AI agent platforms compared: OpenAI, Anthropic, Google

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 on October 9, 2026. The models this article led with in September are still on the price lists, but neither OpenAI nor Anthropic opens its table with them any more, and OpenAI has added a hosted agent runtime it did not document then. 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, in all three cases here, 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: four entry points, one of them hosted

OpenAI's agents guide now lists four starting points rather than two. The Agents API is the new one and the one that changes the comparison: the page says OpenAI runs a managed Codex harness, saves the agent's progress between tasks, and offers an OpenAI-hosted sandbox, a self-hosted sandbox or none. The Agents SDK runs inside your own application and handles the agent loop and handoffs; its own page offers TypeScript and Python repositories and describes orchestration and handoffs, guardrails and human review, sandbox agents, results and state, and MCP servers. The Responses API is for calling models directly or building an agent from scratch. ChatKit is an embedded chat experience. The guide rates integration effort as low, medium and high across the first three.

The API pricing page opens on three flagship models. Per million tokens, short context: gpt-6-astra at $10.00 input and $50.00 output, gpt-6.1-sol at $2.00 and $10.00, gpt-6-luna at $0.10 and $0.50. Long context is a separate, higher column on the same table: $20.00 and $75.00 on astra, $4.00 and $15.00 on sol, $0.20 and $0.75 on luna. The GPT-5.6 generation has not been withdrawn — clicking "All models" expands the same table to show gpt-5.6-sol at $4.00 and $20.00, gpt-5.6-terra at $2.00 and $12.00, and gpt-5.6-luna at $0.20 and $1.20, and the page still prints the caveat "GPT-5.6 Sol’s promotional pricing is available at least through November 21, 2026."

One rule of thumb this article used in September no longer covers the whole table. Cached input is still 10% of the standard input rate on every row that prints one — astra, luna, gpt-6-sol, gpt-5.5, gpt-5.4 and the whole GPT-5.6 generation — with one exception: gpt-6.1-sol prices cached input at $0.10 against a $2.00 input rate, which is 5%. Two older rows, gpt-5.5-pro and gpt-5.4-pro, print no cached-input price at all, so the discount does not apply to them in either direction.

The same table also carries five speed tiers — Standard, Batch, Flex, Fast and Ultrafast — and they are not rounding differences. On gpt-6-astra, every one of the eight price columns comes out at exactly half of Standard under Batch, exactly double under Fast, and exactly six times under Ultrafast: $60.00 input and $300.00 output per million. The page notes that "Priority processing was renamed Fast mode on July 30, 2026."

Hosted tools are billed separately. Web search has four rows. The two marked "(all models)", one of them for image search, are $10 per 1,000 calls with search content tokens billed at model rates on top, as is the preview row for reasoning models; the preview row for non-reasoning models costs $25 per 1,000 calls instead but states that "Search content tokens are free." File search costs $2.50 per 1,000 tool calls plus $0.10 per GB-day of storage, with the first gigabyte free. The hosted shell and Code Interpreter are billed per 20-minute session per container, from $0.03 for 1 GB to $1.92 for 64 GB. One line to read before picking an endpoint: "Regional processing (data residency) endpoints are charged a 10% uplift for models released on or after March 5, 2026." FedRAMP endpoints carry the same uplift. If data residency is a requirement rather than a preference, see what data residency actually commits you to.

Anthropic: Agent SDK, CLI, Client SDK and Managed Agents

Anthropic splits the same territory four ways, and its Agent SDK overview sets them out in one table. The Agent SDK is a library you embed in your own Python or TypeScript application, in a process you operate; the page lists built-in tools, hooks, subagents, MCP, permissions, sessions, skills and plugins as the capabilities it carries over from the CLI. The CLI itself is the terminal interface. The Client SDK is direct API access where you write the tool loop yourself. Managed Agents is the hosted option: Anthropic runs the agent loop, with sessions either in an Anthropic-managed cloud sandbox or in a self-hosted one on your own infrastructure, reachable from a language SDK, a CLI or the REST API. For any language other than Python or TypeScript, the documented route is running the CLI as a subprocess.

The Claude pricing documentation, checked October 9, 2026, leads with four models. Per million tokens: Claude Fable 5.1 at $10 input and $50 output, described on the page as being for "demanding reasoning and long-horizon agentic work"; Opus 5.5 at $4 and $20; Sonnet 5.5 at $2 and $10; and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens, rising to $0.50 and $2.50 for prompts over that. An "Additional models" block underneath keeps sixteen further rows on the same page. Among them: Mythos 5.1, which matches Fable 5.1 in all five columns; Fable 5 and Mythos 5, also at $10 and $50; Opus 5 at $5 and $25; Sonnet 5 at $2 and $10; and Haiku 4.5 at $1 and $5.

Two things to notice before you treat the new line as a straight price cut. First, cache reads are no longer a flat 10% of input across the board: the table prints an explicit per-model rate, and it works out at 2.5% on Fable 5.1, 5% on Opus 5.5 and Sonnet 5.5, and 10% on Haiku 5.5 in both of its bands. The 1.25x and 2x surcharges on five-minute and one-hour cache writes do still hold on all four.

Tool charges: web search is $10 per 1,000 searches plus tokens, and a search that errors is not billed. Code execution is free when used alongside web search or web fetch. Used without them it is billed by execution time, with a five-minute minimum, 1,550 free hours of usage a month per organization and $0.05 per hour per container after that — and the page warns that if files are included in the request, execution time is billed even when the tool is never called, because the files are preloaded onto the container. Managed Agents adds $0.08 per session-hour on top of standard token rates, metered only while the session is actually running rather than idle, and it replaces container-hour billing rather than stacking on top of it. Anthropic applies its own residency surcharge too: pinning a Managed Agents inference region to the US bills tokens at 1.1x the standard rates.

Google: ADK and Runtime

Google's Agent Development Kit is described on its own page as an open-source agent development framework, and it is the only one of the three with first-party libraries in four languages: the page states that "ADK is available in 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 Runtime, Cloud Run or Google Kubernetes Engine.

Model tokens are billed through the Gemini API pricing page, whose Google rows were re-read on October 28, 2026. Gemini 4 Argon now heads the list, and it carries the nearer of the two scheduled increases on this page: $2.00 per million input and $10.00 per million output through November 30, 2026, then $4.00 and $20.00 starting December 1. It is also the only model in this article with no free tier at all — every cell in its free column reads "Not available" — and it supports implicit caching only, at $0.10, with no explicit-cache option and no hourly storage charge. 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. Gemini 3.5 Flash-Lite is $0.30 and $2.50. Gemini 3.1 Pro Preview, the Pro model on the list, is $2.00 input and $12.00 output for prompts up to 200k tokens, and $4.00 and $18.00 above that. Search grounding gives "5,000 free search requests per month (shared across all Gemini 3 and newer models), then $14 per 1,000 requests."

Google is the only one of the three that also bills you for the machine, and the page that documents it has moved: the old Vertex AI pricing URL now redirects to Gemini platform pricing, where agent hosting is billed on three SKUs. Agent Compute is $0.085 per vCPU-hour after 50 free hours a month per account. Agent Memory is $0.009 per GiB-hour after 100 free GiB-hours. Agent Storage is metered at $0.000410959 per gibibyte-hour after the first free gibibyte-month, which the page's own Memory Bank section restates as $0.30 per GiB-month. One- and three-year Flexible Savings Plans discount all three; on compute they bring the hourly rate to $0.0765 and $0.068. The page adds that "All usage is rounded to the nearest second. For Runtime, idle time spent waiting for the next prompt between turns is not billed."

The expensive end, and why "the top model" is two questions

Ask which model costs the most and you get a different answer from asking what the current top tier costs, because all three vendors keep older rows on the same page. On OpenAI's expanded table the highest input price belongs to gpt-5.5-pro and gpt-5.4-pro at $30.00 short-context and $60.00 long, with gpt-5.5-pro at $180.00 output. On Anthropic's, Opus 4.1 and Opus 4 are still listed at $15 and $75. Neither is the headline tier any more.

The current headline tier lands in the same place at two of the three vendors. Claude Fable 5.1 and Mythos 5.1 both list at $10 and $50, and gpt-6-astra lists at $10 and $50 short-context. OpenAI's gpt-5.6-cyber sits above them at $12.50 and $75.00, in a separate "Cyber models" table, and it is the only row there with no long-context prices: the page prints a dash in all four of those cells.

Google has not followed the other two up to $10 on input, and the two Google rows above are there for contrast rather than because they qualify. What is worth noticing is where Google does put its headline model. Gemini 4 Argon at $2.00 and $10.00 is priced exactly where the other two put their mid-tier: against the October 9 figures above, those are gpt-6.1-sol's rates to the cent, and Claude Sonnet 5.5's. On December 1 it becomes $4.00 and $20.00 — Claude Opus 5.5's rate on that same reading, and gpt-5.6-sol's. Google's flagship is sold at its competitors' middle price today and at their second tier in five weeks. Gemini 3.1 Pro Preview, meanwhile, carries a separate cache storage charge of $4.50 per million tokens per hour that the other two do not bill that way. What none of that means is that Google is five times cheaper for the same work. Settling that would take a benchmark of the three models on your own workload, and this site does not run benchmarks.

Side by side

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 October 2026 rates above that is about $0.004 on gpt-6-luna or Claude Haiku 5.5, about $0.03 on Gemini 3.8 Flash, about $0.08 on gpt-6.1-sol or Claude Sonnet 5.5, and about $0.40 on gpt-6-astra or Claude Fable 5.1. Run it a thousand times a month and the same job costs you $4 or $400 depending only on which row of the price list you picked. That hundred-fold spread inside a single vendor's table is the number worth taking to a planning meeting; the full per-model arithmetic is here.

That arithmetic is the easy part. The number that moves the total is the loop. An agent that makes eight tool calls resends the whole conversation each time, so the tokens you pay for are a multiple of the tokens in the task, and the multiple grows with every step rather than staying flat. Caching is the lever: at between 2.5% and 10% on cache reads, a stable system prompt and tool schema stop being the expensive part of every step.

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 whose own documentation lists first-party libraries in those languages; OpenAI's Agents SDK page offers TypeScript and Python repositories, and Anthropic's documents running the CLI as a subprocess for anything outside Python and TypeScript. If you want someone else to run the sandbox and the session store, all three now do that — Anthropic's Managed Agents, Google's Runtime, and now OpenAI's Agents API, which its guide did not document in September and which is the change most likely to upset a comparison written earlier in the year.

If cost is the deciding factor, price the cheap tier, not the flagship. Inside Anthropic's own table the gap between Haiku 5.5 at $0.10 input for prompts up to 100,000 tokens and Fable 5.1 at $10 is exactly a hundred times; above that band Haiku 5.5 goes to $0.50 and the gap narrows to twenty. The same shape shows up at OpenAI, where gpt-6-luna and gpt-6-astra are a hundred times apart on input. A lot of back-office agent work — classifying, extracting, routing — does not need the top model.

One more practical note, and this month is the argument for it: all three publish price changes on the same pages linked above, both OpenAI and Anthropic replaced the models at the head of their tables inside six weeks, OpenAI added a hosted runtime in the same stretch, and both of Google's URLs in the September version of this article now redirect somewhere else. 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.