GPT-6 Astra becomes generally available on Microsoft Foundry

Timeline of the Astra availability arc (all dates verified):

1. 2026-09-03 — OpenAI announces GPT-6 Astra ("A new generation of intelligence"); "the world's most intelligent and aligned model." Rolling out that day to a limited set of organizations; over the coming days to all ChatGPT Plus/Pro/Business/Enterprise users, the OpenAI API, Microsoft Azure and AWS Bedrock. OpenAI help-center release notes same day: "Astra is not yet generally available." Astra is positioned as the strongest model for computer use, coding, professional work, science, cybersecurity (FACT — openai.com; corroborated by VentureBeat, InfoQ, 9to5Google, CNBC).
2. 2026-09-03 — Microsoft publishes "GPT-6 Astra: Frontier intelligence for work, now available in Microsoft Foundry" (Azure blog, Steve Sweetman — VP Product Management Foundry Models — and Naomi Moneypenny): Astra begins rolling out through the Microsoft Foundry Limited Access Program, availability expanding to participating customers over the coming days (COMPANY CLAIM — Microsoft; consistent with OpenAI's own "limited set of organizations" framing).
3. 2026-09-04 — Astra joins Copilot Cowork and Copilot Studio (Microsoft 365 Copilot blog), grounded in Work IQ; availability administered via the M365 admin center. Also rolls into GitHub Copilot the same week (Microsoft 365 blog; SD Times Sep 8 covers Foundry + GitHub Copilot).
4. 2026-09-15 — AWS makes GPT-6 Astra generally available on Amazon Bedrock (IT Brief Asia — INDEPENDENT EVIDENCE), with ChatGPT Work/Codex configurable to use Astra on Bedrock.
5. 2026-09-17 — GA on Microsoft Foundry. The Azure blog becomes "GPT-6 Astra: Frontier intelligence for work, now generally available in Microsoft Foundry": "GPT-6 Astra, OpenAI's newest frontier model, is now generally available for all customers in Microsoft Foundry" (COMPANY CLAIM — Microsoft). The Foundry model catalog lists gpt-6-astra, version 2026-09-03, Lifecycle: Generally available (GA) (COMPANY CLAIM — Microsoft catalog, verified live). An independent daily release roundup covering the Sep 16–17 window states: "Microsoft made it generally available in Azure AI Foundry Models the same day (Sep 17)" (INDEPENDENT EVIDENCE — aggregator, lower weight, but consistent with catalog state and the updated blog).

The same week, OpenAI published "GPT-6 Astra: The next generation in intelligence for work" (page dated 2026-09-17 in search index; text: "Last week we introduced GPT‑6 Astra … now available in ChatGPT Work, Codex, and the API"), with enterprise admin controls, enterprise plugins (Oracle Analytics, Power BI, Navan, Avalara), and confirmation that Astra is the first model to reach the Critical cybersecurity capability threshold under OpenAI's Preparedness Framework (COMPANY CLAIM — OpenAI; framework classification cross-covered in S07).

• Enterprise deployment gate removed: before Sep 17, running Astra on Azure required admission to the Foundry Limited Access Program; from Sep 17 any Foundry customer can deploy gpt-6-astra via Foundry Models with Standard (pay-as-you-go) and Provisioned Throughput options in Global and US Data Zone geographies (COMPANY CLAIM — Microsoft; consistent with catalog + pricing pages).
• Frontier compute leaves the consumer subscription moat: Astra previously reached most organizations through ChatGPT Work/Codex subscription allowances (with metered credit purchases) or the OpenAI API; Foundry GA gives enterprises a governed, SLA-backed, contract-free path on Microsoft's enterprise platform (INTERPRETATION grounded in Microsoft's own "experimentation to production with speed and trust" positioning).
• Azure/OpenAI commercial surface expands: GPT-6 Astra is sold "Direct from Azure" — unified billing, governance, PTU portability, single license/support, no third-party dependency (COMPANY CLAIM — Microsoft catalog).
• Catalog lifecycle state: on the Foundry catalog, gpt-6-astra is the flagship OpenAI entry marked GA, while peers visible on the same catalog (claude-mythos-5-1 "Gated Research Preview", grok-4.6 "Preview") are not yet GA at this snapshot (observation from catalog listing; INTERPRETATION for competitive positioning).
• Multi-cloud parity: Astra became GA on AWS Bedrock (Sep 15) and Microsoft Foundry (Sep 17) within 48 hours, plus OpenAI API — the "exclusive cloud" era is over for OpenAI's flagship (FACT of dates; INTERPRETATION of significance).
• Pricing surfaced for enterprise planners: Foundry Standard Global $10 input / $50 output per M tokens (short context), $20/$75 long context; US Data Zone +10% ($11/$55, $22/$82.50); cached input $1/$2; cache writes $12.50/$25 ($13.75/$27.50 US DZ); Provisioned Throughput at ~$1.00/PTU-hour global ($1.10 US DZ) per Microsoft pricing (COMPANY CLAIM — Microsoft pricing documentation).

• Deployment mechanics (COMPANY CLAIM — Microsoft): via Foundry Models, choose Standard (serverless, pay-per-token, global or data-zone) or Provisioned Throughput (reserved capacity, consistent latency; priced per PTU). Both in Global and US Data Zone geographies at GA. Model version 2026-09-03.
• Model card (OpenAI API docs + Foundry catalog — primary): model ID gpt-6-astra; Chat Completions and Responses API; input text + image, output text; 1,050,000-token context window (922K max input), 128K max output; knowledge cutoff Apr 30, 2026; reasoning effort low/medium/high/xhigh/max; no temperature while reasoning enabled.
• Pricing model (FACT of published rates; COMPANY CLAIM as vendor rates): OpenAI API Standard $10 in / $50 out per M tokens (cached $1, cache writes $12.50); prompts over 272K input tokens billed at 2x input/cache and 1.5x output for the full request; Batch/Flex at 50%, Fast mode at 2x. Foundry Standard mirrors this: long-context tier ≈ 2x input, 1.5x output; US Data Zone adds 10%.
• Enterprise governance (COMPANY CLAIM — Microsoft): Microsoft Entra identity/access management, encryption in transit and at rest, private networking, RBAC, content filtering, safety evaluations, monitoring and governance tooling; prompts and outputs are not used to train models. Microsoft's design guidance for agentic work: scoped credentials, approved resources, human checkpoints for consequential actions, activity records. Content in apps may be "incomplete, misleading, or designed to influence an agent's behavior" — i.e., prompt-injection exposure acknowledged by Microsoft itself.
• OpenAI-side safety stack (COMPANY CLAIM — OpenAI system card): most aligned model to date; internal computer-use safety benchmark shows unintended outcomes 89% less often than GPT-5.6 Sol (74.7% less than Claude Fable 5.1); confirmation policies and automated review for tool calls; misalignment monitoring (flagged tasks stop); first model at Critical cybersecurity capability threshold under the Preparedness Framework; enterprise access off by default (admins enable).

• Frontier-model access escapes consumer/subscription lock-in (the discovery thesis, refined): Foundry GA is the first broadly available, governed enterprise deployment path for OpenAI's strongest model; the enterprise inference race for "agent-grade" frontier models just got a defined price and SLA surface on Azure (INTERPRETATION; supported by Microsoft's "enterprise AI will not be defined by chat experiences" framing — COMPANY CLAIM).
• The "AI worker" workload finally has a deployable substrate: 1M-token context + computer use + tool use + enterprise governance is the combination that lets a compliance team accept "an agent completes a unit of work" (echoed by Microsoft MVPs/practitioners — e.g., Vesa Nopanen's independent hands-on-assessment; INTERPRETATION).
• Cloud strategy inflection: Astra GA on Foundry (Sep 17) 48 hours after AWS Bedrock GA (Sep 15) confirms OpenAI's multi-cloud distribution and ends the "Azure-exclusive OpenAI" period; Microsoft responds by making Foundry the richest governance/compliance wrapper for OpenAI models (FACT of dates; INTERPRETATION).
• Competitive pacing: at GA, Astra is ahead of Claude Mythos 5.1 (gated research preview) and Grok 4.6 (preview) in the Foundry catalog — the strongest model is also the first to be enterprise-GA on Azure (observation; INTERPRETATION).
• Agent-risk era: the same week anchors enterprise deployment of the most capable computer-use model, its Critical cybersecurity classification (S07), and OpenAI's misalignment-incident disclosure framework (S02) — capability and governance now ship together (linking interpretation).

• Deploy OpenAI's top reasoning/computer-use model inside an Azure subscription without Limited Access admission, partner agreements, or consumer plans (FACT — Microsoft announcement).
• Build governed agentic workflows — cross-application task execution with scoped credentials, human checkpoints and activity records — via Foundry Agent Service plus Astra (COMPANY CLAIM — Microsoft).
• Run agent workloads on reserved PTU capacity with guaranteed throughput for latency-sensitive production systems.
• Cost-engineer agentic pipelines: cached input at $1–$2/M, long-context tier for >272K-token windows, batch at 50% — enterprises can now model per-task economics in procurement.
• Deploy the same model API-identically across OpenAI API, Azure and AWS — portable agent stacks (FACT — model id gpt-6-astra across providers; Bedrock GA Sep 15 per IT Brief Asia).
• Give Astra read/write access to enterprise systems (Power BI dashboards, CRM record updates, codebases, document pipelines) under admin-controlled surface restrictions (enterprise plugins: Oracle Analytics, Power BI, Navan, Avalara — COMPANY CLAIM — OpenAI).

• When do EU/APAC data zones become available for Astra on Foundry (parity with Global/US DZ)?
• What quota defaults apply per Azure tier for gpt-6-astra, and how fast can PTUs be provisioned?
• Will Foundry's Model Router pools include gpt-6-astra, and when?
• Does the Critical national-security classification (S07) change Microsoft-side access, monitoring or contractual terms for enterprise deployments?
• What independent benchmark replication will confirm Astra's computer-use/coding claims at production configuration (agent harness, not raw model)?
• How will OpenAI's misalignment-disclosure framework (S02) apply to enterprise Foundry deployments — do enterprises get the same telemetry into monitor flags?
• Will Azure batch and Flex pricing (50%) and Fast mode (2x) land on Foundry for Astra, matching the OpenAI API?
• Whether enterprise demand shifts to Bedrock (AWS) vs Foundry (Azure) changes Microsoft's pricing/PTU strategy within quarters.

• Short term (days–weeks): expect OpenAI/Microsoft blog follow-ups on Astra enterprise adoption metrics; catalogue/watch for EU data-zone and Model Router availability; enterprise agent platforms (Copilot Studio, Agentforce alternatives, Bedrock Agents) race to feature Astra first (PREDICTION).
• Mid term (this quarter): rival frontier GA announcements on Foundry (Claude Mythos 5.1, Grok 4.6 exit preview) pressure Astra pricing/PTU terms; AWS/Azure enterprise inference pricing converges; OpenAI publishes next evals and usage economics (PREDICTION).
• Structural: the Critical classification (S07) and misalignment-disclosure framework (S02) become the template governing future frontier GA releases; expect enterprise contracts to start encoding misalignment-monitor telemetry and stop-semantics SLAs (PREDICTION).
• Watch item: whether Foundry adds Astra to batch/Flex (50%) and Fast (2x) modes matching the OpenAI API — a direct margin lever for high-volume agent workloads (OPEN QUESTION / PREDICTION).

"The world's strongest model just got an enterprise front door." GPT-6 Astra's Foundry GA on Sep 17 is the moment frontier capability became a governed, purchasable production resource rather than a gated flagship demo — on Microsoft's cloud, 48 hours after AWS did the same. The week's deeper story is the collision this creates: agents capable of doing real work in real systems (and, per its own classification, real damage) are being deployed at industrial scale while lab CEOs and regulators argue about pacing. For enterprises, the takeaway is concrete and actionable: the moat is no longer model access — it is the governance, cost-engineering, and evaluation discipline around it. Deploy small, contain hard, measure everything.

Evidence labels used

• FACT — event dates, prices as published, model card specs, catalog lifecycle states, quote attributions (verified in primary pages during research).
• COMPANY CLAIM — OpenAI/Microsoft capability, safety, alignment and benchmark assertions, plus all pricing as vendor rates.
• INDEPENDENT EVIDENCE — InfoQ, VentureBeat, SD Times, IT Brief Asia coverage; the Sep 17 aggregator roundup dating the Foundry GA; Nopanen's independent practitioner assessment.
• INTERPRETATION — strategic/competitive/regulatory readings, clearly marked as analysis.
• PREDICTION — forward-looking statements in §17/§20, labeled as such.