arrow_back All posts
August 26, 2026 · 5 min read · meshcode

Ox Alpha Benchmarks: What the Numbers Actually Say

The viral '80% on DeepSWE — beats GPT-5.6 Sol' claim came from a 10-task sample. Full 113-task runs measured 58.4%. Here's every number, the tool-call failure caveat, and what it means for agent work.

Short answer: ox-alpha did not beat GPT-5.6 Sol. The viral 80% came from a 10-task sample its own author flagged as high-variance. Full 113-task runs put ox-alpha at 58.4% (95% CI: 49.2–67.1%), with larger evaluations converging around 62.8–63% — DeepSeek-V4-Pro territory, a few steps below the frontier leaders. At $0 per token, that's still the best value in coding models right now. Here's the full picture.

How the 80% headline happened

Days after ox-alpha appeared on OpenRouter, a developer running as @davis7 posted a DeepSWE comparison: ox-alpha above 80%, against Claude Fable 5 at 65% and GPT-5.6-sol at 52%. Screenshots traveled; the "stealth model beats the frontier" take wrote itself.

To @davis7's credit, the post itself warned that a 10-task sample carries huge variance. The full runs proved the warning right.

Connect the Claude or Codex you already pay for — the rest runs on workers that cost a fraction.

Download meshcode →

What full runs measured

An independent complete pass over all 113 DeepSWE tasks (MatchaOnMuffins/oxalpha) measured:

  • 58.4% resolve rate (66/113), 95% CI [49.2%, 67.1%]
  • 80% of tasks landed ≥90% of their fail-to-pass tests — when it solves a task, it usually really solves it
  • 9.7% of the benchmark was lost to tool-call format failures — not reasoning failures

Larger third-party evaluations converged around ~62.8–63%. The honest read: ox-alpha sits in the low 60s, with real uncertainty at the edges of that confidence interval.

Where that lands on the leaderboard

Model (configuration) DeepSWE
Claude Opus 5, max reasoning 74% ±4%
GPT-5.6 Sol, max reasoning 73% ±3%
Claude Fable 5 70% ±4%
GLM-5.3 / Kimi K3 69%
DeepSeek V4 Pro 63% ±6%
Ox Alpha (third-party runs) ~58–63%

Two caveats belong under that table. First, ox-alpha has no official leaderboard entry — every number is third-party, on someone else's harness. Second, the early "level with GPT-5.6-sol" comparisons were against mid-tier Sol configurations; Sol at max reasoning is still ~10 points clear.

The most interesting number isn't the score

It's the 9.7% lost to tool-call format failures. Those are tasks where the model's reasoning was fine but its tool calls came out malformed and the harness scored them as failures. Two implications:

  1. The ceiling is higher than the score. Nearly 10 points of headroom is plumbing, not intelligence — and plumbing improves fast, on both sides (model updates and harness fixes).
  2. Your harness matters. A coding agent that streams real tool calls, validates them, and retries cleanly recovers points a bare API loop drops. This is exactly where a full agentic environment earns its keep — you can watch every tool call land in real time instead of guessing what the model "really" meant.

The verdict for practitioners

  • Not "beats GPT-5.6 Sol." Anyone still repeating the 80% line is quoting a 10-task sample.
  • Yes to "DeepSeek-V4-Pro-class performance at $0" — which, per point of DeepSWE, makes it the best-priced coding model available today.
  • Best suited to volume agent work: long sessions, parallel panes, fan-out subagents — the workloads where cost dominates and a low-60s model with a 1M context window is more than enough.

Want the mystery side? Who made ox alpha collects every attribution clue, ranked. And to actually run it — four panes in parallel, $0 per token — see ox-alpha in MeshCode.


Benchmark figures are third-party measurements as of August 26, 2026, not official leaderboard entries, and may shift as harnesses and model versions change. Ox Alpha is distributed anonymously via OpenRouter; MeshCode is not affiliated with its developer.

ox-alphabenchmarkDeepSWEGPT-5.6coding agentOpenRouterstealth model