What an AI agent really costs: we ran the numbers on Anthropic, Kimi, OpenAI, DeepSeek

What this is about

For a long time, AI labs competed mainly on their models. Since fall 2026, they've also been selling the runtime environment, known in the industry as the "harness": the software wrapped around the model. That includes a sandbox (an isolated machine where the agent runs code), tools, memory, logs and error handling. Only with all of that in place can an agent work autonomously for days or weeks.

The week around November 5 alone was busy. Moonshot officially launched Kimi Hosted Agents. That same evening, DeepSeek published user numbers for DeepSeek Harness for the first time and announced a plugin marketplace. AWS added crew bots to Kiro, and Perplexity launched Teammates. Western comparisons such as The New Stack's ("the harness is the product … they disagree on the price") cover Anthropic, OpenAI, Google and Microsoft. The Chinese providers are missing. So we ran the numbers on a concrete scenario to see what each provider charges and where you end up tied to a vendor.

The offerings at a glance

Runtime: small amounts, plenty of fine print

Anthropic. The official pricing page lists $0.08 per session hour. According to Anthropic, this time is "measured to the millisecond and only accrues while the session's status is running". When the agent is waiting for the next message or an approval (idle), no fee applies. The same goes for the rescheduling and terminated statuses. The often-quoted $58 a month is therefore a ceiling for an agent that computes without a break. An agent that spends most of its time waiting costs far less. The fee replaces the container hours you'd otherwise pay for code execution ($0.05/hour after 1,550 free hours). Batch discounts don't apply within sessions. Anthropic plans to offer Managed Agents only through its own API, not through partner clouds such as Bedrock.

Kimi. The official formula on the pricing page is: sandbox hours × ¥0.072 + model costs + plugin calls × unit price. According to Sina, a sandbox running for 30 days costs ¥51.84. At the November 5 exchange rate (6.667 yuan per dollar), that's $7.78, about one-seventh of Anthropic's maximum rate. According to Biggo Finance, the sandbox has just 1 CPU core and 1 GB of RAM; we found no mention of this on the official site. Whether idle time is billed isn't documented. Plugins cost between ¥0.001 and ¥0.45 per call: a web search ¥0.03, an image search ¥0.045, some financial data plugins ¥0.35 or ¥0.45. At launch there are more than 140 skills, over 30 plugins with data sources such as Wind, iFinD and S&P Global, and eight templates for finance agents. The beta had been running since September for verified enterprise customers in China.

OpenAI. OpenAI makes no money directly from the Agents SDK or the Agents API (public beta since September 10). The forum states: "There are no additional fees for using the Agents API." But if you use the sandbox hosted by OpenAI, you pay container rates. According to the pricing page, that's $0.03 (1 GB) to $1.92 (64 GB) per 20-minute session. A footnote, however, mentions per-minute billing with a 5-minute minimum for "eligible" sessions. According to tokencost.app, whether idle time counts hasn't been published. Alternatively, you can bring your own sandbox or use partners such as DigitalOcean, E2B, Modal or Cloudflare.

For comparison: pure infrastructure. DigitalOcean charges $0.0535 per hour for 1 vCPU and 1 GB for its harness runtime. CPU costs drop to zero while waiting; only memory ($0.0095/GB-hour) keeps running.

AWS Kiro. Crew bots are named, long-lived agents with their own memory, for tasks like build monitoring or code review. An AWS spokesperson told The Register: "All Kiro bots are billed together through the existing Kiro plan." According to kiro.dev, plans cost $20 to $200 with 1,000 to 10,000 credits, and each additional credit is $0.04. How many tokens a credit is worth depends on the model, so there's no way to convert it to dollars per hour. The Register warns that because every bot has its own memory, the same context can show up twice in your costs.

Perplexity. Teammates is only available on the $200 Max plan. According to the help page, it includes 10,000 computer credits a month, and the agents draw from those.

Our sample calculation

Scenario: Agent A runs 24/7 for 30 days (720 hours). Agent B works two hours a day (60 hours). That's 780 agent hours in total. For tokens, we assume 120 million input tokens a month, 80 percent of them cache hits (previously processed context that's reused at a lower price), plus 6 million output tokens. For simplicity, we left out cache write costs. The assumptions are ours; the prices come from the official pricing pages. Euro figures are based on the ECB rate of November 5 ($1.1224).

Web searches come on top. At 900 searches a month, that's $9 each at Anthropic and OpenAI ($10 per 1,000), about $4 at Kimi and $0.68 at DeepSeek ($0.75 per 1,000).

One important caveat: the table compares equal token volumes, not equal work. According to Anthropic, Claude 4.7 and newer models produce about 30 percent more tokens for the same text with their new tokenizer. Kimi K3.1 Pro only runs in thinking mode, set to the highest level by default, which costs more output tokens. An independent test by the Prompt Engineering channel still found Kimi about 55 percent cheaper than Opus 5.5, but also more error-prone on precision tasks.

Our take: The runtime fee is the smallest line item. Even in the worst case, it makes up a third to a fifth of the bill at Anthropic and about six percent at Kimi. An analysis by Scott Armbruster in April came to a similar conclusion: Anthropic's runtime fee amounts to 10 to 15 percent of a typical session bill. The big differences come from model pricing.

Where the lock-in hides

Lock-in means you're tied to a vendor in a way that makes switching expensive or painful. The runtime is cheap; the ties come from elsewhere.

• Session state: At Anthropic, every session lives as a continuous event log on Anthropic's servers and only works with Claude models. Verdent's guide calls this the real price: switching is "not trivial". OpenAI's Agents API also manages context and long sessions itself. According to third-party reports, it offers only US data residency and no Zero Data Retention, meaning no guarantee that data won't be stored. Check this before you deploy.
• Plugin marketplaces: Kimi's edge is licensed financial data (Wind, iFinD, S&P) that you can't simply take with you. DeepSeek is planning an official marketplace. According to IT Home, there are already around 35,000 plugins, and about 60 percent of active users have built or installed at least one.
• Security hooks: Kimi's "Inference Hook" checks inputs and tool outputs against company-specific rules before every model request. But it's only available by invitation to verified enterprise customers. Control layers like this are hard to replace once compliance processes depend on them.
• Usage terms: According to its pricing page, Kiro prohibits third-party harnesses such as OpenClaw that route requests around Kiro's own interfaces. Kiro Crew is Apache 2.0-licensed, but reportedly runs on the proprietary Kiro CLI.

What you can self-host

The most open option is DeepSeek Harness: MIT license, around 274,000 stars on GitHub, launched locally via npx. Built-in providers are DeepSeek, Anthropic and OpenAI, plus any OpenAI-compatible interface, all the way down to local models via Ollama. But the project is a developer preview, and the README warns of breaking changes to existing setups. Here, the lock-in comes from the ecosystem. The 1.54 million user figure counts only users of the official DeepSeek API, and the announced rewards for plugin authors are paid out in DeepSeek tokens. Do the math: 92 trillion tokens a day divided by 1.54 million users comes to around 60 million tokens per user per day. So the free harness is first and foremost a sales channel for the API. On September 30, the figure stood at about 271,000 daily users.

OpenAI's Agents SDK can run with your own sandbox or in your own VPC, an isolated network in your own cloud. GitHub took a different route: it ported the Copilot runtime from around 430,000 lines of TypeScript to 832,378 lines of Rust. According to the GitHub blog, the goal was to embed it directly in applications without Node.js, from VS Code to Word and Outlook. According to a GeekNews summary of the related talk, the port consumed about 136 billion tokens, roughly $120,000 at list prices. GitHub hasn't announced any pricing changes for Copilot.

Our verdict

If you only look at hourly rates, the picture is clear: Kimi costs about one-seventh of Anthropic, and OpenAI and DeepSeek charge nothing for the runtime. But these rates measure different things. Anthropic bills only active compute time, Kimi bills the sandbox runtime of a very small machine, and at OpenAI, idle billing is unclear. Our take: The model determines your monthly bill, not the runtime. Your long-term costs depend on where your sessions, memory and plugins live.

For teams in Europe, there's more to consider: during the beta, Kimi Hosted Agents targeted Chinese enterprise customers, and the data question applies to every hosted offering. If you want to stay flexible, your safest bet is an open harness (DeepSeek, OpenAI SDK) with swappable models. The trade-off is that you handle operations and security yourself. One open question is how Microsoft will bill: according to Sina, billing for its Copilot runtime in Foundry starts on December 1, with resource-based pricing announced.