Arcade.dev puts tool search to a 7.6% workflow cost test

The cost of showing an agent its options

Arcade is not a chatbot or a language model. Its platform connects AI agents to workplace software, checks whether users and agents are allowed to take actions, handles credentials and records what the agents do.

The new feature addresses a cost that grows with an agent’s menu of possible actions. Arcade says its catalog contains more than 8,000 tools. Supplying descriptions for the entire catalog would take about 1.75 million tokens before a user even sends a request.

Tool Recommendation searches and ranks tools when an agent needs them, returning a shorter list instead of loading the whole catalog. Small specialized models handle the search; the agent’s underlying model still decides which suggested action to take.

Arcade says the feature works through its MCP gateway, between agents and the tools they can access. Customers do not have to rebuild existing integrations or switch away from Claude, ChatGPT or their own agents to use it.

The approach sits alongside products from Microsoft, Anthropic and OpenAI that also defer loading tool descriptions until they are needed. Arcade’s distinction is that it combines search with a layer for permissions, corporate rules and audit logs across models.

The headline savings are not the whole bill

Arcade’s percentages describe different things, and they should not be treated as interchangeable.

The maximum 37% saving is a calculation of input-token costs for an assistant with 100 predefined tools. For a typical conversation of three user requests and about eight tool calls per request, Arcade estimates savings of about 20%, depending on the model. Sam Partee, Arcade’s co-founder and CTO, gave a similar estimate: 18–20% of tokens in a ten-turn session.

That is an estimate of token reduction in a typical session, not independently verified savings on a company’s full bill. It excludes output tokens and assumes customers keep large tool results out of the conversation context. If those results remain in context, Arcade estimates savings of only 5–7%.

The 91% figure is narrower still: it covers input tokens in function descriptions and recommendation responses during a simulated ten-turn conversation, not the conversation itself or data returned by business systems. In Arcade’s calculation, tokens fall from 256,650 to 22,164 for GPT-5.6, and from 411,729 to 38,170 for Claude Opus 5.

The more complete test covered 408 sessions across 64 workflows using Codex and Claude Code. Arcade reports a 7.6% drop in total token costs, including input and output, with a 95% confidence interval of 2.5% to 12.6% and no change in task completion. The reduction was 9.7% for Claude Code and 5.5% for Codex; the interval for Codex includes zero. VentureBeat did not reproduce the tests independently.

The company also estimates annual savings of about $119,000 for 1,000 Claude Opus 5 users, assuming each has 20 conversations a month. That is a modeled scenario, not a result reported by a named customer.

Partee acknowledged a basic failure mode: repeated searches can consume enough resources to erase the savings or make tool use more expensive. He said such cases are rare, but did not provide independently verifiable data on how often they occur.

The comparison is not a head-to-head test

Arcade also compares its tool search with Microsoft Foundry using ToolRet’s Recall@10 metric, which measures how many needed tools appear in the first ten results. It does not measure whether an agent completes the task.

The comparison uses Arcade’s own benchmark image and Microsoft’s published estimate, collected under different conditions. Arcade’s production version adds a ranking stage and benchmark instructions; Microsoft’s test omitted the instruction string to avoid the cost and delay of an extra model call. Microsoft also describes a search method that can avoid using a resource-intensive ranking model at runtime.

Arcade acknowledges that the comparison suggests a possible advantage, not a controlled result. It does not establish that Arcade completes tasks better or reduces total costs more than Foundry.

That caveat matters because tool search is already appearing across major platforms. Microsoft Foundry Toolboxes gives agents ways to search for and run tools. Anthropic’s Tool Search Tool loads descriptions when Claude needs them, and Anthropic reports an 85% token reduction in its own example, using a different baseline. OpenAI’s Responses API supports deferred function loading and MCP tools, with search handled by a hosted service or the application.

Arcade’s broader pitch is that search belongs in the same control layer as permissions and execution. The platform can check user rights, enforce rules before and after calls, keep credentials out of model context and pass audit logs through OpenTelemetry. Those capabilities may matter more to some buyers than the token reduction.

What the customer still has to measure

Arcade says Tool Recommendation will be included in Arcade Cloud from October 15, but customers will still pay the usual platform and usage fees. The platform’s public pricing is based on usage, not employee count:

Those charges are separate from the cost of the large language models running the agents. Enterprise deployment of the new feature is planned, but Arcade has not named its price or exact timing. Partee said large organizations are Arcade’s main commercial segment and that enterprise deployments would be paid.

I think the useful question is not whether a tool catalog can be made smaller; it is whether the search layer saves more than it costs across a customer’s actual workflows. Arcade’s own calculations say recommendation is less attractive with a fixed list of fewer than about 25–28 tools. Customers also need to count platform fees, tool calls, any enterprise license and their own infrastructure.

The 7.6% workflow result is the strongest evidence in Arcade’s material because it includes both input and output tokens across complete tasks. Even that is a company-reported test, not a customer’s audited bill. The business case will depend on whether savings survive contact with each organization’s tools, search frequency and licensing costs.