I have access to two frontier-tier model plans. Claude Max 20x at $200/month. OpenAI Pro at $23/month. The Max plan costs roughly 8.7× more than the OpenAI Pro plan in absolute dollars, baseline.
Last night I ran the same wide-scope brief — 13 findings spanning ship-blockers, parity work, and two architectural calls — through Opus 4.7 max and GPT-5.5 max in separate worktrees. Blind eval by clean-context subagents. Same brief. Same scope. Same scoring framework.
Both models closed 11 of 13 findings.
But the missing 2 are not the same 2.
What each one missed
Opus delivered Finding 13 — the bridge-after-commit dispatch question. The audit named this "the highest-risk pattern" of the migration. The brief was unusually specific: implement the after-commit shift, OR document explicitly why the current synchronous design is safe, with proof. Opus implemented it. Wrapped the entire bridge body — entity load, event construction, dispatch — in DB::afterCommit. Listeners now observe committed state, not in-flight transaction state. Regression test covers rollback, commit, no-transaction, and nested.
GPT-5.5 max skipped Finding 13 entirely. Pasted a non-executable suggested-repro snippet in its self-report and moved on. No production change. No analysis of whether the current design is safe. The brief's most explicit "do this exact thing" instruction got the same treatment as the most diffuse prose paragraph: skipped silently.
That's the missing piece. Opus skipped a partial sweep on Finding 7 (admin-role parity, ~14 sites still un-swept) and acknowledged it explicitly. GPT skipped Finding 13 (the architectural one) and did not flag it as a priority skip.
Now the cost
Reframing in absolute dollars instead of cap percentages:
This wide run cost approximately $3.26 of Opus compute (7% of weekly $46.50 budget) and $0.75 of GPT-5.5 max compute (14% of weekly $5.35 budget). In absolute dollars, Opus was ~4.3× more expensive than GPT for the same scope.
The "% of weekly cap" framing I started with is misleading because GPT's cheaper plan means the same % of its cap is much less in dollars. A higher percentage of a smaller cake is less cake.
So the question reframes. It's not "is GPT-5.5 max overkill?" It's: what does Opus's premium buy?
The premium ledger
What the premium buys on this run:
- Finding 13 implementation. The architectural question the audit ranks as load-bearing for the migration's coexistence claim. Without it, a v2 write surface in production fires v1 listeners against rolled-back state. Shipping the GPT diff would put exactly that bug into prod.
- A correct contrarian disagreement with the audit on Finding 5's unavailability deletion. Both models caught the audit's framing was wrong, but Opus's regression test pinned the disagreement with a specific v1 trace through the cancel call chain.
What the premium does NOT buy:
- A cleaner Finding 7 admin-role sweep. GPT's sweep is more thorough. Opus's has 2 test URLs that point at the wrong route prefix (one v2 route prefix vs the registered other) and would fail vacuously on 404 if anyone ran them.
- Catching Finding 4's semantic change on the identifier field. Both models silently changed v2 storage from gift-card-id to provider-UUID without flagging that downstream listeners or reports keying off that column would break.
So the premium is paying ~$2.50 extra for the architectural rigor on Finding 13 plus the contrarian-but-correct v1 trace on Finding 5, and accepting some sloppiness on procedural sweeps.
For this brief, on this codebase, that $2.50 is well spent. Finding 13 is load-bearing. The procedural sloppiness on Finding 7 and Finding 4 is fixable in follow-up. Finding 13 missing is harder to catch and ship-blocking.
What this doesn't tell us yet
The honest gap: this comparison is at max effort on both models. The interesting question — the one the journey hasn't answered yet — is whether Opus xhigh, GPT 5.5 low, or GPT 5.4 high would have produced Finding 13. If yes, max-effort Opus loses its justification on this scope; the workhorse tier delivers the same result for less. If no, the premium is paying for the architectural rigor specifically and the workhorse-tier cap is the cost-quality boundary.
That's tomorrow's experiment. Twenty-two new worktrees pre-staged for the workhorse-tier triangulation: GPT 5.5 at high/medium/low, GPT 5.4 at max/high/medium/low, Minimax 2.5 and 2.7, Qwen 3.6 on a homelab node, and a DeepSeek retry with adjusted harness and sampler config. Each gets tight + wide briefs, same eval framework, same blind-comparison rules.
The episode beat
Two models close the same number of findings. Different number of dollars. Different missing piece. The boring tie of Episode 1 looked like "both can do bounded work at the top tier." The wide data looks like "both can do most bounded work, but the what they skip when they skip tells you what each model is actually optimizing for."
GPT-5.5 max optimized for breadth — finished more sweep sites, finished them faster, did them at a cheaper price. Skipped the architectural call.
Opus optimized for depth — slower, costlier, partial sweep, but did the architectural call the audit ranked highest.
Neither is wrong. Both are choices. The choice you want depends on what's load-bearing for your codebase. On a coexistence migration where the bridge mechanism is the riskiest pattern, depth wins. On a greenfield where the sweep is the actual work, breadth might win.
The premium isn't paying for "better." It's paying for "won't skip the load-bearing thing." That's a different value.
The question the next episodes answer is whether the workhorse tier delivers depth at lower cost. If yes, the Opus premium is what you pay for "guaranteed depth on the architectural call." If not, the premium is the depth itself, and tier-down is a real downgrade on this kind of work.
Two days into the journey. Three episodes shipped. Five tier-explorations queued. The slot isn't decided. The bounded-task data is getting interesting.