Five faults frontier models don't fix (and a $0.45 fine-tune does)

On 2026-07-19 I ran GPT-4o, Claude Sonnet 4 and Claude Sonnet 4.5 against the 35 hard tasks of my held-out set — the same prompt, temperature 0, and the same hidden checkers that grade my own models. Full disclosure first:

The frontier models win overall. GPT-4o fixed 19/35, Sonnet 4 fixed 19/35, Sonnet 4.5 fixed 22/35. My tuned 7B fixed 12/35. This page is not "a 7B beats GPT-4o." It's the five specific faults where the big models fail and the small tuned one doesn't — with every transcript published so you can check me.

faultgpt-4osonnet-4sonnet-4.5tuned 7B
202f4a13 template re-renderfailfailfailpass
73db72e9 revenue report crashfailfailpasspass
99b8b877 UTC-naive timestampspassfailfailpass
d2e3f44f CRLF breaks CSV→JSONfailpasspasspass
f62d646f idempotency violatedpassfailpasspass

The star is 202f4a13: all three frontier models fail it — and they don't even agree on which file the fault is in. GPT-4o rewrites renderer.py; both Sonnets rewrite parser.py. All three fixes look plausible. All three fail the checker. The tuned 7B fixes the right file.

The five, in one line each

202f4a13 — the template engine remembers the previous render

Rendering "Hello, {{ name }}!" with "Alice" works; rendering again with "Bob" still prints "Alice". State leaks between renders across two files. Multi-file: fixing one file provably still fails.

73db72e9 — monthly revenue report crashes with AttributeError

CSV loading is fine; the report module blows up on a type mismatch between what the loader stores and what the report expects.

99b8b877 — timestamps lose their timezone

Events come back with naive datetimes; the contract says UTC-aware. The fault sits in a helper other code silently depends on.

d2e3f44f — Windows line endings poison the output

CSV→JSON conversion fails on CRLF input with a stray carriage return inside a value.

f62d646f — "idempotent" import inserts duplicates

Running the same import twice should insert zero new rows the second time. It inserts three.

Check me

The full bundle — each fault's issue report, broken code, the (now-public) checker, all frontier transcripts verbatim, and my model's passing fix — ships alongside the model release. Grading rule, for every model equally: the proposed fix is applied to the broken tree and the hidden checker runs; pass means the checker genuinely completes. Publishing these five retires them from my held-out set — a public checker is no longer hidden, so they will never be part of a future headline number.

Share this: X Hacker News Reddit LinkedIn Email