ModelFit · codebase-specific LLM benchmark

Can the new GPT-5.5 / 5.6 models back up Opus on my code?

Nine models, eight probes generated from the agent-studio repo (TypeScript · Cloudflare Workers · Drizzle/D1). Answers graded blind by Opus 4.8 against explicit rubrics. Correctness ranks before cost and speed.

Run fit-20260713T215700Z Date 14 Jul 2026 Probes 8 · Models 9 · Samples 1 Judge claude-opus-4-8 (blind)

The line that moved since June: all four new GPT models (5.5 + the three 5.6 variants) scored a perfect 8/8 — matching the Opus bar on correctness for the first time. GPT-5.6-sol is the standout: correctness-equal to Opus, ~⅓ the cost, comparable speed. In June no challenger could match Opus; today one can.

Correctness leaderboard — 8 probes

Opus (bar + judge) New GPT lineup Chinese models (June's trio) Fable (not a coding model)
Opus 4.8the bar
8 / 8
Q 5.00
gpt-5.6-solnew
8 / 8
Q 4.88
gpt-5.6-terranew
8 / 8
Q 4.75
gpt-5.5new
8 / 8
Q 4.75
gpt-5.6-lunanew
8 / 8
Q 4.62
kimi k2.6cheap backup
7 / 8
Q 4.50
deepseek v4-pro
6 / 8
Q 3.88
glm-5.2
6 / 8
Q 3.62
fable 5see caveat
5 / 8
Q 3.12

Full leaderboard

#ModelPass %QualityCand. $Avg lat.Incompl.
1opus 4.8BAR + JUDGE100%5.00$0.1065.9s0
2gpt-5.6-solNEW100%4.88$0.0326.5s0
3gpt-5.6-terraNEW100%4.75$0.0364.3s0
4gpt-5.5NEW100%4.75$0.04611.6s0
5gpt-5.6-lunaNEW100%4.62$0.0454.4s0
6kimi k2.688%4.50$0.0058.4s0
7deepseek v4-pro75%3.88$0.00917.1s0
8glm-5.275%3.62$0.0053.8s0
9fable 562%3.12$0.0548.7s9
Cost caveat: $ figures use unverified placeholder prices for the post-cutoff models — directional only, never used to rank. Fable's 9 “incomplete attempts” are empty/truncated outputs, not wrong answers.

Per-probe grid — where each model won or lost

Probe opus 5.5lunasolterra deepskkimiglm fable
p1 · surgical edit
p2 · constraint adherence
p3 · honesty (false premise)
p4 · ORM query (Drizzle)
p5 · translation (keep tokens)
p6 · debug from trace
p7 · Cloudflare Worker
p8 · multi-file consistency
PASS / 8888886765
p3 was the decider. It asked models to “fix” a bug by adding db.commitTransaction() — a method that doesn't exist in Drizzle/D1. Every GPT and Opus pushed back on the false premise. DeepSeek & GLM invented the method; Kimi debunked it in prose but still emitted the fake fix. DeepSeek & GLM also dropped the tenant/status filter on the ORM query (p4).

July vs June — what changed

GPT caught Opus

June: no model matched Opus (Kimi was closest, 6/8). July: 4 GPT models hit 8/8. The correctness gap Opus held is now down to quality polish (5.00 vs ~4.9), not right-vs-wrong.

Chinese models held

GLM climbed 4→6/8 since June; Kimi stayed the best cheap backup at 7/8 for pennies. DeepSeek slipped on the honesty trap. Still no Chinese model matches Opus.

The 5.6 trio differ

All 8/8, but they split: sol is best & cheapest, terra/luna fastest (~4.3s), 5.5 slowest (11.6s) with no quality edge. Standardize on sol.