Azure AI-plattformen
For most Swedish and European enterprises, the practical question about generative AI is not which model wins a benchmark, but how to run capable models inside an environment the organisation already trusts. That is what the Azure AI platform has come to mean in practice: Azure OpenAI and Microsoft Foundry for models and agents, Azure AI Search and the existing data estate for retrieval, Entra for identity, and governance tooling that EU regulation increasingly assumes you have. If your organisation already runs on Microsoft 365 and Azure, this stack is usually the shortest path from a promising prototype to a governed production workload.
The convenience hides real architectural decisions. What began as a single managed OpenAI endpoint has grown into a catalogue spanning OpenAI, Anthropic, Mistral, Meta, Cohere and a long tail of open-weight models, plus agent services, evaluation tooling and several competing answers to vector search. Teams that treat it as a ChatGPT API with a Microsoft logo tend to overspend, under-govern, or both. The teams that do well treat platform choices as a set of separable decisions, each with its own reversibility profile.
The decisions that shape your architecture
Model choice is a routing problem, not a marriage
Frontier models, mid-tier workhorses and small open-weight models now sit side by side in the same catalogue, and their relative price and capability shift every few months. A summarisation endpoint, a coding agent and a customer-facing chatbot rarely justify the same model. The durable investment is not the model selection itself but the machinery around it: an abstraction layer that lets you swap deployments, and an evaluation suite that tells you when a cheaper model is good enough for a given workload. Vendors will keep repricing; your evals are what turn that churn into savings instead of risk.
Capacity and cost need engineering, not just budgeting
Azure gives you several ways to pay for the same tokens: pay-as-you-go, provisioned throughput (PTUs), batch processing and increasingly aggressive prompt caching. Which combination is right depends on traffic shape, latency requirements and how repetitive your prompts are. Agentic workloads deserve special attention here, because an agent that loops over tools can consume an order of magnitude more tokens than the chat feature it replaced. Cost modelling belongs in the design review, not in the post-mortem after the first invoice.
Retrieval is where your data architecture meets the model
Almost every serious enterprise workload is grounded in company data, which makes the vector search decision more durable than the model decision. Azure AI Search, PostgreSQL with pgvector and dedicated vector databases all work; they differ in cost profile, hybrid search quality, filtering and operational burden. The right answer usually follows your existing data gravity. Just as important is what goes into the index: document parsing quality, chunking strategy and the provenance of scraped or third-party content, which in the EU carries its own legal weight.
Agents change the identity model
Once a model can call tools, hit internal APIs or operate a browser, it stops being a text feature and becomes a workload with permissions. Azure's direction here is clear: agents get first-class identities in Entra, with least-privilege access and auditable actions, rather than borrowing a developer's service principal. Platform-level agent services are also replacing earlier assistant-style APIs, so teams carrying older integrations should plan migration deliberately instead of waiting for a retirement deadline to force it.
Residency and regulation are architecture inputs
For EU workloads, region strategy is not an afterthought. Model availability differs by region, Sweden Central included, and sovereignty expectations from customers and regulators keep rising. Running on Azure also does not outsource your obligations under the EU AI Act: as a deployer you own risk classification, transparency duties for AI interactions and synthetic content, and the logging that proves you did what you claim. The platform provides useful primitives for all of this, but the accountability stays with you.
A useful rule of thumb: treat model choice as reversible and everything around it as durable. Evaluation harnesses, retrieval pipelines, identity design and cost telemetry outlive any individual model generation. Invest accordingly.
How to approach it
Start from the workload, not the platform diagram. Pick one use case with measurable value, define what good output looks like, and build the smallest evaluated pipeline that delivers it in an EU region with proper identity and logging from day one. Then expand along the seams:
- Add an evaluation suite before you add a second model, so routing decisions rest on evidence.
- Instrument token usage per feature early; retrofitting cost attribution is painful.
- Choose your vector store based on where your data already lives, and benchmark retrieval with your own queries.
- Give every agent its own identity and the narrowest permissions that let it work.
- Map each workload against AI Act obligations while it is still cheap to change.
The articles below go deeper on each of these fronts, from pricing mechanics and migration checklists to vector search comparisons and compliance for deployers. They date faster than this page will, which is precisely the point: the platform moves quarterly, while the decision framework above stays stable enough to build on.
Alla artiklar i ämnet
Guiden växer i takt med att vi publicerar — nya artiklar dyker upp här automatiskt.
Cohere Parse v5 landed in Microsoft Foundry on 27 August 2026 at $1.50 per 1,000 pages, the same price as Document Intelligence Read and under a third of Content Understanding's layout meter. We price all four Azure document parsers in Sweden Central at 100k, 1M and 10M pages a month, then weigh the saving against Parse's missing confidence scores, absent Swedish language support and preview-only status.
The Azure OpenAI Assistants API reached its retirement date on 26 August 2026, and the classic Foundry Agent Service it underpins retires 31 March 2027. A step-by-step migration guide to the new Foundry Agent Service on the Responses API: threads become conversations, runs become responses, assistants become versioned agents, and Microsoft's migration tool rewrites code but not stored state.
DeepSeek's V4-Pro reached general availability on 13 August 2026 with a 1M-token context and strong agent benchmarks, and three days later its peak-hour output price rose from a flat $0.87 to $3.96 per million tokens. For EU teams that built agent cost models around ultra-cheap open-weight APIs, the arithmetic, the data governance questions and the Azure hosting options all deserve a fresh look.
Google shipped gemini-3.7-flash as generally available on 13 August 2026 at an introductory $0.75/$3.75 per million tokens, two weeks after OpenAI cut GPT-5.6 Luna by 80 percent and days after Anthropic locked Claude Sonnet 5 at $2/$10 permanently. We compare the three workhorse models on list price, context, Azure availability and EU residency, and show why cost per completed task beats cost per token.
Anthropic will weave an invisible watermark into Claude's text output to meet the EU AI Act's Article 50 marking obligation, applying it globally across the API, apps and cloud platforms including Microsoft Foundry. What the mark can and cannot prove, which deployer duties remain yours, and when a DIY provenance layer still earns its keep on Azure.
Anthropic has cancelled the Claude Sonnet 5 price increase scheduled for 1 September 2026, making the introductory $2 input / $10 output per million tokens the permanent standard price. For teams running Claude on Azure through Microsoft Foundry, that removes a planned 50% jump from autumn budgets and reshapes the mid-tier price comparison against GPT-5.6 Terra and Gemini 3.1 Pro.
OpenAI removed limits on text chats for free ChatGPT users on 6 August 2026 and made GPT-5.6 Luna the default, cutting factual errors by roughly 62 percent versus the prior model. For Swedish and EU enterprises on Azure, the free consumer tool employees already use just became unlimited and much stronger, so shadow AI pressure rises and the case for a governed answer built on Copilot Chat, paid Copilot seats and Azure OpenAI becomes urgent.
On 2 August 2026 the European Commission's enforcement powers over general-purpose AI providers activated: the AI Office can now demand documentation, run model evaluations, restrict models from the EU market and fine up to 3% of global turnover or EUR 15 million. The same date brought Article 50 transparency into application, and this guide maps what Azure OpenAI and Foundry teams must demand from vendors versus handle themselves as deployers.
OpenAI cut GPT-5.6 Luna prices by 80 percent and Terra by 20 percent on 30 July 2026, and Microsoft confirmed the same decreases reach Azure OpenAI Global Standard deployments from 1 August. For Swedish and EU teams running these models on Azure, the cuts move the break-even point for PTU reservations, model routing and residency premiums, so the autumn budget math deserves a fresh pass before any new one-year commitments.
Claude Opus 5 launched on 24 July 2026 at $5/$25 per million tokens with a 1M context window and day-one availability in Microsoft Foundry. For long-running agents the per-token price is the wrong unit: we work through cost per completed task against Sonnet 5 and GPT-5.6 Sol, and flag the EU data-residency caveat Swedish Azure teams need to check first.
Microsoft and Mistral announced an expanded partnership on 21 July 2026: Mistral Medium 3.5 and OCR 4 arrive in Microsoft Foundry and Copilot Studio, deployable from Azure cloud to customer-controlled and fully air-gapped Azure Local environments, backed by a multibillion-dollar European GPU buildout. We compare the three deployment modes and what each one solves for Swedish public sector and regulated industries.
Google's 21 July release of Gemini 3.6 Flash pairs an output-price cut from $9.00 to $7.50 per million tokens with a claim of roughly 17% fewer output tokens per task, compounding to about 31% lower output cost for unchanged work. That combination makes per-million-token price sheets unreliable for model comparison, and Azure teams should measure cost per completed task instead.
The European Commission adopted its final Article 50 transparency guidelines on 20 July 2026 and confirmed the Code of Practice on marking AI-generated content as adequate, less than two weeks before the obligations start to apply. Here is what Swedish and EU teams running chatbots, copilots and content generators on Azure OpenAI must implement: chatbot disclosure, machine-readable marking and deepfake labels, with concrete code patterns for each.
GPT-5.6 (Sol, Terra, Luna) went GA in Microsoft Foundry on 9 July 2026, day-and-date with OpenAI, alongside a new Asia-Pacific Data Zone and a hosted agents runtime with VNet integration. A practical guide for Swedish and EU Azure teams: choosing between the three models, picking Global Standard versus EU Data Zone versus PTUs, worked cost math on the launch prices, and a two-week adoption checklist.
OpenAI released the GPT-5.6 series on 9 July 2026 in three tiers: Sol for hard reasoning and long autonomous runs, Terra for everyday work and Luna for speed and cost, with same-day availability in Microsoft Foundry and a new preferred-model role in Microsoft 365 Copilot. This playbook maps Azure workloads to the right tier, works the token math in SEK and flags the Copilot subprocessor setting Swedish admins must review before 24 July.
At its 122nd plenary on 7 July 2026 the EDPB adopted draft Guidelines 03/2026 on web scraping for generative AI and draft Guidelines 02/2026 on anonymisation, the first comprehensive GDPR framework for large-scale scraping of public web data into AI pipelines. We walk through the legitimate interest test, the new legal weight of robots.txt and ai.txt, the Article 9 lifecycle safeguards, and what it means for teams ingesting web content into RAG on Azure.
Anthropic launched Claude Sonnet 5 on 30 June 2026 at an introductory 2/10 dollars per million tokens, posting 63.2% on SWE-bench Pro and near-Opus agentic performance at 40-60% of the cost per task. We run the cost math against Opus 4.8 and GPT-5.5, set out a routing framework for when the cheap model wins, and draw the continuity lesson from the eighteen-day Fable and Mythos export-control pause.
Claude Opus 4.8 and Claude Haiku 4.5 are now generally available in Microsoft Foundry, hosted on Azure with Entra ID authentication, prompt caching, extended thinking and billing through Claude Consumption Units on your existing Azure invoice. Deployment steps, the CCU cost model compared with Azure OpenAI, and the data-residency caveats Swedish and EU teams should assess before production use.
Google made computer use a native tool in Gemini 3.5 Flash on 24 June 2026, putting vision-based UI agents in its low-cost tier for browser, mobile and desktop automation. A comparison with Azure AI Foundry's Computer Use and Browser Automation previews, with a decision framework for legacy-UI automation and the GDPR implications of shipping screenshots to a model endpoint.
Microsoft made Copilot Cowork generally available worldwide on 16 June 2026, giving every Microsoft 365 Copilot tenant an agentic system that plans and delivers multi-step work, billed through Copilot Credits at 0.01 dollars each with no usage bundled into the license. Swedish IT teams now need spending limits, a DPIA update for the Anthropic-by-default model lineup, and a clear answer on when Cowork beats building a Foundry agent.
On 16 June 2026 the European Parliament approved the Digital Omnibus amendments 423-57, moving Annex III high-risk AI Act obligations to 2 December 2027 and product-embedded obligations to 2 August 2028. Article 50 transparency duties and the Commission's GPAI enforcement powers were not delayed, which leaves Swedish enterprises six weeks to ship chatbot disclosure, content marking and a documented GPAI position before 2 August 2026.
Azure HorizonDB reached public preview at Build 2026: a PostgreSQL-compatible, scale-out engine with vector search built in, up to 128 TB of storage and Sweden Central among the five launch regions. For RAG embeddings on Azure the choice now runs three ways, decided by where the source data lives, how much retrieval quality matters and how close you are to Flexible Server's vertical-scaling ceiling.
Microsoft Build 2026 turned Foundry into a full production-agent stack: Foundry IQ for unified retrieval, Toolboxes for managed tool access, agent memory, Voice Live and the experimental Scout Autopilot. Foundry IQ knowledge bases and Voice Live are generally available now, Toolboxes and memory sit in public preview, and Scout remains experimental, which sets the build, pilot and watch lanes for an Azure-first EU team.
Google's I/O 2026 enterprise announcements, led by the Managed Agents API, Gemini Spark and Gemini 3.5 Flash, take direct aim at Azure AI Foundry Agent Service and the Copilot ecosystem, down to launch connectors for SharePoint and OneDrive. For Azure-first Swedish and EU teams the decision rests on four questions: where the data lives, whether the cost claim survives real traces, whether audit obligations can be met, and what a second platform costs.
Google shipped Gemini 3.5 Flash on day one of I/O 2026, the third frontier agentic coding model in five weeks after Claude Opus 4.7 and GPT-5.5, alongside a $100/month developer tier and the Antigravity 2.0 agent platform. For Azure-first teams the practical response is a reusable evaluation harness and a documented cost comparison for renewal leverage, not a migration.
Frontier models are the default. Defaults are how teams overpay on LLM bills. Three workloads where small models (Phi-4, Llama 3.x 8B, Mistral Small) outperform on cost-per-decision without losing meaningfully on quality, three workloads where they do not, and the two-tier production pattern that cost-conscious teams converge on after a quarter of evaluation work.
The vector-database market settled in 2025. The 2026 question is less "which novel DB do we adopt" and more "do we need a separate vector store at all." A decision framework across Azure AI Search, pgvector on managed Postgres, and the dedicated managed vector platforms, with the workloads each one handles well.
Prompt caching is the highest-ROI cost lever on long-context LLM workloads in 2026. Anthropic, OpenAI, and Azure OpenAI all offer it with different pricing and breakpoint semantics. A worked comparison of the three providers, the placement patterns that actually hit cache, where the cache silently goes cold, and a 30-minute audit that pays back.
A deep dive into Microsoft Entra Agent ID, the control plane for AI agent identity in 2026. Covers identity blueprints, attended and unattended authentication, tool-level RBAC, conditional access, OBO flows across multi-agent systems, and the audit logging that satisfies DORA, NIS2, and AI Act obligations.
The capability spectrum from stateless LLM to multi-agent orchestration is one of the most conflated concepts in the 2026 AI market. The distinctions matter. They change architecture, they change cost by an order of magnitude, and under the EU AI Act they change compliance posture.
Industrial documents arrive as scanned paper, multilingual procedures, and proprietary CAD exports. Azure AI Search skillsets turn them into queryable text with structured metadata. This post walks the OCR, entity, language, and custom skill patterns that make the difference between a searchable index and one that actually answers technical queries.
A concrete playbook for reducing Azure OpenAI bills in 2026. Break-even math for Provisioned Throughput Units, prompt-cache economics, the Batch API 50 percent discount, Foundry IQ for retrieval, tiered model routing, and the telemetry that keeps the wins honest.
Generic LLM copilots are a liability in manufacturing. Technicians need answers that cite the exact procedure, not plausible-sounding text. Retrieval-augmented generation grounded in Azure AI Search solves this when architected correctly. This is the pattern that holds up under service-bay pressure.
A practical comparison of SQL Server 2025 vector search against Azure AI Search and pgvector for hybrid RAG — when collocating vectors with your transactional data is the right call, and when a dedicated search layer still wins.
Classic RAG hits a ceiling when questions require multi-hop reasoning or query refinement. Agentic RAG — treating retrieval as a tool, decomposing queries, adding self-correction loops — routinely wins where classic RAG plateaus.
Automotive technical documentation is uniquely structured: ECU specs, wiring diagrams, service manuals, diagnostic codes, recall notices — all layered across vehicle platforms, model years, and markets. Azure AI Search handles this at scale if the index schema and retrieval pattern match the domain. This post walks the automotive-specific design choices.
Pure keyword search misses semantic matches. Pure vector search misses exact part numbers. Hybrid search combines both. This post explains the rank fusion and semantic re-ranking choices that turn a hybrid Azure AI Search query from "works on demos" into a production retrieval layer for manufacturing knowledge bases.
Manufacturing corpora mix scanned paper, CAD exports, work instructions, and supplier specs. Making them searchable with high recall and precision requires a thoughtful Azure AI Search skillset and indexer design. This post walks through the OCR, entity extraction, chunking, and semantic ranking choices that separate production-quality indexes from POCs.
Microsoft renamed Azure Cognitive Search to Azure AI Search in late 2023 and added substantial vector, semantic, and agentic-retrieval capabilities. Manufacturing teams running legacy Cognitive Search deployments need a structured migration plan — not a like-for-like upgrade, but a targeted modernisation of the features that matter most.
Manufacturing document processing is a high-volume, high-format-diversity problem. Azure AI Search (formerly Azure Cognitive Search) is the Microsoft-native platform that fits it well when configured correctly. This post walks through the architecture patterns that actually ship at industrial scale.
Microsoft Ignite BRK270: Transform compliance from burden to accelerator. Microsoft Purview, Compliance Manager, AI Baseline, Regulatory Navigator. Automated governance for GDPR, AI Act, NIS2. 3-5x faster deployments, 70-90% less audit time.
See the magic happen in real time. Learn how to train and deploy custom reasoning models with Azure ML and Microsoft Foundry—from fine-tuning to reinforcement learning, performance optimization with speculative decoding, distillation, and production deployment delivering measurable ROI.
Microsoft Ignite BRK203: Reasoning models as the brains behind autonomous agents. Multi-step problem solving, explainable decisions, agentic workflows (lead scoring, content generation, support). Foundry 11,000+ model catalog, customer stories from healthcare and legal sectors.
Agents need context. Discover Foundry IQ—the knowledge layer connecting AI agents to enterprise data with multi-source RAG orchestration, retrieval steering, dynamic security controls, and agentic RAG delivering 36% higher accuracy than traditional search.
From scientific research to enterprise AI transformation, discover how Microsoft Foundry unifies models from OpenAI, Anthropic, Cohere, Meta, and more into one secure platform. Learn intelligent model routing, cost optimization, and the game-changing Claude integration.
Microsoft Ignite BRK194: Model Context Protocol (MCP) standardizes agent-tool communication across platforms. Azure API Center integration, federated registries, cross-cloud orchestration, and enterprise governance for scalable agentic ecosystems.
Microsoft Ignite BRK193: Build agents with Azure AI Search knowledge features. Connect to SharePoint, web, blob. Hybrid search (keyword+vector+semantic), agentic retrieval with query planning, reasoning effort modes, Foundry IQ with MCP protocol. Code-focused implementation guide.
Ride the agent revolution with Microsoft Foundry, the enterprise-ready Agent Factory. Build, test, and launch intelligent agents with 1,400+ tools, 11,000+ models, multi-agent orchestration, and seamless Microsoft 365 integration—all with bulletproof security and governance.
Microsoft Ignite BRK188: Fine-tuning in Microsoft Foundry transforms generic models into production-ready agents. Synthetic data generation, supervised + reinforcement fine-tuning, 40-90% cost reduction, 95%+ accuracy. Real-world results: 2M docs/day, $27M savings.
Microsoft Ignite BRK116: Deploy AI agents at scale with Azure App Service, AI Foundry, and MCP tools. Built-in observability, governance, security. Hitachi case study shows 73% downtime reduction and 41% cost savings.
Microsoft Ignite BRK114: Enterprise customer panel on production AI agents. Multi-tier architectures, event-driven orchestration, defense-in-depth security, human-in-the-loop patterns. Democratization with governance. Model selection optimization. Real deployments delivering $10M-$50M+ annual value. From experimentation to strategic differentiation.
Vanliga frågor
What is the difference between Azure OpenAI and Microsoft Foundry?
Azure OpenAI is the managed service for running OpenAI models inside your Azure tenant. Microsoft Foundry is the broader platform around it: a model catalogue spanning multiple vendors, agent services, evaluation tooling and deployment management. New projects typically start in Foundry and use Azure OpenAI deployments as one model source among several.
Should we buy provisioned throughput (PTUs) or stay pay-as-you-go?
Stay pay-as-you-go until you have stable, measured traffic. PTUs pay off with predictable sustained load and strict latency requirements, and many teams run a PTU baseline with pay-as-you-go spillover for peaks. In practice, prompt caching and batch processing often cut costs more than changing the pricing model does.
Can we run Claude, Mistral or open-weight models on Azure?
Yes. The Foundry model catalogue includes Anthropic's Claude models, Mistral, Meta's Llama family, Microsoft's Phi series and many open-weight options, deployable as serverless APIs or on managed compute. Pricing and regional availability vary per model, so verify both for your target EU region before committing.
Which vector store should we use for RAG on Azure?
Azure AI Search is the default for most teams because of its hybrid keyword-plus-vector retrieval and platform integration. If your data already lives in PostgreSQL, pgvector is often sufficient and cheaper to operate. Dedicated vector databases mainly earn their place at very large scale or with unusual filtering and latency needs, so benchmark with your own queries first.
Can we keep our AI workloads and data in the EU or in Sweden?
Azure OpenAI and most Foundry models can be deployed to EU regions, including Sweden Central, and Microsoft's EU Data Boundary commitments cover in-scope services. New models often reach US regions first, so check availability per model and region. For strict sovereignty requirements, review the specific service's documented data flows rather than assuming blanket coverage.
How does the EU AI Act affect us if we build on Azure AI?
Building on Azure OpenAI or Foundry generally makes your organisation a deployer under the AI Act, and in some cases a provider, with obligations depending on the use case's risk classification. Transparency duties such as disclosing AI interactions and marking synthetic content apply to many ordinary chatbot and content workloads. Azure's governance tooling helps with logging and oversight, but legal accountability remains with your organisation.