LangGraph vs CrewAI vs Semantic Kernel (Now Agent Framework) vs OpenAI Agents SDK 2026
AutoGen is archived and Semantic Kernel is now Microsoft Agent Framework. A comparison of the open agent frameworks that are left, plus a worked hosting-cost example from each vendor's own price list.
The short answer
Pick LangGraph if you need durable, checkpointed state and human approval steps. Pick CrewAI if role-based crews suit the problem and Python-only is acceptable. Pick Microsoft Agent Framework, the 1.0 successor to Semantic Kernel and AutoGen released on 3 April 2026, if you run .NET on Azure. Pick the OpenAI Agents SDK for the thinnest code path. All four are MIT-licensed. LangChain's managed LangGraph hosting starts at $39 per seat per month, including one free small deployment.
What changed in 2026
AutoGen is archived. The AutoGen README now opens with "Project archived and no longer maintained" and says "Microsoft is not providing bug fixes, security fixes, new features, or support for this project." It also says AutoGen "was developed for research and experimentation" and that "we continue to advise against production use." New users are sent to Microsoft Agent Framework, and existing users to a migration guide.
Semantic Kernel has been folded into the same product. Its README says: "Semantic Kernel is now Microsoft Agent Framework! Microsoft Agent Framework (MAF) is the enterprise-ready successor to Semantic Kernel." Microsoft's 1.0 announcement, dated 3 April 2026, says the team set out to "unify the enterprise-ready foundations of Semantic Kernel with the innovative orchestrations of AutoGen". It also says: "Coming from AutoGen or Semantic Kernel? Now is the time to migrate to Microsoft Agent Framework."
OpenAI is retiring the visual half of AgentKit. Its deprecations page says Agent Builder "is scheduled to shut down" on 30 November 2026; ChatKit remains available. The migration path is the Agents SDK or ChatGPT Workspace Agents. If your team prototyped in Agent Builder, migrate before 30 November 2026.
LlamaIndex has changed focus too. Its README says "our primary focus has shifted towards LlamaParse", its document platform, while the framework remains "an open toolkit". LlamaIndex.TS is "deprecated and no longer maintained".
Paid platforms and published prices
Prices and plan names are as published on each vendor's own pricing page; last verified 10 November 2026.
| Vendor | Billing unit | Published entry price | What it's for |
|---|---|---|---|
| LangGraph on LangSmith Deployment | Seats, plus LSU-metered compute ($1.00 per LSU) | Plus: $39 per seat per month, billed monthly. Includes 1 free Small serverless deployment | Hosting long-running, stateful LangGraph agents |
| CrewAI AMP | Workflow executions | Basic: free, 50 workflow executions per month. Enterprise: custom | Visual builder, deployment and governance for CrewAI |
| Microsoft Agent Framework on Foundry Agent Service | vCPU-hours and GiB-hours for hosted agents | No added charge for Foundry-native prompt and workflow agents. Hosted agents: $0.0994 per vCPU-hour and $0.0118 per GiB-hour (East US) | Running Agent Framework or LangGraph agents on a managed Azure runtime (the Agent Framework integration shipped as a 1.0 preview feature) |
| OpenAI Agents SDK and Agents API | Per tool call or session, plus model tokens | Web search $10 per 1,000 calls. File search $2.50 per 1,000 calls. Code containers from $0.03 per 20-minute session | Built-in tools. The SDK is self-hosted; the managed Agents API has no separate platform fee |
| LlamaIndex | Framework is free | Paid product is LlamaParse, a document platform, priced separately | Document-centric agents |
| Microsoft Copilot Studio | Copilot Credits | $200 per pack of 25,000 credits per month | Low-code agents for Microsoft 365 tenants |
CrewAI publishes no Enterprise price or overage rate, only "Flexible overage". LangSmith bills deployments by requested resources: "Each running billable container is charged for that interval based on requested resources, not actual utilization."
How the frameworks actually differ
LangGraph: a graph of nodes with saved state
LangGraph calls itself a "low-level orchestration framework for building stateful agents". You write nodes and edges yourself. Persistence is built in: according to the docs, "checkpointers persist a thread's graph state as checkpoints", and they are used for "human-in-the-loop workflows, time travel, and fault tolerance". To get human approval, you call interrupt() inside a node. The graph pauses and resumes when you send it a Command. You wire more than elsewhere, and in exchange see every state transition. There are Python and JS/TS versions.
CrewAI: roles, tasks and crews
CrewAI describes itself as a "standalone Python framework with its own primitives for agents, tasks, crews, flows, tools, and orchestration". Crews are "teams of AI agents with true autonomy and agency", and Flows are "event-driven workflows that deliver precise control". Use Flows for deterministic parts and Crews for open-ended ones. The README lists human-in-the-loop, memory and checkpointing, but it does not describe the storage behind checkpointing. The framework is Python-only and requires Python 3.10 to 3.13. See CrewAI vs LangGraph.
Microsoft Agent Framework: Semantic Kernel plus AutoGen
The Microsoft Agent Framework README describes "graph-based workflows supporting sequential, concurrent, handoff, and group collaboration patterns; includes checkpointing, streaming, human-in-the-loop, and time-travel", plus "built-in OpenTelemetry integration". Microsoft's announcement lists first-party connectors for Foundry, Azure OpenAI, OpenAI, Anthropic Claude, Amazon Bedrock, Google Gemini and Ollama. Durable Task and Azure Functions hosting are in a separate durable extension. There are two caveats. Microsoft says preview features "may evolve" before they become stable, and the same post says "A2A 1.0 support coming soon". If you are still choosing between LangGraph and Semantic Kernel, compare LangGraph with Agent Framework instead; see LangGraph vs Semantic Kernel.
OpenAI Agents SDK: the smallest set of primitives
The SDK has agents, handoffs, guardrails, sessions and tracing. It describes itself as "provider-agnostic", and you reach "100+ other LLMs" through LiteLLM or any-llm. For durable execution, its docs point to Dapr, Temporal, Restate and DBOS integrations for runs that "may span long waits, retries, or process restarts". Separately, OpenAI's Agents API is a managed runtime: "OpenAI manages sessions, orchestration, context compaction, and recovery", billed at model, tool and container rates. The model is portable; the tools are not. Web search, file search and code containers are OpenAI-hosted tools billed on OpenAI's price list. See LangGraph vs OpenAI AgentKit.
LlamaIndex: document agents
LlamaIndex Agent Workflows is "an event-driven orchestration library where steps are async Python functions". You can "persist state, recover from failures", and llama-agents-server adds "persistence, and human-in-the-loop support". Pick it when the agent's main job is parsing and extracting documents.
Get one new AI workflow every week
Practical playbooks, prompts, and tool stacks — for the SMB operators and enterprise teams actually shipping AI.
Side by side: licence, durability, lock-in
| Framework | Licence | Languages | Durable state | Human-in-the-loop | Observability | Model choice |
|---|---|---|---|---|---|---|
| LangGraph | MIT | Python, JS/TS | Checkpointers per thread | interrupt() and resume | LangSmith | Any provider |
| CrewAI | MIT | Python | Flow state. Checkpointing listed, storage not described | Supported | Tracing and OpenTelemetry listed on the CrewAI platform | OpenAI default, others via connectors |
| Microsoft Agent Framework | MIT | Python, .NET (1.0); Go SDK in a separate repo | Checkpointing, pause and resume, durable extension | Approvals built in | OpenTelemetry built in | Seven first-party connectors |
| OpenAI Agents SDK | MIT | Python, JS/TS | Via Dapr, Temporal, Restate or DBOS, or the managed Agents API | Built in | Built-in tracing | 100+ via LiteLLM. Hosted tools are OpenAI-only |
| LlamaIndex | MIT | Python (TS deprecated) | Pluggable, file or database | Via agents server | Not stated in README | 300+ integration packages |
| AutoGen | MIT code, CC-BY-4.0 docs | Python, .NET | — | — | — | Archived. Do not start here |
Tracing costs are compared in Langfuse vs LangSmith vs Braintrust.
Worked example: hosting one production agent
Assumptions: one agent, deployed all month (730 hours), handling 30,000 runs a month with one web search per run, and one developer seat. Model tokens are excluded because they depend on the model. Prices are monthly list prices.
LangSmith Deployment, Plus plan: one seat is $39, and the first Small serverless deployment is free. Plus includes 10,000 base traces a month; if all 30,000 runs are traced, the other 20,000 bill at 0.005 LSU each, which is $100. Total: $139 per month, or $39 with tracing sampled below 10,000 traces.
LangSmith, a second Small serverless deployment kept running: the billing docs price Small (0.8 vCPU, 2 GiB) at 0.072 LSU per running hour. 0.072 × 730 = 52.56 LSU, which is $52.56. Serverless can scale to zero, so this is the ceiling, not the floor.
LangSmith, Small dedicated deployment: the runtime minimum is 3 vCPU and 6 GiB, so (3 × 0.0675) + (6 × 0.009) = 0.2565 LSU per hour. The database is 0.277 LSU per hour. 0.5335 × 730 = $389.46 per month, and the docs say dedicated "runs continuously".
Foundry hosted agent, East US, sized at 1 vCPU and 2 GiB: (1 × $0.0994) + (2 × $0.0118) = $0.123 per hour. × 730 = $89.79. Foundry lists no container sizes, so this size is illustrative.
CrewAI: 30,000 runs is far beyond Basic's 50 executions, and Enterprise is unpriced.
OpenAI: the SDK runs on your own infrastructure. The managed Agents API publishes no platform fee; you pay model, tool and container rates.
The web-search tool fee is the same whichever framework calls it. On OpenAI it is 30,000 × $10 / 1,000 = $300, plus search content tokens at model rates. On Foundry's Web Search tool it is 30,000 × $14 / 1,000 = $420.
For one agent, tool fees ($300 to $420), tracing and model tokens outweigh hosting, so model those first.
When a no-code platform is the better answer
If the agent lives inside Microsoft 365, answers employees, and calls connectors that already exist, Copilot Studio, one of the enterprise assistant platforms, usually beats any code framework on time to ship. That is what people searching "LangGraph vs Copilot Studio" are really asking. Copilot Studio costs $200 per 25,000 credits per month, and licensed Microsoft 365 Copilot users can "build and use internal agents within Microsoft 365" with the Copilot Chat and Standard harness at no additional cost. Pick a code framework when you need custom state, non-Microsoft hosting, or a customer-facing agent. Credit-based billing across Agentforce, Copilot Studio and others is broken down in Five Billing Units for One Enterprise Agent.
When not to use each
LangGraph: skip it for a single prompt plus two tools; the graph and checkpointer are overhead. LangSmith dedicated deployments bill around the clock.
CrewAI: avoid it if your team writes TypeScript or .NET. Also avoid it if you need a published price before you sign, because CrewAI publishes no Enterprise price or overage rate.
Microsoft Agent Framework: avoid it if you need A2A 1.0 today; Microsoft says it is "coming soon". Avoid it too if you are moving away from Azure, because the hosted runtime is Foundry.
OpenAI Agents SDK: on its own it does not persist runs; for durability add Dapr, Temporal, Restate or DBOS, or move to the Agents API. Avoid both if you need provider-neutral built-in tools.
LlamaIndex: avoid it for general agents in TypeScript, since the TS port is deprecated.
AutoGen: don't start new work on it. Microsoft says it provides no security fixes.
Frequently asked questions
Is AutoGen still maintained?
No. The AutoGen README says the project is archived and that Microsoft provides no bug fixes, security fixes, new features or support. Microsoft directs new and existing users to Microsoft Agent Framework.
Has Semantic Kernel been replaced by Microsoft Agent Framework?
Yes. The Semantic Kernel README calls Microsoft Agent Framework "the enterprise-ready successor to Semantic Kernel". Agent Framework reached version 1.0 for .NET and Python on 3 April 2026, and Microsoft publishes a migration guide.
What is the difference between LangGraph and CrewAI?
LangGraph has you define an explicit graph with persisted state, checkpointers and interrupt() for human approval, in Python or JS/TS. CrewAI is Python-only and organises work as role-based agents in crews, with Flows for deterministic steps. Both are MIT-licensed.
Does the OpenAI Agents SDK only work with OpenAI models?
No. The SDK says it is provider-agnostic and reaches more than 100 other LLMs through LiteLLM or any-llm. Its hosted tools, such as web search at $10 per 1,000 calls, are OpenAI services billed on OpenAI's price list.
How much does LangGraph Platform cost?
LangChain sells managed LangGraph hosting as LangSmith Deployment. The Plus plan is $39 per seat per month and includes one free Small serverless deployment. Additional Small serverless deployments use 0.072 LSU ($0.072) per running hour, about $52.56 for a full month, and traces beyond 10,000 a month cost 0.005 LSU each.
Should I use LangGraph or Copilot Studio?
Use Copilot Studio when the agent serves Microsoft 365 users through existing connectors. It costs $200 per 25,000 credits per month, and licensed Microsoft 365 Copilot users can build internal agents at no extra cost. Use LangGraph when you need custom state, approval steps, or a customer-facing agent hosted outside Microsoft.
Next step
Compare every framework we track in Agent Development tools. Before your first agent ships, set up an agent evaluation and regression harness so that a framework or model change cannot break it silently.
Related Workflows
AI Spend Guardrails & Cost Allocation (FinOps for Agents)
Put a gateway in front of every model call so each token is tagged to a team and a use case, routed to the cheapest adequate model, capped by budget, and alerted on before the vendor cuts you off.
View workflowBrowser-Using Agent for Vendor Research & Pricing Sweeps
Agentic browser session that visits 50 vendor sites, captures pricing, terms, and feature differences, and produces a comparison sheet — overnight.
View workflowPredictive Maintenance Alerting
Monitor equipment sensor data and production metrics to predict failures before they happen — reducing downtime and emergency repair costs.
View workflowKeep Reading
Vapi vs Bland vs Retell (2026): AI Voice Agent Pricing, Limits and Real Monthly Cost
Vapi, Retell and Bland price AI phone agents in three different ways. We read each vendor's own pricing page and worked out what 2,000 inbound minutes a month actually costs on each.
Composio vs Arcade vs MCP in 2026: Agent Tool Integration Pricing, Auth and Security
The managed tool layers for AI agents price the same action four different ways. Here is what Composio, Arcade.dev and Pipedream Connect publish, what running your own MCP servers costs you instead, and a worked monthly bill.
Agentforce vs. Copilot Studio vs. Gemini Enterprise vs. OpenAI Agents API vs. Claude Managed Agents: Five Billing Units for One Enterprise Agent
Salesforce bills actions or conversations, Microsoft bills credits and switches you off at 125%, Google bills seats, OpenAI bills tokens plus a sandbox, Anthropic bills tokens plus session-hours. One 900-conversation-a-day agent, priced on each where the published numbers allow.
Found this helpful?
Get weekly AI workflow ideas in your inbox.