4 MIN READ
Two years ago the AI stack was a free for all. By Q2 2026 it has settled into a shape. Foundation models, orchestration frameworks, vector databases, agent frameworks, evaluation tools, deployment platforms, observability tooling. Buyers know what they are buying, vendors know what they are selling, and the market has stopped being the wild west.
Quick map of who is in the room. Foundation models: OpenAI (GPT-5 and the o series), Anthropic (Claude 4 with the Sonnet and Haiku split), Google DeepMind (Gemini 2.5 Pro and Flash), Meta (Llama 4 in the open weights), xAI (Grok 3), and the Chinese players (DeepSeek, Qwen, the Baidu ERNIE team). Orchestration frameworks: LangChain, LlamaIndex, Pydantic AI, Vellum. Vector databases: Pinecone, Weaviate, Qdrant, and the pgvector extension on Postgres. The choices sit clear, which is the change worth noting.
Where the stack has actually matured
Four layers have shipped production ready tooling. The foundation model layer, where the major labs (OpenAI, Anthropic, Google, Meta) now ship the production ready models, with context windows in the millions of tokens, cost per million tokens in the single digit dollars, and latency under 500 milliseconds for the typical prompt. The orchestration layer next, where LangChain, LlamaIndex and Pydantic AI ship the production ready abstractions, with agent loops, tool calling, structured output, and streaming all mature. Then the evaluation layer, where Braintrust, LangSmith, and Honeycomb (for the AI traces) ship the production ready evaluation, with model grading, human review, and regression testing all mature. Finally the deployment layer, where AWS Bedrock, Azure AI Foundry, Google Vertex AI, and vLLM (for the self hosted option) ship the production ready deployment, with autoscaling, A/B testing, and canary deploys all mature. Add them up and these four layers handle most of what a typical AI workload needs in 2026.
Where the work is still open
Three areas where the stack is not done. Agent reliability sits as the single largest gap. The agent frameworks (LangChain, the Anthropic SDK, the OpenAI Assistants API) can produce an agent, but agents run for 10 steps before losing the context, and 50 steps before going off the rails. Cost predictability is the second gap. Foundation model pricing sits published, but the actual cost of running an agent (model cost, tool cost, retry cost, human review cost) adds up to a different number than the published price. Procurement budgets the published price and gets the agent cost. Regulation is the third. The EU AI Act sits in force, the US executive order on AI sits in force, but the documentation requirements still get interpreted differently by every auditor. Most procurement teams do not have a stable answer yet.
What to actually buy
Three moves if you are buying the AI stack in 2026. First, pick the foundation model for the use case, not the benchmark leader. The benchmark winner and the production winner usually end up at different labs. The production winner sits at whichever lab has the best fit for the specific workload being built. Second, pick the orchestration framework the engineering team can actually operate. A framework nobody understands ends up abandoned. A framework the team can read and extend ends up running the whole production stack. Third, pick the deployment platform that integrates with the existing infrastructure. A platform that needs new IAM, new networking, and new monitoring will not survive contact with the operations team. A platform that drops into the existing VPC, uses the existing CI/CD, and ships logs into the existing SIEM tends to be the one that survives the first quarter in production.
The bottom line
Mature layer, framework the team can actually run, deployment that fits the existing stack. That is the buyable shape of AI in Q2 2026. The rest, agent reliability, cost predictability, regulatory clarity, is still owed, and pretending otherwise is how the AI project ends up shelved by Q4.
Sources & Further Reading
All claims in this article are sourced from primary documentation, vendor advisories, and reputable security researchers.
Spotted an error? Email the editor. Corrections are issued with a visible correction note.
Editorial standards. Every article on humanrequired.org is reviewed by a human editor before publication. AI may assist with drafting or research; final editorial control is human. Read the full standards.
