AI agent platforms compared: OpenAI, Anthropic, Google
All three sell a model that calls your tools in a loop, plus a framework to keep that loop in check. The difference that shows up on your invoice is not the model price but the extras: hosted search, sandbox time, and in Google's case the server the agent runs on.
OpenAI, Anthropic and Google all sell roughly the same thing: a model that can call your tools in a loop, plus a framework to keep that loop from running away. The real differences are in how they bill the loop — per token in every case, with extra charges bolted on for hosted search, sandboxes and, in Google's case, the server the agent runs on.
Prices below were checked against each vendor's own pricing page in September 2026. They change often, so treat the numbers as a snapshot with a date on it, not a permanent rate card.
What "agent platform" actually means here
An agent, in the sense all three vendors use the word, is a program that gets a goal, decides which tool to call next, calls it, reads the result, and repeats until it thinks it's done. The vendor supplies the model and usually a framework that manages the loop, the conversation state and the permission checks. You supply the tools — your CRM, your database, your invoicing system.
That structure matters for your bill. A single "task" is not one API call. It is a chain of calls, each one carrying the growing conversation back into the model as input tokens. Agents are input-heavy, which is why every vendor now sells caching discounts.
OpenAI: Agents SDK and the Responses API
OpenAI documents two paths. The Agents SDK is a code-first framework in Python and TypeScript that runs the loop for you and adds handoffs between specialist agents, input and output guardrails with resumable approval flows, session state and tracing across model calls and tools. The Responses API is the lower-level option for teams that want to write their own orchestration. Both can call Model Context Protocol servers and platform-hosted tools, per OpenAI's agents guide.
On the API pricing page, as of September 2026, GPT-5.6-Sol is listed at $2.00 per million input tokens and $10.00 per million output; GPT-5.6-Terra at $1.00 and $6.00; GPT-5.6-Luna at $0.10 and $0.60. Long-context requests are priced higher — $4.00 input and $15.00 output on Sol. Cached input tokens are billed at 10% of the standard input rate.
Hosted tools are billed separately. Web search is listed at $10 per 1,000 calls on top of the model tokens the results consume. File search costs $2.50 per 1,000 tool calls plus $0.10 per GB-day of storage, with the first gigabyte free. The Code Interpreter and hosted shell are billed per session, listed in a range from $0.03 to $1.92 depending on container size.
Anthropic: Agent SDK, Client SDK and Managed Agents
Anthropic splits the same territory three ways. The Claude Agent SDK is a Python and TypeScript library that runs the agent loop inside your own process, with built-in file and command tools, web search, subagents, hooks, MCP connections and a permission layer that decides which tools run automatically and which need approval. The Client SDK is for teams writing their own loop. Managed Agents is a hosted REST product where Anthropic runs the agent and the sandbox. Other languages have to shell out to the CLI as a subprocess.
Per the Claude pricing documentation, checked September 2026: Claude Opus 5 is $5 per million input tokens and $25 output. Sonnet 5 is $2 and $10. Haiku 4.5 is $1 and $5. Batch processing takes 50% off, and cache reads cost 10% of the base input price, with a 1.25x surcharge on the write for a five-minute cache or 2x for a one-hour cache.
Tool charges: web search is $10 per 1,000 searches plus tokens. Web fetch adds no charge beyond tokens. Code execution is free when used alongside web search or web fetch; used on its own it comes with 1,550 free container-hours a month per organization, then $0.05 per container-hour. Managed Agents adds $0.08 per session-hour of runtime on top of standard token rates.
The toolset definitions themselves cost tokens. Anthropic's docs put computer use at roughly 4,500 extra input tokens per request just to describe the tools, and browser use at roughly 6,600, before any screenshots are counted.
Google: ADK and Agent Engine
Google's Agent Development Kit is open source under Apache 2.0 and, unusually, ships in four languages — Python, TypeScript, Go and Java. It covers multi-agent orchestration, a third-party tool ecosystem, built-in evaluation, and both fixed pipelines and dynamic routing. Agents can run locally or deploy to Agent Runtime, Cloud Run or Google Kubernetes Engine.
Model tokens are billed through the Gemini API pricing page. As of September 2026, Gemini 3.8 Flash is $0.75 per million input and $3.75 per million output through December 31, 2026, rising to $1.50 and $7.50 on January 1, 2027 — a scheduled increase worth planning around. Gemini 3.5 Flash-Lite is $0.30 and $2.50. Gemini 3.1 Pro Preview is $2.00 input and $12.00 output up to 200k tokens, and $4.00 and $18.00 above that. Search grounding gives 5,000 free requests a month, then $14 per 1,000.
Google is the only one of the three that also bills you for the machine. Vertex AI pricing lists Agent Engine compute at $0.085 per vCPU-hour after 50 free vCPU-hours a month, memory at $0.009 per GiB-hour after 100 free GiB-hours, and Sessions and Memory Bank storage at $0.30 per GiB-month, effective September 1, 2026. Usage is rounded to the nearest second, and idle time spent waiting for the next prompt between turns is not billed.
Side by side
| Item | OpenAI | Anthropic | |
|---|---|---|---|
| Mid-tier model, input / output per 1M tokens | GPT-5.6-Terra $1 / $6 | Sonnet 5 $2 / $10 | Gemini 3.8 Flash $0.75 / $3.75 (to Dec 31, 2026) |
| Framework languages | Python, TypeScript | Python, TypeScript | Python, TypeScript, Go, Java |
| Hosted web search | $10 / 1,000 calls | $10 / 1,000 searches | 5,000 free/month, then $14 / 1,000 |
| Cache read discount | 10% of input rate | 10% of input rate | Published per model |
| Separate runtime charge | Per code-interpreter session | $0.08/session-hour (Managed Agents) | $0.085/vCPU-hour after free tier |
What a real task costs
Take an agent that reads a long vendor contract and returns a structured summary: call it 30,000 input tokens and 2,000 output tokens per run. On the September 2026 rates above, that is about $0.042 on GPT-5.6-Terra, about $0.08 on Claude Sonnet 5, and about $0.03 on Gemini 3.8 Flash. Run it a thousand times a month and you are looking at $30 to $80.
That arithmetic is the easy part. The number that surprises people is the loop. An agent that makes eight tool calls resends the whole conversation each time, so a task that looks like 30,000 tokens on paper can bill closer to 150,000. Caching is the lever: at 10% on cache reads, a stable system prompt and tool schema stop being the expensive part of every step.
Model price is the number on the page. The agent loop is the number on the invoice.
How to choose without a bake-off
If your team already writes Go or Java, Google's ADK is the only one of the three with first-party libraries in those languages, and everyone else is shelling out to a subprocess. If you want someone else to run the sandbox and the session store, Anthropic's Managed Agents and Google's Agent Engine both do that; OpenAI's hosted piece is narrower, covering search, file search and the code interpreter rather than the whole agent.
If cost is the deciding factor, price the cheap tier, not the flagship. The gap between Gemini 3.5 Flash-Lite at $0.30 input and Claude Opus 5 at $5 input is more than sixteen times, and a lot of back-office agent work — classifying, extracting, routing — does not need the top model.
One more practical note: all three publish price changes on the same pages linked above, and Google has already posted a specific increase for January 1, 2027. Put a calendar reminder on whichever page you depend on, and re-check before you sign anything that assumes today's rate.