fails before you pay for more. At $0.10 per million input tokens, you can run an evaluation set of 10,000 prompts for a few cents of input cost.
2. Agents that call tools. Sol is the sensible middle. Microsoft calls GPT-6.1 Sol the default for production agents, and its 5x price gap against Astra gives you room to retry failed steps. Our Microsoft Foundry agents walkthrough shows the setup.
3. Regulated data. Pick Data Zone from day one. Switching later means redeploying and re-testing, and the 10% premium is small next to a compliance rework.
4. Seasonal or bursty traffic. Stay on pay-as-you-go. Provisioned capacity bills every hour, so a spiky workload leaves you paying for idle PTUs between peaks.
5. Fine-tunable narrow tasks. Consider a tuned open-weight model. Microsoft’s extraction study took gpt-oss-20b to an F1 of 0.8115 and gpt-oss-120b to 0.8573 on 20 documents, which may beat a frontier model’s cost-per-correct-answer on a repeated, bounded task.

Migration Guide: Moving From GPT-5.6 or Direct OpenAI to GPT-6 on Foundry

Two migrations come up most often. The first is a team on GPT-5.6 in Foundry that wants GPT-6 pricing. The second is a team calling OpenAI directly that wants Azure billing and service-level agreements. Both follow the same shape.

1. Inventory current spend. Pull a month of usage by deployment and model. Note input, cached and output tokens, because the GPT-6 price ratios differ from older models.
2. Map old models to new ones. A rough mapping is GPT-5.6 Sol to GPT-6 Sol (or 6.1 Sol), GPT-5.6 Luna to GPT-6 Luna, and your hardest GPT-5.6 Sol tasks to Astra. Price drops on the Sol and Luna lines make the first two moves cheap.
3. Check region and quota. Confirm Global Standard or Data Zone quota in the region you use. Quota documentation lists the GPT-6 models under these tiers, and Tier 5 and Tier 6 subscriptions reportedly get quota by default.
4. Deploy in parallel. Create the new deployment next to the old one with its own name. Do not overwrite the live deployment.
5. Replay real traffic. Run a sample of logged prompts through both and compare quality, latency and token counts. GPT-6 models may use reasoning tokens differently from GPT-5.6, so measure actual output length.
6. Restructure prompts for caching. Move stable content to the front and dynamic content to the back so cached reads kick in at the 10% price.
7. Route gradually. Shift 5%, then 25%, then 100% of traffic. Keep the old deployment live until error rates settle.
8. Re-run the cost model. Compare against the first step’s baseline, and decide on Data Zone, Flex or PTU only after you have real GPT-6 numbers.
9. Retire the old deployment. GPT-5.6 has no announced retirement date, so you can move on your schedule. Note that the GPT-5.6 Sol promotional price runs through at least November 30, 2026, so there is no price pressure to rush before then.

For the direct-OpenAI move, the main change is billing and support. Foundry models sold directly by Azure bill through an Azure subscription, come under Azure service-level agreements and get Microsoft support. Astra’s Global Standard price ($10 and $50) matches OpenAI’s listed API price, so the move can be price-neutral on Global and costs 10% more only if you choose Data Zone. Our GPT-6.1 Sol API setup guide covers the direct-API side if you need to compare endpoints.

Here is a deployment command for a Global Standard Luna deployment with the Azure CLI. Flags change between CLI versions, so run the command with --help first to confirm them.

az cognitiveservices account deployment create \
  --name my-foundry-resource \
  --resource-group my-rg \
  --deployment-name luna-prod \
  --model-format OpenAI \
  --model-name gpt-6-luna \
  --model-version 2026-09-22 \
  --sku-name GlobalStandard \
  --sku-capacity 100

Swap GlobalStandard for DataZoneStandard when residency rules apply. The model version string matches the September 22 catalog date for gpt-6-luna.

Pros and Cons of Foundry’s Pricing Model

Foundry’s strengths and weaknesses follow from how Microsoft built the sheet. Here is the honest accounting.

Pros

• The model ladder is clean. Each GPT-6 tier is a fixed multiple of the next, so you can estimate cost from token counts alone.
• Cached reads cost 10% of fresh input, which rewards well-structured prompts.
• Luna’s $0.10 and $0.50 pricing makes high-volume workloads close to free at the model layer.
• Data Zone adds only 10%, a small price for residency.
• Billing, service-level agreements and support sit inside Azure, which suits companies with existing Microsoft commitments.
• The 1,050,000-token context window is available on every GPT-6 tier.

Cons

• Output tokens cost 5x input in short context, which punishes chatty responses.
• Long context doubles input price, and the cutoff is easy to cross by accident.
• Provisioned Throughput has no universal public per-token price, so you need a quote.
• Batch and Priority prices for GPT-6 were not confirmed in the material I retrieved.
• Naming churn (Azure AI Foundry, Microsoft Foundry, Azure OpenAI in Foundry, gpt-6-sol versus gpt-6.1-sol) makes it easy to compare the wrong rows.
• Flex is still a preview, so it is not yet fit for customer-facing paths.

Verdict: Which Foundry Model and Deployment to Pick

Start with Luna on Global Standard for any workload you have not yet proven needs more. At $0.10 input and $0.50 output, it prices tokens at 1% of Astra’s rate. Move up to Sol, which Microsoft positions as the default for production agents, when Luna fails your evaluation set. Reserve Astra, at $10 and $50, for the hardest reasoning tasks where a single wrong answer costs more than the token bill, and pair it with aggressive caching, because cached reads cost $1 per million.

On deployment, Global Standard should be your default and Data Zone your exception. The 10% premium is only worth it when a legal or contractual rule requires in-zone processing. Provisioned Throughput earns its keep only when traffic is steady enough to keep PTUs busy, and the October 16 and 17 normalized-token change makes that sizing easier to do. Keep Batch and Flex for offline work, and verify the GPT-6 batch price before you rely on it.

The numbers back a clear call. A 300M-input, 40M-output support workload costs $50 on Luna, $1,000 on Sol and $5,000 on Astra, and caching can take the Astra figure down to $2,840. Pick the cheapest model that passes your evals, cache everything stable, and keep the premium tier for the narrow slice of requests that truly need it. For the wider picture across providers, browse our cloud computing coverage, and recheck the live Microsoft price sheet before committing, since the GPT-5.6 Sol promotion ends on November 30 at the earliest and Ignite runs November 17 to 20.

FAQ

What is Azure AI Foundry called now?

Microsoft’s current name is Microsoft Foundry. Older documentation and many third-party posts still say Azure AI Foundry, and OpenAI model access is described as Azure OpenAI in Microsoft Foundry. The price sheet you want is the Foundry Models pricing page.

How much does GPT-6 Astra cost on Foundry?

On Global Standard with short context, Astra costs $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens. Long context costs $20 input and $75 output. US Data Zone Standard adds 10%, taking short-context rates to $11 and $55.

Which Foundry model is the cheapest?

GPT-6 Luna. It lists at $0.10 per million input tokens and $0.50 per million output tokens in short-context Global Standard. That is 20x cheaper than GPT-6 Sol and 100x cheaper than GPT-6 Astra, which makes it the natural starting point for high-volume work.

How much extra does Data Zone cost?

Microsoft prices Data Zone Standard at a 10% premium over Global Standard, so Astra’s short-context rates go from $10 and $50 to $11 and $55. The same 10% premium applies to Provisioned Throughput and Priority Processing. EU Data Zone prices run higher than US ones.

Does Foundry offer a batch discount?

Microsoft lists Batch as one of three deployment categories, alongside Standard and Provisioned. Flex processing, a preview, prices at 50% of Standard rates. I could not confirm a GPT-6 batch price in the material I retrieved, so verify the live table before you budget a batch run.

Is GPT-5.6 retired on Foundry?

No retirement has been announced. GPT-5.6 Sol, Terra and Luna remain in the catalog, and the GPT-5.6 Sol promotional price of $4 input and $20 output runs from September 1 through at least November 30, 2026. GPT-6 is now the newest family.

Does the Data Zone premium apply to Provisioned Throughput?

Yes. Microsoft states that US Data Zone Provisioned Throughput carries a 10% premium over Global Provisioned Throughput. Microsoft does not publish a single per-token price for Provisioned Throughput, because pricing varies by deployment type.

When does Provisioned Throughput beat pay-as-you-go?

It wins when your traffic keeps the provisioned capacity busy most of the time. Because PTUs bill for reserved capacity rather than tokens, spiky workloads waste money. Size against peak load using Microsoft’s normalized-token guidance, and let overflow requests fall back to Global Standard.