Benchmarks

We gave both agents the same ten bugs.

Real bugs from django, taken from SWE-bench Lite. Same model on both sides, same instructions, same code, the same freedom to search. A grader nobody controls decides whether the fix worked. Below is what happened, including every bug we did not fix.

SWE-bench Lite, django subsetClaude Sonnet 510 bugsNetwork open, both sidesOfficial grader
Fewer tokens68%119k against 374k on a typical issue
Lower cost per issue47%$0.0714 against $0.1343 on the issues both agents fixed
Cheaper worst case1.7xOur dearest issue $0.1508, theirs $0.2510
Bugs fixed8 vs 8Too close to call at ten issues. We explain why below.

Did it actually fix the bug?

Each square is one bug. Filled means the fix passed the project's own tests, checked by the official grader. Superbrain fixed 8, Claude Code fixed 8. One of those is a coin flip, not a gap, and we show the working further down.

Superbrain8 of 10 fixed
Claude Code8 of 10 fixed

What each fix cost

The seven bugs both agents fixed. Same bug, same model, so the difference is how much context each one moved to get there. Superbrain was cheaper on six of seven.

SuperbrainClaude Code
Uploaded files got inconsistent permissions2.7x cheaper
$0.0915
$0.2510
A form field rejected a computed folder path2.4x cheaper
$0.0468
$0.1144
Multi line SQL broke query ordering1.0x cheaper
$0.1508
$0.1542
Migration output wrapped in a transaction it should not have1.3x cheaper
$0.0513
$0.0671
A duration field printed the wrong format in its error1.9x cheaper
$0.1151
$0.2190
A username validator accepted a trailing newline4.6x cheaper
$0.0360
$0.1663
HTTP responses mangled memoryview content1.0x cheaper
$0.0975
$0.1009
Deleting an object left a stale primary key2.4x cheaper
$0.0285
$0.0675

The bad days are what cost you

Every dot is one bug, placed by what it cost. Superbrain's dots stay bunched. Claude Code is quick and cheap on easy bugs, then spends $0.2510 on the one it finds hard. If you are budgeting a team, you pay for the worst day, not the average one.

Superbraincheapest $0.0285 · dearest $0.1508
Claude Codecheapest $0.0671 · dearest $0.2510
$0$0.25 per issue

Where it went wrong

Three bugs went unfixed. Two of them beat both agents. Here is what happened in each.

both failed

Both missed the fix. Only one of them broke the codebase doing it

Neither agent passed the new test. What the resolve column hides is what else happened: Claude Code's patch reads schema_editor.connection.alias, the grader calls that migration with schema_editor set to None, and all eight tests that were passing before died on 'NoneType has no attribute connection'. It scored 0 of 8 on the previously passing suite. Superbrain scored 8 of 8 — it did not fix the bug, but it did not break the eight things that already worked. Two identical failures in the table, one of which would have shipped a regression.

both failed

Neither agent solved the media merge

The hardest bug in the set, needing 16 tests to pass. Claude Code got 13, Superbrain 12, and neither broke anything that already worked. Both independently arrived at the same upstream shape, down to the exact separator in the warning message that has decided this bug in every previous run. It is the most expensive instance for both of them and the one neither has ever reliably solved.

What this does not prove

Ten bugs is a small sample. These are the things we would want someone to hold against the numbers.

Ten instances from one repository. A leaderboard figure covers 300 across eleven repos, so this is not comparable to one. The interval on a rate this small is roughly 30 points either way.

One run per instance. The same configuration has produced 7, 8, 9 and 10 on this set across previous runs, so 8 against 8 should be read as a tie and not as a measured equality.

Neither agent had a working Python environment, so neither could run the test suite while working. Both were handicapped the same way.

Claude Code was given a tool we do not have. It ran with WebSearch enabled; Superbrain ships no search tool at all. Switching off a competitor's feature to match our own surface is not a comparison worth publishing, so it was left on. In the event it went unused.

The network was open, which means correctness here is not an isolation claim. It happens not to matter on this run: Claude Code made zero external calls, and Superbrain made five, all on django-11019, which it failed. Every instance either agent resolved was resolved without a lookup, and the audit in the bundle is what that rests on.

How it ran

Everything held identical except the agent.

The task

The GitHub issue text, word for word, plus one line saying do not edit the tests. Same string to both agents.

The model

Claude Sonnet 5 on both sides, pinned, both billed to the same API key.

The code

A fresh django checkout per bug, at the commit before the fix landed, with no later history in it.

Network open, both sides

The fix for every one of these bugs is public, and both agents were free to go and find it — the same freedom, on both sides. What is being measured is how efficiently each harness drives the model, not whether the model remembers the bug. Every transcript was audited and the bundle names each URL either agent reached for.

The scoring

The official SWE-bench grader, in Docker. A bug counts as fixed only if the failing tests pass and the passing ones still do.

The pricing

One rate card for both. We do not use either agent's self reported cost, because they price cached tokens differently.

Check it yourself

We ran this, so do not take our word for it. Rerun the official grader against our patches and see if you get the same 8 and 8.

python -m swebench.harness.run_evaluation --dataset_name princeton-nlp/SWE-bench_Lite   --split test --predictions_path predictions/superbrain.jsonl --run_id verify
Download the evidence

75 KB. Patches, grader output, container logs, per bug usage and the isolation audit. No agent transcripts, since those carry both companies' internals.