Discover the real Azure identifiers + capabilities for Sol/Terra/Luna

[Kenny-Heitritter]

Question

Resolve the codenames Sol / Terra / Luna (OpenAI "GPT 5.6") into everything the downstream layers need. Deliver a markdown summary linked to this issue answering:

• Cardinality: are these 3 distinct models, or variants/tiers of one? (Sets deployment count everywhere downstream.)
• Azure identifiers per model: the real Azure OpenAI model_name + model_version (as consumed by azurerm_cognitive_deployment), and the deployment name to use.
• Region + capacity: availability in eastus2 (the region of qbraid-staging-ai-services / qbraid-prod), the sku_name (GlobalStandard?), and the max sku_capacity / quota ceiling per model — this sets the achievable "much higher" prod limit.
• Capabilities (for branding/qbraid/models.json): context window, max output, modalities, reasoning, tool_call, temperature.
• Pricing (for qbraid-api MODEL_PRICING): input / output / cache $ per 1M tokens.

Primarily research (Azure docs + the existing staging config as evidence). If confirming region/quota requires the live qBraid Azure subscription/portal, spin that portion out as a task.

Root unblocker: Azure deploy (staging + prod), gateway routing, billing, and the Codeq catalog all key off these facts.

[Kenny-Heitritter]

Sol / Terra / Luna (OpenAI GPT‑5.6) — Azure identifiers, capacity, capabilities & pricing

Ticket: Discover the real Azure identifiers + capabilities for Sol/Terra/Luna (map #31)
Type: research (AFK). One live‑only piece — the qBraid subscription's quota tier — is spun out as a task (see §7).
Confidence: High on identifiers/capabilities/pricing (Azure Foundry docs + OpenAI + LiteLLM agree). One field flagged for a deploy‑time smoke test: temperature.

---

1. Cardinality — 3 tiers of ONE generation, but 3 separate Azure deployments

Sol / Terra / Luna are three tiers of a single generation (GPT‑5.6), not three unrelated base models and not effort settings on one model. GPT‑5.6 moved away from "one model + adjustable effort" to three durably‑tuned tiers on the cost/speed/capability curve:

• Sol — flagship; most advanced reasoning + agentic workloads.
• Terra — balanced everyday tier; ~½ the cost of Sol, competitive with GPT‑5.5.
• Luna — fastest, cheapest tier.

Operational answer for downstream deployment count: on Azure each tier is an independent deployable model with its own model_name (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna). → 3 azurerm_cognitive_deployment resources, 3 gateway route entries, 3 models.json entries, 3 MODEL_PRICING rows. Conceptually one family; operationally three.

2. Azure identifiers per model (for azurerm_cognitive_deployment)

Deploy on the AIServices / AI Foundry account (qbraid-staging-ai-services), matching the existing gpt-5.x pattern in terraform/environments/staging/azure/. Recommend deployment name == model name (as the existing config does).

Deployment name	model.name (model_name)	model.version (model_version)	model.format
gpt-5.6-sol	gpt-5.6-sol	2026-07-09	OpenAI
gpt-5.6-terra	gpt-5.6-terra	2026-07-09	OpenAI
gpt-5.6-luna	gpt-5.6-luna	2026-07-09	OpenAI

All three GA on 2026‑07‑09; single stable version 2026-07-09.

Ready-to-paste ai_foundry_model_deployments entries (staging capacity per §3):

"gpt-5.6-sol"   = { model_name = "gpt-5.6-sol",   model_version = "2026-07-09", sku_capacity = 250 }
"gpt-5.6-terra" = { model_name = "gpt-5.6-terra", model_version = "2026-07-09", sku_capacity = 250 }
"gpt-5.6-luna"  = { model_name = "gpt-5.6-luna",  model_version = "2026-07-09", sku_capacity = 250 }

(model_format and sku_name default to OpenAI / GlobalStandard in the existing variable schema — no override needed.)

3. Region + capacity

• eastus2 availability: CONFIRMED from docs. All three (gpt-5.6-sol/terra/luna, v2026-07-09) are ✅ in eastus2 on Global Standard (and also Data Zone Standard, Standard/Regional, and Provisioned Managed). eastus2 is the region of both qbraid-staging-ai-services and qbraid-prod, so no region change is needed.

• sku_name: GlobalStandard — matches the existing staging convention and the map's decision.

• Capacity ↔ TPM mapping: for GlobalStandard Azure OpenAI, sku_capacity is in thousands of TPM (1 unit ≈ 1,000 TPM). The existing staging gpt-5.4 (sku_capacity = 1999 ≈ 1.999M TPM) and gpt-5.3-codex (3414 ≈ 3.414M TPM) confirm this.

• Quota ceiling is per region, per subscription, per model, and set by the subscription's auto‑assigned quota tier. GPT‑5.6 GlobalStandard ceilings (identical for all three tiers):

  Quota tier	RPM	TPM	sku_capacity equiv.
  Tier 1	1,000	1,000,000	1,000
  Tier 2	2,000	2,000,000	2,000
  Tier 3	4,000	4,000,000	4,000
  Tier 4	7,000	7,000,000	7,000
  Tier 5	10,000	10,000,000	10,000
  Tier 6	15,000	15,000,000	15,000

  Tier 5 and Tier 6 subscriptions have gpt‑5.6 quota by default; lower tiers may need a quota request to deploy at all. Beyond Tier 6, request more via the quota form or move to Provisioned Throughput Units (PTU). (Data Zone Standard ceilings are lower: 333K→5M TPM across tiers.)

• The "much higher" prod limit (map destination / Bootstrap prod Azure inference + deploy the 3 models at a high rate limit #34): the achievable ceiling = qBraid's prod subscription quota tier. Which tier qbraid-prod's subscription sits in is subscription state, only visible live → task in §7. Recommended prod target: Tier 5–6 (sku_capacity 10,000–15,000 = 10–15M TPM), filing a quota‑increase request if the subscription is below that.

4. Capabilities (for branding/qbraid/models.json) — identical across all three tiers

Field	Value
limit.context	1,050,000 (input 922,000 / output 128,000)
limit.output	128,000
modalities.input	["text", "image"] (text + image/vision processing)
modalities.output	["text"]
attachment	true (image input)
reasoning	true — reasoning effort none / low / medium / high / xhigh / max
tool_call	true — functions, parallel tool calling, structured outputs, computer use
temperature	false (recommended) — reasoning‑first family, controlled via reasoning effort, consistent with the GPT‑5 / o‑series API convention. ⚠️ Single field to smoke‑test against a live Azure completion when building #39.
APIs	Responses API + Chat Completions API
Knowledge cutoff	2026‑02‑16 (OpenAI). Note: Azure Foundry's "training data" column lists June 2026 for the 2026-07-09 snapshot — cosmetic only; doesn't affect deployment.
release_date	2026-07-09 (GA)
family	suggest gpt-5.6

5. Pricing ($ per 1M tokens)

GlobalStandard = OpenAI list price (no regional uplift; the +10% uplift applies only to azure/us/* / azure/eu/* regional deployments, which qBraid does not use). Standard/short‑context tier = input ≤ 272K tokens:

Model	Input	Cached input (read)	Cache write	Output
gpt-5.6-sol	$5.00	$0.50	$6.25	$30.00
gpt-5.6-terra	$2.50	$0.25	$3.125	$15.00
gpt-5.6-luna	$1.00	$0.10	$1.25	$6.00

• Cache read = 90% off uncached input; cache write = 1.25× uncached input.
• Long‑context tier (> 272K input tokens): input & output prices double. Sol $10 / $45, Terra $5 / $22.50, Luna $2 / $9 (cache read scales the same: Sol $1.00, Terra $0.50, Luna $0.20). Relevant because the context window is 922K input — large requests cross into the 2× tier.

6. Schema‑gap flags for the downstream billing/catalog tickets

• Add billing pricing + quota for the 3 models (qbraid-api) #38 (qbraid-api MODEL_PRICING) — current ModelPricing (src/features/billing/ai-chat/ai-chat.types.ts) has only inputPricePerMillionTokens / outputPricePerMillionTokens and provider: 'bedrock' | 'vertex'. To bill GPT‑5.6 correctly it must (a) add 'azure' to the provider union, and (b) decide whether to model cache pricing and the >272K long‑context tier — the current flat schema tracks neither, so cache discounts and long‑context uplift would be mis‑billed if ignored.
• Add Sol/Terra/Luna to the Codeq model catalog (branding/qbraid/models.json) #39 (models.json) — the cost block already supports cache_read / cache_write, and Gemini's entry shows the nested‑tier pattern (context_over_200k). Use an analogous context_over_272k block for GPT‑5.6's long‑context doubling. No gpt-5.x entry exists in models.json today, so these are new family: "gpt-5.6" entries.

7. Spun‑out task (live‑only)

Region availability is doc‑confirmed, so no task is needed for region. The only piece that genuinely requires the live qBraid Azure subscription/portal is the quota tier (which fixes the achievable prod ceiling for #34): confirm the qbraid-prod subscription's assigned quota tier for gpt‑5.6 GlobalStandard in eastus2, and file a quota‑increase request if it's below Tier 5–6. Created as a task ticket blocking #34.

---

Sources

• Azure Foundry — Models sold by Azure (GPT‑5.6 §, ids/versions/capabilities/cutoff): https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
• Azure Foundry — Region availability (eastus2 × GlobalStandard confirmation): https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
• Azure Foundry — Quotas & limits (per‑tier TPM ceilings): https://learn.microsoft.com/en-us/azure/foundry/openai/quotas-limits
• OpenAI — GPT‑5.6 (family, GA): https://openai.com/index/gpt-5-6/
• LiteLLM — Day‑0 GPT‑5.6 (pricing incl. cache + short/long tiers, azure/* names): https://docs.litellm.ai/blog/gpt_5_6
• Simon Willison — GPT‑5.6 family (reasoning effort levels, knowledge cutoff): https://simonwillison.net/2026/Jul/9/gpt-5-6/
• Independent pricing confirmation: https://www.aipricing.guru/openai-pricing/ , https://openrouter.ai/openai/gpt-5.6-terra
• Local evidence: qbraid-infrastructure terraform/environments/staging/azure/{main,variables}.tf (gpt‑5.x deployment pattern, GlobalStandard, sku_capacity↔TPM).