AI agent platforms in 2026: what a business should choose

How to read this overview

Agent platforms fall into four classes, and comparing them head to head makes no sense. Orchestrators wire services together. Frameworks give you code and leave the infrastructure to you. Self-hostable visual platforms give you a builder inside your own perimeter. Enterprise ones are tied to an ecosystem and sold with it.

For each we look at three things: what is being charged for, where the ceiling runs and whether it works from Russia. A price without the first item says nothing: €20 for 2,500 executions and €20 for 2,500 messages are two different products.

The year's main news: OpenAI shut down two products from this list

Worth starting here, because many overviews still recommend things that no longer exist. Per OpenAI's official deprecations page: the Assistants API was shut off on 26 August 2026, replaced by the Responses API and Conversations API. Agent Builder, the visual canvas inside AgentKit, was deprecated on 3 June 2026 with shutdown set for 30 November 2026; the replacement is the Agents SDK or ChatGPT Workspace Agents.

The practical lesson is wider than one vendor. The layer you assemble an agent on has a life of its own and can disappear within a year. So it is worth separating, in the architecture, what belongs to you — the process, the tools, the data, the authority rules — from what you are renting.

Orchestrators: n8n, Make, Zapier

These are not agent platforms by birth — they are automation engines with AI nodes bolted on. Strong where the task is “wire five services together”, and predictably weaker where access rights, idempotency and tests are required.

• n8n. Counts whole-workflow executions: Starter €20/month billed annually for 2,500 executions and 5 concurrent, Pro €50 for 10,000 and 20 concurrent, Business €667 for 40,000 and 30 concurrent. No permanent free tier; the trial is 14 days and 1,000 executions. Self-hosting is free under the Sustainable Use License, but the Community edition has no SSO, environments, Git versioning or external secrets.
• Make. Since 19 January 2026 the billing unit is credits (formerly operations, converted one to one). Free gives 1,000 credits a month, at most 2 active scenarios and a minimum 15-minute interval; Core is $12, Pro $21, Teams $38 a month with 10,000 credits and a one-minute interval. AI agents are billed twice over: 1 credit per operation plus credits for provider tokens. No self-hosting.
• Zapier. The largest integration catalogue and a separate agent line with its own currency: Free gives 400 activities a month and at most 10 per run, Pro is $400 billed annually for 1,500 activities and 40 per run. An activity counts for a trigger, a knowledge-base answer, an action, web browsing and web search. The main platform counts tasks: Free 100 a month, Professional from $19.99.

Frameworks: LangGraph, CrewAI, OpenAI Agents SDK

Here you get code and full freedom, and with it all the infrastructure work. The vendor earns not on the framework but on the observability and hosting around it — and that is exactly where the billing surprises hide.

• LangGraph. The framework itself is MIT and self-hosting is free. The paid layer is LangSmith: Developer at $0 per seat with 5,000 base traces a month, Plus at $39 per seat per month. A base trace costs 0.05¢, an extended one 0.5¢ — ten times more — and the upgrade switches on automatically with online evaluators or an automation rule. Self-hosted and hybrid deployments are Enterprise-only, and a standalone server requires a licence key.
• CrewAI. MIT-licensed framework. The commercial tier: Basic is free with 50 executions a month and 2 automations, no self-hosting; after that it is Enterprise on request, deployable in their cloud, your VPC or your own infrastructure. There are no mid-tier plans with a public price. Fifty executions a month is demo scale, not pilot scale.
• OpenAI Agents SDK. MIT-licensed and provider-agnostic: it works with Responses and Chat Completions as well as other models. There is no separate AgentKit fee — per the announcement the tools are included in standard API pricing, so you pay for tokens and tool calls. The main risk here is not technical: Russia is not on OpenAI's list of supported countries, and the company explicitly warns that access from unsupported countries may lead to account suspension.

Visual platforms in your own perimeter: Flowise and Dify

A compromise between a builder and code: a visual editor you can host yourself. Both licences look open and both carry commercial caveats, which are worth reading before rather than after.

• Flowise. Core code under Apache 2.0, the enterprise directory under a commercial licence. Cloud: Free gives 2 flows and 100 predictions a month, Starter is $35 for 10,000 predictions, Pro $65 for 50,000 plus $15 per user beyond five. Self-hosting the core is free and removes the payment question from Russia entirely.
• Dify. A modified Apache 2.0 licence with two restrictions: you may not use the source to operate a multi-tenant environment without written permission, and you may not remove the logo from the console and apps. Cloud: Sandbox is free with 200 message credits, Professional is $590 per workspace per year, Team $1,590. The Community edition is free, self-hosted, one workspace. The multi-tenancy ban is a hard block for anyone planning to build their own SaaS on Dify.

The enterprise contour: Microsoft Copilot Studio

A class of its own: agents welded to Power Platform, Dataverse and the Microsoft 365 tenant. Billing in its own currency: pay-as-you-go at $0.01 per message, or a prepaid pack at $200 per tenant per month for 25,000 Copilot Credits.

The word “message” is misleading here: per the official table a classic answer costs 1 credit, a generative one 2, an agent action 5, tenant graph grounding 10, and a single generative answer with graph grounding debits 12 credits at once. Separately, Microsoft suspended new sales of products and services in Russia in a statement dated 4 March 2022.

The Russian contour: Yandex AI Studio and GigaChat

For a company paying in roubles and unwilling to depend on someone else's jurisdiction, the choice today is real rather than forced.

• Yandex AI Studio. Resource-based pricing, no subscription. Per the official pricing page: YandexGPT Lite ₽0.2 per 1,000 tokens, YandexGPT Pro 5.1 ₽0.8, Alice AI LLM Flash ₽0.1 on input. Agent tool calls: web search ₽915 per 1,000 requests, file search ₽300, code execution and MCP not billed. Voice: inbound audio ₽0.0264 per second. A useful detail: a request that fails with an internal server error is not billed.
• GigaChat. Every model supports function calling, so an agent scenario is buildable. Pay-as-you-go for legal entities: Lite ₽0.065 per 1,000 tokens, Pro ₽0.5, Max ₽0.65, with a ₽600 monthly minimum; packages too — for example 300M Lite tokens for ₽19,500 a month. Individuals get a twelve-month freemium totalling 365M tokens, but generation runs single-threaded, so concurrent load cannot be tested on it.
• A caveat on the Yandex figures: the documentation page is closed to automated requests, so these numbers come from an archived snapshot of the official page dated 7 August 2026, and the pricing rules changed twice during 2026. Verify by hand before budgeting.

What the research says — and how not to misquote it

Two Gartner figures bear directly on the choice. Press release of 25 June 2025: over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value and inadequate risk controls. Press release of 26 May 2026: by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after production incidents.

A MIT report is often quoted alongside: 95% of generative AI pilots with no measurable P&L impact. The wording there is about the absence of measurable impact, not about “never reaching production”, and it covers generative AI broadly rather than agents. Swapping one for the other is a mistake — it is where trust in the rest of the text breaks.

A second signal from the same Gartner release: of the thousands of vendors calling themselves agentic, analysts count roughly 130 as real. The rest is repackaged assistants, RPA and chatbots. In practice that means one rule: do not ask “do you have agents”, ask “what exactly does the agent do inside my system, and under whose permissions”.

Five questions that settle the choice

Answer these before comparing plans — the list will shorten on its own.

• Does the agent act or answer? If it changes data you need an authority policy, a log and rollback. No platform provides those for you — that part is yours in every scenario.
• Where does the data live? A requirement to keep documents inside your perimeter immediately narrows you to self-hosting: n8n Community, Flowise, Dify Community, or Russian clouds.
• How many events at peak, not on average? Concurrency is fixed by plan: 5, 20 or 30 simultaneous executions on n8n; on Make the webhook queue is capped by licensed credits; on Zapier a hundred simultaneous events trigger flood protection and are replayed at one per second.
• Who pays and with what? From Russia, Visa and Mastercard cards issued by Russian banks have not worked abroad since 5 March 2022 per both networks' own statements. None of the foreign vendors on this list publishes anything about Russian cards — payment needs planning up front, not at launch.
• What do you lose if the platform disappears? The Assistants API story shows this is not hypothetical. Logic assembled in someone else's editor does not export: decide up front which part of the system is allowed to live there.