Unreal Labs has developed Unreal Agent, a system that moves tool calls outside the model’s mind, and claims the shift reduces costs by up to 40% against Codex on production tasks and coding tests. The company describes the arrangement as a “frontier cost-efficiency” design, and has made the code available on GitHub.
The Cost Claim
Unreal Labs claims Unreal Agent delivers up to 40% cost savings over Codex and up to 20% cost savings over Pi when measured against real-world workloads and agentic benchmarks. The firm put GPT-6 Astra xhigh through its paces against both Codex and Pi, with the test results made public in its own benchmarks.
The Harness Design
Unreal Agent handles tool calls entirely asynchronously, which frees the underlying model from having to manage waits, polls, and heartbeats for tools. This design delivers two key advantages:
- Users can always guide the agent without pausing for tool calls to complete.
- The agent can fit in more useful tool call work between model calls.
SDK Trade-Offs
Unreal Labs has developed several agentic products and picked up lessons about widely used SDKs as it went. The SDKs that lean on CLIs, such as Claude’s Agent SDK, come with assumptions about local sessions, subprocesses, and resource limits that don’t move over easily into production work. Keeping completion, cancellation, and background tasks running reliably usually means building your own lifecycle management around these tools.
Moving between API modes risks breaking tools or ruining compaction, and raising the SDK changes how messages are built, which forces integration code to be rewritten. Approvals and security measures that depend on harness hooks and custom tools usually need more upkeep and offer weaker protection compared with environment or sandbox limits set outside the harness.
The Cost Savings Breakdown
According to Unreal Labs, the cost reduction comes from two factors. The first is a small footprint for the harness combined with deliberate management of how tools are used, including basic prompts, results tuned for efficiency, and no sub-agents or workflows. The second factor is that more work gets done through tools during each model turn. Because the model operates asynchronously, the agent can make multiple heavy tool calls without spending tokens on checking for completion or waiting.
How The Harness Works
When Unreal Agent makes a tool call, it records the event-log entry for the tool returning in the “in-progress” state at once, even though it keeps going in the background. The actual finish of a tool is what adds its result to the session log and triggers a call to an LLM. Getting this whole process to function without disrupting the cache was an interesting engineering challenge on its own.
The Benchmark Results
The testing of GPT-6 Astra xhigh by Unreal Labs pitted it against Codex and Pi. The outcomes were posted on the company’s website.
| Benchmark | Codex (lb) | Unreal Agent | Pi |
|---|---|---|---|
| Terminal-Bench 4.0 | Codex (lb) | Unreal Agent | Pi |
| SWE-Atlas Codebase QnA | Codex (lb) | Unreal Agent | Pi |
| DeepSWE 1.1 | Codex (lb) | Unreal Agent | Pi |
| Agents’ Last Exam | Codex (lb) | Unreal Agent | Pi |
Getting Started
Unreal Labs has released an SDK for Unreal Agent that comes with a Go library built for direct integration into your codebase, a runner executable along the lines of claude -p /codex exec, and a benchmark runner that works with Harbor. Anyone wanting more information should reach out to contact@unreallabs.ai.
The API Limits
The Responses API documentation does not specify how to use two tool-call result items within a single context, according to Unreal Labs. The company found that some models rejected such requests during testing, and the function_call_output status field appeared to make no difference in those situations.
The team behind Unreal Labs has developed several agentic products, and they know firsthand that software development kits come with hidden expenses. Managing the product’s lifecycle, keeping APIs updated, and dealing with dependency trees are all tasks that don’t appear in a vendor’s promotional material. Anyone can check the benchmarks and inspect the GitHub code for themselves.
See the video the story is built around at unreallabs.ai.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.