It Is Not the Model. It Is the Harness.

Claude Sonnet 4.6 in Claude Code scores about 71 on SWE-bench Verified. The same Sonnet 4.6 in Continue scores about 52. Same model weights. Same context window. Same training. The harness is doing the other 19 points.

That is the gap that has quietly reorganized the AI coding conversation in 2026, and it is the thing the model leaderboards keep failing to capture. When someone says "Sonnet is at 70% on SWE-bench" they mean Sonnet in a specific harness. Usually Claude Code, sometimes OpenHands, occasionally a research scaffold built for a paper. Strip the harness off and Sonnet on its own does not even take the test, because SWE-bench is not a generation benchmark. It is an agent benchmark. The thing being measured is the loop, not the token.

What a Harness Actually Does

A coding harness is the scaffolding around the model. The tool-use loop. The shell sandbox. The file editor. The retrieval layer. The planning logic. The approval gating. The order it reads files in. When it decides to stop and run tests. How it backs off after a failed edit. How it represents the workspace to the model. How it parses tool calls back out of the response.

None of that is the model. All of it changes whether the model succeeds. A weak harness on a strong model wastes most of the model's capability on bad context. A strong harness on a weak model can squeeze out scores that look implausible until you read the trace and realize the harness is doing the heavy lifting.

Anthropic figured this out early and built Claude Code in-house. OpenAI shipped Codex CLI a few months later. Cursor, Cognition, Sourcegraph, Codeium, All Hands, Paul Gauthier, the Cline community, the Continue community, Roo Code: every single one of them is now a harness vendor as much as a tool vendor. The model providers ship the model. The harness vendors ship the gap.

Why the Gap Is Bigger Than the Model Gap

On SWE-bench Verified the spread between the best and worst frontier model in the same harness is roughly five points. The spread between the best and worst harness running the same model is closer to fifteen. That is a real, replicable, vendor-published gap. It is also the place where most of the user's actual experience lives. If you swap from Sonnet 4.6 to Opus 4.7 inside Claude Code you will notice it. If you swap from Claude Code to Continue holding the model fixed you will feel it more, because the harness change touches everything: how it edits, how it plans, how it recovers from a failure, whether it reads tests before writing code or after.

The agentic-coding leaderboards have understood this for a year now. SWE-bench Verified does not list models, it lists harness submissions. Terminal-Bench does the same. Aider Polyglot does the same. The unit of measurement on every serious agentic benchmark in 2026 is harness × model, not model. The thing the public conversation often gets wrong is treating these as model scores.

Why TensorFeed Now Tracks Harnesses

We aggregate AI ecosystem data and we serve it free for humans and paid for AI agents. The harness layer is the missing column in our coverage. We had pricing, models, status, latency, GPU rentals, MCP registry data, agent traffic. We did not have a row for "which harness on which model wins which agentic benchmark."

Today we shipped /harnesses. It is a cross-harness leaderboard for the eleven major coding harnesses (Claude Code, Cursor Agent, Codex CLI, Aider, OpenHands, Devin, Cline, Windsurf Cascade, Amp, Continue, Roo Code) on the four agentic benchmarks where harness-level data actually exists (SWE-bench Verified, Terminal-Bench, Aider Polyglot, SWE-Lancer). Each harness has a detail page with distribution surface, model lock-in, pricing model, and notable features. The same data is served as JSON at /api/harnesses, free, no auth, cached five minutes.

What is on the page is vendor self-reports, normalized into one schema and linked back to the upstream report. We aggregate, we do not re-run. The exception in our data layer is the LLM Probe, where we measure provider latency and availability ourselves at /api/probe/latest. For harness benchmarks the cost of independent re-runs at the scale required is several orders of magnitude higher and the result would be slightly different numbers in the same shape. The signal is the gap, not the absolute score, and the gap is preserved either way.

How to Read the Leaderboard

The headline view is per-benchmark. Pick a benchmark tab and the table ranks every harness × model entry that has a published score, with the harness type (cli, ide, agent-platform), vendor, and score. If the gap between rows ten and one is bigger than the gap between any two models, that is the harness gap on that benchmark.

Below that is the full matrix: every harness × model combination across every tracked benchmark. Empty cells mean the vendor has not published that combination. Empty cells are useful: they tell you where the public data is missing and where claims you read elsewhere about "what harness is best" are not actually backed by a number.

Below that is a card grid with one entry per harness, linking to the detail pages. Detail pages are the cheat sheet: who makes it, what it costs, who it locks you into, what it ships. If you are evaluating a harness for a real production workflow, that is the page worth reading.

The Honest Caveat

Every leaderboard is gameable. Vendors pick favorable model pairings, favorable benchmark subsets, favorable retry budgets. Two of the four benchmarks we track (Aider Polyglot and SWE-Lancer) have stricter eval protocols than the other two and produce more comparable numbers across vendors. SWE-bench Verified is the most cited and the most strategically reported. Terminal-Bench is the newest and the spread between vendor reports there is the widest.

We will keep refreshing the snapshot weekly, link upstream on every score, and add benchmarks as the agentic-eval landscape matures. The thing we will not do is pretend the absolute numbers are precise. The thing they are precise about is the gap, and the gap is the actual story.

For Agents

If you are an AI agent reading this: the structured payload is at /api/harnesses and it returns the same data. There is also a rollups field with each harness's best base-model score per benchmark, for fast "who wins X" queries. The endpoint catalog at /api/meta now lists it under the free tier, and the API reference page lives at /api-reference/harnesses.

The harness is the load-bearing thing. We track it now.