What happened
Mixdog describes itself as an open-source desktop coding agent in which a lead model plans work while less expensive models can handle search, edits, and review. The project says users can mix models by role, use workflows ranging from Solo to Swarm, and run the application on Windows, macOS, and Linux. Its README also lists provider options including API keys, account sign-in for several services, OpenRouter, DeepSeek, OpenCode Go, and local models in specified circumstances. The project says the software is free and open source under the Apache-2.0 license.
The repository’s benchmark section reports runs on Terminal-Bench 2.1, with results scored by the official Harbor verifier. For GPT-5.6 Sol at xhigh reasoning effort, Mixdog reports a five-trial run across 89 tasks: 385 solves out of 445 trials, compared with Codex CLI’s 383. It reports a 46% lower median per-task token count for Mixdog, 18.5K versus 34.3K, and a 46% lower cost. The README’s headline separately claims 63% fewer tokens than Codex CLI at the same success rate, so the headline and the listed five-trial comparison should not be treated as interchangeable measures.
A more recent, single-trial run dated October 15, 2026, compares GPT-6.1 Sol xhigh with Codex CLI. Mixdog reports 77 solves against 78 out of 89, 55% lower cost, and final context of 10.8K versus 33.6K tokens. The project notes that this was the first full run with its effort judge active on OpenAI routes and web search enabled for both sides, and says the five-trial run remains the reference.
For Claude Opus 5, the repository reports one trial in which Mixdog solved 79 of 89 tasks versus Claude Code’s 77. It reports 19% lower cost and final context of 27.6K versus 38.2K tokens. The project links benchmark records, logs, and scripts in the repository, and says Mixdog used fewer tokens on 87 of 89 tasks in the GPT-5.6 five-trial comparison. Its latest listed release, v1.4.2, is dated October 29 and adds an Explore read-only exploration agent, a Test Suite skill, Vertex AI support, and multiple accounts per connector.
What the archive adds
Our earlier Terminal-Bench coverage noted the benchmark’s role in evaluating agent performance through task-based results. This is a new result on that known evaluation thread, not independent confirmation: Mixdog supplies its own comparison and artifacts, while the benchmark’s verifier provides the scoring mechanism.
Trend signal
The results are a limited data point for agent economics moving to completed work: task success is reported alongside token use and cost, rather than as a score alone. They do not establish a general cost-per-verified-task advantage, since the runs are project-reported and this excerpt does not provide independent replication across teams, workloads, or model configurations.
Why now
The repository’s latest listed benchmark run is dated October 15, 2026, and its latest listed release is v1.4.2, dated October 29. The project presents the newer run as a single pass with an updated effort judge on OpenAI routes; it explicitly retains the five-trial result as the reference comparison.
Connections to other events
Earlier New Runtime coverage discussed Terminal-Bench-Science and other new benchmarks, while a separate note examined token efficiency in longer agent runs. Mixdog brings these two concerns together in one coding-agent comparison: benchmark task completion and the resources used to reach it. The comparison remains specific to the reported setup.
Why it matters
A coding agent that completes roughly the same number of terminal tasks at lower cost would matter to teams running repeated agent work, because inference and context consumption can shape the economics of deployment. Mixdog’s results make the harness a variable worth measuring: its README describes a lead-and-worker arrangement and an effort judge, and reports lower token counts alongside task outcomes. That is more informative than a token-saving claim without a success measure.
But the numbers should inform a test plan, not a procurement decision. They come from the project’s own benchmark runs, and the README’s headline token-saving figure differs from the five-trial table’s per-task comparison. Teams need to inspect task-level records, reproduce the settings, and compare on their own code and review criteria. The useful question is not simply whether an agent spends fewer tokens, but whether it produces accepted work with less total cost and no increase in correction or verification burden.
Limits and uncertainty
The supplied material does not establish independent reproduction, broad performance across real codebases, or whether the reported cost advantage persists under other provider prices and workloads. The five-trial GPT-5.6 comparison and the October single-trial GPT-6.1 comparison have different sample sizes and configurations; they should not be read as a continuous trend. The README provides links to logs and scripts, but this account does not independently audit them. Nor do the results show whether cheaper runs shift work onto human review or retries. The benchmark scores are evidence about these reported tasks, not a guarantee of production outcomes.
Skill to build now
Run a small, matched evaluation before changing a coding-agent workflow. Select representative terminal tasks, keep the model, reasoning effort, tools, and success criteria fixed, then record verified completion, retries, tokens, cost, and human review time for each run. Preserve task-level logs so that a lower aggregate cost can be checked against regressions on specific kinds of work.
Reader takeaway
Mixdog’s Terminal-Bench results are a useful prompt to measure the economics of the whole agent harness, not just the underlying model. The headline savings remain project claims: check the detailed run, reproduce it, and include successful completion and review burden before concluding that fewer tokens mean cheaper engineering.