Azure Startup Credits for OpenAI GPT-6 Models: Costs Compared

Running Costs · Cloud & Infrastructure · ~18 min read

Microsoft will give an early-stage startup up to $150,000 in Azure credits, and those credits can be spent on OpenAI’s newest models. That turns a model choice that normally costs real money into one that costs, at least for a while, nothing out of pocket. We asked Puffin Notes editor Owen Calder how the credits actually work, which GPT-6 model to point them at, and where founders quietly waste them.

Q&A WITH OWEN CALDER · NOVEMBER 2026 · PRICES CHECKED THIS MONTH

$ cat summary.txt

$150K

Maximum Azure credits through Microsoft for Startups, unlocked in stages

$5,000

Unlocked after business verification — the amount most founders actually start with

80×

Output price gap between GPT-6.2 Astra ($40) and GPT-6 Luna ($0.50) per million tokens

$0

Cash cost of running GPT-6 models on Azure while credits last

// part one · how the money works

Credits are not cash. They behave like a budget with an expiry date.

Q01

Let’s start simple. What do startups actually get from Microsoft?

OC

Under the current Microsoft for Startups programme, you get approved, you get a starter amount of credit to build and test with, and then you unlock up to $5,000 once Microsoft has verified that your business is real. After that, more credits unlock as your Azure usage grows and stays steady, up to a ceiling of $150,000. Microsoft describes it as one programme that grows with you, and that’s a fair summary.

The part people skim past is the shape of it. The $150,000 is a ceiling, not a deposit. Most teams I talk to are living on the first few thousand for months, and the larger tranches only show up once you’re running a real workload on Azure. If you plan your first year around the headline figure, you’re planning around money you haven’t earned yet.

Q02

And those credits really cover OpenAI models?

OC

Yes — that’s the bit that makes the programme interesting in 2026. Microsoft lists Azure OpenAI alongside Foundry Tools and Azure Machine Learning as eligible AI services, next to the boring infrastructure you’d expect: virtual machines, storage, networking, App Service and Kubernetes. So the same credit balance that pays for your database can pay for GPT-6 tokens.

That’s a genuinely different deal from going direct. On OpenAI’s own platform you pay by card, by the token, from day one. On Azure, inside the credit window, the token bill is real but it’s deducted from credits, not from your bank account.

Q03

What happens when the credits run out?

OC

Nothing dramatic, which is exactly the danger. Microsoft is clear that your workloads keep running and you move to normal pay-as-you-go pricing. There’s no cliff, no shutdown, no migration. There’s also no moment that forces you to look at the bill. I’ve seen teams discover a five-figure invoice because the credits ran out mid-month and the model they’d been happily burning on was the most expensive one in the catalogue.

My advice: set a budget alert in Azure Cost Management at 50%, 75% and 90% of your credit balance on day one. It takes five minutes and it’s the cheapest insurance you’ll buy all year.

Table 1 — How Microsoft for Startups credits unlock
StageWhat triggers itCreditsWhat it realistically pays for
ApprovalAccepted into the programmestarter amountPrototyping, a dev VM, first API calls
VerifiedBusiness verificationup to $5,000An MVP with modest model traffic
GrowingMore Azure services, sustained usageadditional tranchesEarly production, real users
CeilingScaling on Azure over timeup to $150,000A serious production workload — for a while

Source: Microsoft for Startups programme FAQ, November 2026. Exact tranche sizes depend on eligibility and usage; check your portal.

// part two · the model menu

Four models, one family, very different bills

Q04

OpenAI’s lineup has changed a lot this year. What’s actually on the menu?

OC

The current generation is easy to read once you ignore the naming. There are four general-purpose models, and they’re named after a size ladder rather than a version number. GPT-6.2 Astra is the flagship for the hardest work. GPT-6.2 Sol is the everyday workhorse. GPT-6.1 Terra is the fast, efficient one with a strong line in visual reasoning and computer use. GPT-6 Luna is the cheap, high-volume option.

What surprised me is how much they share. All four have the same 1.05 million token context window, the same 128K maximum output, the same tool support — functions, web search, file search, computer use — and all of them do text and image input. The difference is how hard they think and what they charge for it.

Table 2 — The GPT-6 lineup at OpenAI’s direct prices (per 1M tokens)
ModelPositioningInputCached inputOutputKnowledge cutoff
GPT-6.2 AstraHardest projects, start to finish$8.00$0.40$40.0018 May 2026
GPT-6.2 SolEveryday work, agents at scale$2.00$0.10$10.001 Jun 2026
GPT-6.1 TerraFast; visual reasoning, computer use, writing$1.00—$5.001 Jun 2026
GPT-6 LunaFocused, high-volume tasks$0.10$0.01$0.5018 May 2026

Source: OpenAI API model catalogue, November 2026. All four: 1.05M context, 128K max output. Terra’s cached rate wasn’t in the catalogue view we used.

Q05

So when should someone actually reach for Astra?

OC

When a wrong answer costs more than the tokens. Long, multi-step engineering work where the model has to hold a plan for an hour. Legal or financial analysis where you’d otherwise pay a human to check the output. Agentic tasks with computer use where a mistake halfway through means starting over. Astra’s reasoning goes all the way up to a “max” setting, and on genuinely hard problems that extra thinking pays for itself.

When not to: anything you’d describe as “just”. Just summarise this. Just tag these tickets. Just rewrite the email. I see Astra used as a default far more than it should be, usually because someone picked it in a prototype and never revisited the choice. At $40 per million output tokens, defaults are expensive.

Q06

And Sol and Terra? They sound close.

OC

They are close, and that’s where most of the interesting decisions are. Sol is what I’d put behind most production features: a chat assistant, a coding helper, an agent that calls a handful of tools. It’s a quarter of Astra’s price and, in my experience, good enough for the large majority of real work.

Terra is half the price of Sol again, and it’s the one I’d test first for anything visual — reading screenshots, filling in forms through computer use, extracting data from scanned documents — and for writing that needs to sound natural rather than clever. It also has a “none” reasoning setting, which Astra and Sol don’t, so for quick interactive replies you can turn the thinking off entirely and keep latency low.

Q07

Luna at ten cents per million input tokens sounds almost free. What’s the catch?

OC

There isn’t much of one, if you use it for what it’s built for. Classification, routing, extraction into a fixed schema, moderation pre-filters, short summaries — tasks where the answer is narrow and you can check it automatically. Luna also gets much higher rate limits: 180 million tokens per minute at OpenAI’s top tier, against 40 million for Astra and Sol. For batch jobs, that matters as much as price.

Where it falls down is open-ended reasoning. Ask Luna to plan a migration or debug a race condition and you’ll get a confident, plausible answer that’s wrong in the third paragraph. The best architectures I see use Luna as the front door and only escalate to Sol or Astra when the task needs it.

Table 3 — When to use each model, and when not to
ModelUse it forAvoid it forReasoning settings
GPT-6.2 AstraLong agentic tasks, hard code, high-stakes analysisSummaries, tagging, chat replies, anything “just”low → max
GPT-6.2 SolProduction assistants, coding help, tool-calling agentsBulk classification where Luna is enoughlow → max
GPT-6.1 TerraScreenshots, forms, computer use, natural writing, low latencyDeep multi-step planningnone → max
GPT-6 LunaRouting, extraction, moderation, batch jobsOpen-ended reasoning, debugging, planningnone → max

Puffin Notes guidance, based on OpenAI’s positioning and what we see in practice. Test on your own data before committing.

// part three · openai direct vs azure

Same models. Very different invoice.

Q08

Is running these models on Azure cheaper than going to OpenAI directly?

OC

At list price, no — and occasionally it’s a bit dearer. Microsoft’s published Foundry prices for the GPT-6 family on its Global Standard tier are $2 in and $10 out per million tokens for Sol, and $0.10 and $0.50 for Luna. That’s the same as OpenAI. For the original GPT-6 Astra, Azure lists $10 and $50, which is higher than OpenAI’s current $8 and $40 for GPT-6.2 Astra. Pin a US Data Zone deployment for data residency and Microsoft adds a 10% premium on top.

But that’s the wrong comparison for a startup with credits. The right comparison is cash leaving your account. Direct, you pay every token. On Azure, inside the credit window, you pay zero. The list price only decides how fast the credits burn.

Table 4 — Price per 1M tokens: OpenAI direct vs Azure list vs Azure on credits
ModelOpenAI direct
in / out
Azure Global Standard
in / out
Azure US Data ZoneYour cash cost
on credits
Astra$8 / $40 (6.2)$10 / $50 (6)+10%$0
Sol$2 / $10$2 / $10+10%$0
Terra$1 / $5check catalogue+10%$0
Luna$0.10 / $0.50$0.10 / $0.50+10%$0

Azure prices are Microsoft’s published Global Standard short-context rates for the GPT-6 family (September 2026). Long-context requests above 272K input tokens cost more on both platforms. Model versions reach Foundry on their own schedule, so confirm which version — 6, 6.1 or 6.2 — is deployable in your region before you plan around it.

Q09

Put some numbers on it. What does a real workload cost?

OC

I modelled three workloads I see constantly, at OpenAI’s direct prices with no caching, so the comparison is clean. A support assistant handling 50,000 conversations a month, each about 3,000 tokens in and 500 out. A coding agent doing 2,000 tasks a month at 200,000 tokens in and 20,000 out, reasoning included. And a classification job pushing two million documents a month at 1,500 tokens in and 50 out.

Table 5 — Monthly cost of three workloads by model (OpenAI direct prices)
WorkloadAstraSolTerraLuna
Support assistant
50K chats · 3K in / 500 out
$2,200$550$275$27.50
Coding agent
2K tasks · 200K in / 20K out
$4,800$1,200$600$60
Document classifier
2M docs · 1.5K in / 50 out
$28,000$7,000$3,500$350

Our calculation. Reasoning tokens are billed as output, so real coding-agent costs vary widely with the reasoning setting.

Look at the classifier row. Running that job on Astra costs eighty times what it costs on Luna, and for a narrow labelling task you won’t get eighty times the quality. You probably won’t get twice the quality. That’s the single most common waste I see, and credits hide it, because nobody feels a $28,000 month that’s paid in credits.

Q10

How far does a credit balance actually stretch?

OC

Less far than the headline suggests, if you choose the wrong model. Here’s how many tokens $5,000 buys, assuming a typical mix of three input tokens for every output token.

One thing to keep in mind as you read it: the credit balance isn’t ring-fenced for AI. The same pot pays for your virtual machines, your managed database, your storage and your bandwidth. In the startups I’ve looked at, ordinary infrastructure takes somewhere between a third and a half of the monthly credit draw before a single token is generated. So whatever runway the model numbers suggest, assume the real figure is shorter, and check Cost Management by service rather than by total.

tokens bought with $5,000 of credit (3:1 input:output mix, OpenAI direct prices)

GPT-6.2 Astra
312M
GPT-6.2 Sol
1.25B
GPT-6.1 Terra
2.5B
GPT-6 Luna
25B

Our calculation: blended cost per million tokens is $16 for Astra, $4 for Sol, $2 for Terra and $0.20 for Luna. At the $150,000 ceiling, multiply by 30: about 9.4 billion Astra tokens, or 750 billion Luna tokens.

Table 6 — How many months the credits last for the support assistant
ModelMonthly bill$5,000 lasts$150,000 lasts
GPT-6.2 Astra$2,2002.3 months5.7 years
GPT-6.2 Sol$5509.1 months22.7 years
GPT-6.2 Sol, prompt cached$36013.9 months34.7 years
GPT-6 Luna$27.5015 yearseffectively forever

Our calculation, model token costs only — your VMs, storage and databases draw on the same balance, so real runway is shorter. Cached row assumes 2,000 of each 3,000 input tokens are a reused system prompt.

$ echo $LESSON

“Free tokens don’t make a bad model choice cheap. They make it invisible until the day they run out.”

— Owen Calder

// part four · where credits leak

The quiet ways a credit balance disappears

Q11

You mentioned reasoning tokens. Why do they matter so much?

OC

Because they’re billed as output, and output is the expensive side. When you turn Astra or Sol up to “high” or “max”, the model does its thinking in tokens you never see but absolutely pay for. On a hard coding task the hidden reasoning can be several times longer than the visible answer.

My rule of thumb: start every new feature at the lowest reasoning setting that passes your tests, then raise it only where the tests fail. Most teams do the opposite — they start at max because the demo looked better, and never come back down.

Q12

What about caching? Is it worth the effort?

OC

It’s the best discount in the price list. Cached input on Astra and Sol costs 5% of the normal input price — $0.40 instead of $8 on Astra, $0.10 instead of $2 on Sol. If your app sends the same long system prompt, the same tool definitions or the same reference document on every call, put that stable part at the start of the prompt so it can be cached.

There is a small catch: writing to the cache costs a little more than normal input — $10 instead of $8 per million on Astra. So caching pays off on prompts you reuse, not on one-offs. For the support assistant above, caching the system prompt cuts the Sol bill from $550 to $360 a month. That’s a third off for an afternoon of work.

Q13

Anything Azure-specific people get wrong?

OC

Three things. First, deployment type. Global Standard is the cheapest; Data Zone deployments keep processing inside the US or EU but cost about 10% more. If you don’t need residency, don’t pay for it with your credits. Second, long context. Both platforms charge more once input goes past 272,000 tokens — on Azure, Astra roughly doubles its input price. A 1.05 million token window is a capability, not a target.

Third, versions. OpenAI ships a new point release and the Azure catalogue follows on its own timetable, region by region. Don’t write “we’ll use 6.2 Sol on Azure” into a plan until you’ve actually deployed it in the region you need.

Table 7 — Credit leaks, ranked by how much they usually cost
LeakTypical overspendFix
Flagship model on a narrow taskup to 80×Route simple work to Luna or Terra
Reasoning left on max2–5× on outputStart low, raise only where tests fail
Long system prompt not cached~35% on chat workloadsPut stable content first; reuse it
Data Zone without a residency need+10%Use Global Standard
Prompts over 272K tokens~2× input priceRetrieve the relevant chunks instead
Idle infrastructure on the same balancevariesBudget alerts; delete what you don’t use

Puffin Notes estimates from the published price lists. Your mileage will vary with prompt shape and traffic.

// part five · strategy

Using the window well

Q14

If you were a founder with fresh credits tomorrow, what would you do?

OC

I’d treat the credit window as a research budget, not a subsidy. Use it to run the same real traffic through Astra, Sol, Terra and Luna, side by side, and measure which is the cheapest model that’s good enough for each task. That experiment would cost real money on OpenAI direct. On credits it’s free, and the answer is worth more than the credits themselves, because it sets your cost structure for the next five years.

Then I’d build routing in from day one: a cheap model classifies the request, and only the hard ones go up the ladder. Teams that leave this until after the credits run out have to do it in a panic, with real customers watching.

Q15

And what if a company simply can’t use all of them?

OC

It happens more than you’d think. A team pivots away from AI, moves to another cloud for a customer contract, or just builds something leaner than they expected. The balance sits there until it expires. Before you write that off, check your programme terms and what your options are. Got unused credits? There are marketplaces where you can sell your Azure startup credits rather than letting them lapse.

Whatever you decide, decide it before the expiry date, not after. Every credit programme I’ve seen is generous on the way in and completely unforgiving on the way out.

$ man azure-credits

Frequently asked questions

How much Azure credit can a startup get from Microsoft?

Up to $150,000 through Microsoft for Startups, unlocked over time. Founders start with a smaller amount and can unlock up to $5,000 after business verification; more becomes available with sustained Azure usage.

Which OpenAI models are current in November 2026?

GPT-6.2 Astra (flagship), GPT-6.2 Sol (everyday), GPT-6.1 Terra (fast, visual and computer use) and GPT-6 Luna (high-volume, low cost).

Is Azure OpenAI cheaper than OpenAI direct?

At list price, Sol and Luna cost the same. Azure’s GPT-6 Astra is listed higher than OpenAI’s GPT-6.2 Astra. With credits, your cash cost on Azure is zero until the balance runs out.

What is the cheapest GPT-6 model?

GPT-6 Luna, at $0.10 per million input tokens and $0.50 per million output tokens, with the highest rate limits.

Are reasoning tokens charged?

Yes, as output tokens. Higher reasoning settings can multiply the output bill, so start low and raise it only where needed.

What is the context window of the GPT-6 models?

1.05 million tokens for all four, with up to 128,000 output tokens. Requests above 272,000 input tokens are priced higher.

What happens when Azure credits run out?

Your workloads keep running and move to pay-as-you-go pricing. There is no shutdown, which is why budget alerts matter.

Can I use credits if I already use another cloud?

Yes. Microsoft says you can join while using another provider and move workloads to Azure gradually.

Can Azure startup credits be used for OpenAI models?

Yes. Microsoft lists Azure OpenAI among the eligible AI services, alongside Foundry Tools and Azure Machine Learning.

Is GPT-6.2 available on Azure?

Azure adds new OpenAI versions on its own timetable and region by region. Check the Microsoft Foundry model catalogue for your region before planning around a specific version.

What does GPT-6.2 Astra cost?

On OpenAI direct, $8 per million input tokens, $0.40 cached and $40 per million output tokens.

When should I use GPT-6.1 Terra instead of Sol?

For visual tasks, computer use, natural writing and low-latency replies. It is half Sol’s price and can run with reasoning switched off.

How much does prompt caching save?

Cached input costs 5% of the normal input price on Astra and Sol. On a chat workload with a long, reused system prompt, that can cut the total bill by around a third.

What is the Azure Data Zone premium?

Microsoft prices US Data Zone deployments at a 10% premium to Global Standard, in exchange for keeping processing within the zone.

Do Azure startup credits expire?

Credit grants typically carry an expiry date. Check the date shown in your Microsoft for Startups portal and plan usage around it.

Sources: Microsoft for Startups programme page and FAQ; OpenAI API model catalogue and per-model pricing pages; Microsoft Azure blog on GPT-6 Astra, Sol and Luna in Microsoft Foundry (September 2026). Workload and runway figures are our own illustrative calculations. Prices change often — check the current price lists before you commit to a budget. Puffin Notes has no commercial relationship with Microsoft or OpenAI.

Contact Us

We'd love to hear from you