Azure OpenAI cost check: GPT-5.6 price cuts and PTU math

On 30 July 2026, OpenAI cut API prices for two of the three GPT-5.6 models: Luna dropped 80 percent, from $1.00 per million input tokens and $6.00 per million output tokens to $0.20 and $1.20, and Terra dropped 20 percent, from $2.50 and $15.00 to $2.00 and $12.00. Sol stayed at $5.00 and $30.00. Two days later, on 1 August, Microsoft confirmed on its Q&A forum that Azure OpenAI customers receive the same price decreases for Luna and Terra, effective 1 August. If you run GPT-5.6 workloads on Azure in Sweden or elsewhere in the EU, this is not just a pleasant line-item reduction. An 80 percent cut to your cheapest model moves break-even points across your whole deployment architecture: PTU reservations sized against old token prices, batch jobs routed to Luna because Terra looked expensive, and model-selection logic tuned to a price ratio that no longer exists. The first week of August is the right moment to rerun the math, before autumn budgets and any new reservation commitments lock in.

What changed, exactly

The GPT-5.6 family went generally available in Microsoft Foundry on 9 July 2026, in Global Standard and Global Priority Processing across all 28 global regions, plus Data Zones Standard and Global Provisioned from day one. Microsoft positions the three variants clearly: Sol for the most advanced reasoning, agentic workflows and code-focused scenarios; Terra as the balanced model for everyday work, competitive with GPT-5.5 at lower cost; Luna as the fastest and most affordable option for high-volume, latency-sensitive workloads.

Three weeks after GA, OpenAI announced the cuts. The new per-million-token prices for Global Standard deployments:

GPT-5.6 Luna: $0.20 input / $1.20 output (was $1.00 / $6.00). An 80 percent reduction.
GPT-5.6 Terra: $2.00 input / $12.00 output (was $2.50 / $15.00). A 20 percent reduction.
GPT-5.6 Sol: $5.00 input / $30.00 output. Unchanged. A separate Fast tier for Sol is priced at twice the standard rate.

OpenAI framed the cuts as passing through inference efficiency gains: better hardware routing, improved production inference software and smarter context caching. One detail stood out in the coverage: OpenAI credited Sol with autonomously rewriting production inference kernels, contributing roughly 20 percent of the serving-cost reduction. The competitive context matters too. VentureBeat noted the cuts landed within two weeks of aggressive low-cost releases from Google and Anthropic, and that the new Luna price moves OpenAI directly into the market's low-cost inference tier. Whatever the mix of motives, the direction for buyers is clear: frontier-adjacent inference keeps getting cheaper, and pricing assumptions have a short shelf life.

What Microsoft has confirmed for Azure

On the Microsoft Q&A thread asking whether Azure would follow, a Microsoft employee stated: "Effective August 1st, Azure OpenAI customers will receive the same updated price decreases for GPT 5.6 Luna and Terra models recently announced by OpenAI." That covers Global Standard, where Azure tracks OpenAI's direct pricing. Two practical caveats for the next few weeks:

Verify your meters. Pricing-page updates, the Azure Pricing Calculator and billing meters do not always move in lockstep on day one. Check your early-August usage in Microsoft Cost Management once invoices land, and open a support ticket if metered rates do not match the announcement.
Check the non-global tiers separately. Data Zone Standard and regional deployments carry their own price list at a premium over Global Standard. Microsoft's statement names Luna and Terra decreases; confirm on the official pricing page whether and when your specific deployment type in your EU data zone reflects them before you update internal chargeback rates.

The token math: what the cuts do to a real budget

Straight arithmetic on the published Global Standard prices shows the scale. Take a Luna workload processing 2 billion input tokens and 300 million output tokens per month, typical for a high-volume classification, extraction or summarisation pipeline:

Luna, 2B input + 300M output per month

Before:  2,000 x $1.00  +  300 x $6.00   = $2,000 + $1,800 = $3,800/month
After:   2,000 x $0.20  +  300 x $1.20   = $400   + $360   = $760/month

Saving: $3,040/month (~80%)

The same shape for Terra, at 500 million input and 100 million output tokens per month: $2,750 before, $2,200 after. Real money, but a different order of magnitude. The asymmetry is the point. Luna is now 10x cheaper than Terra on input and output alike, where the ratio was 2.5x on both before the cut. Any routing logic that sends "easy" requests to Luna and "hard" ones to Terra was tuned against the old ratio. With the new one, the threshold for escalating to Terra should move: every request you can keep on Luna is worth four times more in relative savings than it was in July.

PTU vs Global Standard: the break-even just moved

Provisioned throughput deployments bill per PTU per hour regardless of tokens consumed, with Azure Reservations offering discounted effective rates against 1-month or 1-year commitments. The economic case for PTU has always rested on utilisation: dedicated capacity beats pay-per-token only when you push enough tokens through it. The comparison is simple to state:

PTU wins on cost when:

  ptu_count x effective_hourly_rate x hours
      <
  input_tokens x standard_input_price + output_tokens x standard_output_price

Everything on the right side just got 80% smaller for Luna
and 20% smaller for Terra. The left side did not move.

When Standard token prices fall and PTU hourly rates stay put, the utilisation you need to justify provisioned capacity rises in direct proportion. For Luna that proportion is 5x: a Luna PTU deployment that broke even at a given traffic level in July now needs five times that traffic to beat Global Standard on cost. Very few Luna deployments will clear that bar. For Terra the bar rose 25 percent, which will flip marginal deployments but leave well-utilised ones standing. Sol PTU economics are untouched.

Cost is not the only reason to hold PTUs. Provisioned deployments buy you a defined latency target and guaranteed throughput, which Standard does not offer. Priority Processing sits in between: pay-per-token at a priority rate with a defined latency target per model and no long-term commitment. After this cut, a Luna workload that sat on PTU mainly for latency consistency deserves a fresh comparison against Priority Processing, where the pay-per-token side of the ledger is now dramatically cheaper.

Decision framework: what to do with each deployment

For every provisioned deployment serving GPT-5.6 Luna or Terra, work through this sequence:

1. Pull actual utilisation. Microsoft Cost Management shows reservation utilisation across deployments. Get real August token volumes, not sizing-time estimates.
2. Recompute break-even at new prices. Use the inequality above with the post-cut Standard rates. Remember cached input tokens: they do not consume PTU capacity, and prompt-caching discounts also reduce the Standard-side cost, so model your actual cache hit rate on both sides.
3. Below break-even and latency-tolerant? Move to Global Standard (or Data Zone Standard for EU residency). Deleting the deployment releases its capacity, so be sure before you scale down: capacity may not be available if you want it back.
4. Below break-even but latency-sensitive? Price out Priority Processing before renewing any reservation.
5. Above break-even? Keep the PTUs, and consider spillover to a Standard deployment for burst traffic instead of over-provisioning headroom you rarely use.
6. Reservation expiring soon? Do not auto-renew a 1-year term this month. Reservations are financial commitments against the hourly meter, loosely coupled to deployments. A 1-month reservation buys you time to see where prices and your routing mix settle.

The reservation point deserves emphasis. Twice in one summer the pricing floor has moved: GA in July, cuts in August. Committing to a 1-year PTU reservation sized against today's traffic and today's model split assumes both stay stable for twelve months. Recent history argues for shorter commitments and quarterly re-sizing, accepting a somewhat higher effective hourly rate as the price of optionality.

The Swedish and EU angle

Residency tiers now have a clearer price of admission

Swedish organisations rarely get to deploy Global Standard everywhere. Global deployments route inference across Azure regions worldwide, which many DPOs and information-security functions will not accept for personal data or sensitive workloads. Data Zone deployments keep processing within the EU geographic zone, and regional deployments pin it to a single region. GPT-5.6 launched with Data Zones Standard available from day one, so the EU-resident option exists. What changes with the price cut is the size of the premium you pay for residency: if Global Standard falls 80 percent for Luna and your EU Data Zone rate lags or falls by less, the relative cost of your residency requirement grows in percentage terms even as absolute costs shrink. Put the actual numbers for your deployment type in front of the people who own the residency decision. Cheaper global inference is not an argument to abandon residency requirements, but it is exactly the input a risk-based DPIA discussion needs: what does the constraint cost, and what does it buy.

Budget and procurement timing

Early August is when many Swedish organisations begin the autumn budget cycle after industrisemestern. Three concrete inputs for that process. First, restate your per-workload AI unit costs at the new prices before anyone anchors on spring numbers; a forecast built on July's Luna pricing overstates that line by 5x. Second, remember that Azure bills in USD-denominated meters for most Swedish agreements, so the SEK cost you present carries exchange-rate exposure on top of token prices; quote both. Third, for public-sector teams and anyone buying through a ramavtal or upphandling: falling and volatile per-token prices strengthen the case for consumption-based contract structures over fixed-capacity commitments, and for contract clauses that pass provider price reductions through to you rather than leaving them as supplier margin.

Your August cost-check list

Confirm the new rates apply to you. Check the Azure pricing page for your deployment types and regions, then verify August meters in Cost Management when invoices arrive.
Rerun PTU break-even for every Luna and Terra provisioned deployment using actual utilisation and post-cut Standard prices, cache rates included.
Revisit model-routing thresholds. Luna went from 2.5x to 10x cheaper than Terra; escalation logic tuned to the old ratio leaves savings on the table.
Pause long reservation commitments. Prefer 1-month reservations until your post-cut traffic mix stabilises, and never renew a 1-year term without redoing the math.
Re-price the residency premium. Compare Global Standard against your Data Zone or regional rates at the new prices and document the delta for your DPO and budget owners.
Update autumn forecasts in both USD and SEK, and flag consumption-based contract structures in any procurement now in flight.