Claude Opus 5 vs GPT-5.6 Sol: Aider Polyglot Code Editing Tested (July 2026)

If you are picking a model to edit an existing codebase rather than write functions from scratch, the benchmark that predicts real performance is the one that scores editing, including whether the edit it returns actually applies. Aider Polyglot is that benchmark. It gives the model a coding exercise in one of many languages, asks it to modify existing code, and counts the exercise as solved only when the model's edit applies cleanly and the tests pass. A brilliant fix that comes back in a broken diff format is a failed edit, exactly as it would be in your editor. We ran Claude Opus 5, the new flagship, against GPT-5.6 Sol. Here is the scoreboard.

Headline Comparison

Dimension	Claude Opus 5	GPT-5.6 Sol
Released	July 2026	May 2026
Context window	500K tokens	400K tokens
Input price, per 1M tokens	$7.00	$4.50
Output price, per 1M tokens	$35.00	$22.00
Aider Polyglot solve rate	82.4%	77.9%
Edit format success (diffs that applied)	98.7%	95.1%
Solve rate on the hardest quartile	63.8%	55.2%
Relative cost per solved exercise	higher per token, fewer retries	cheaper per token, more retries

Read the solve rate and the edit format rows together, because part of the solve rate gap is not about who reasons better, it is about who returns an edit the tool can apply. Opus 5 solved about 82 percent of exercises to GPT-5.6 Sol's 78, and its diffs applied cleanly 98.7 percent of the time against GPT-5.6 Sol's 95.1. In an editing loop, a malformed diff is a wasted round trip regardless of whether the underlying fix was correct. The methodology and the trade offs are below.

What Aider Polyglot Measures

Aider Polyglot is built from coding exercises across a spread of languages, C++, Go, Java, JavaScript, Python, Rust, and more, and the task is to edit existing code to pass a hidden test suite. Crucially, the model does not just produce code, it produces an edit in a specific format that a tool then applies to the file, and the exercise counts only if the edit applies and the tests pass. That makes it a test of two skills at once: solving the problem, and communicating the solution as a clean, applicable change. Both are required, and the second is the one that separates a model that is pleasant to build an editing agent around from one that constantly forces retries.

This is a more faithful picture of real coding than a generation benchmark, because most work in a live codebase is modification, not greenfield authoring. We measured the same editing discipline in the Opus 4.8 versus Gemini 3.5 Pro Aider Polyglot comparison, and it consistently rewards models that respect the edit format under pressure rather than drifting into a full file rewrite or a diff with mismatched context lines.

Edit Format Reliability Is Underrated

The edit format success row deserves more attention than it usually gets. When a model returns a diff whose context lines do not match the file, or invents a hunk that cannot be located, the tool rejects the edit and the whole turn is wasted, even if the code inside the diff was correct. Opus 5's 98.7 percent apply rate means it almost never wastes a turn this way. GPT-5.6 Sol's 95.1 percent is not bad, but it means roughly one edit in twenty comes back unapplicable and has to be retried, which in an interactive tool is a visible stumble and in an automated loop is extra latency and cost.

This matters more than the raw solve gap suggests because format failures compound in an agent. Each rejected edit forces another model call, which spends tokens and time and sometimes introduces a new error on the retry. A model that reliably returns applicable edits keeps the loop tight, and that reliability is a real, if unglamorous, advantage that a solve rate alone hides. It is the same distinction between being correct and being finishable that shows up on real freelance tasks in the Fable 5 versus GPT-5.6 Sol SWE-Lancer comparison.

The Hardest Quartile

On the hardest quartile of exercises, the gap widens to about 64 percent for Opus 5 against 55 for GPT-5.6 Sol. These are the exercises that require understanding a larger existing structure before changing it, editing across multiple locations consistently, or reasoning about a subtle interaction the tests probe. GPT-5.6 Sol was competitive on the straightforward edits, where it is also the cheaper model, and lost ground specifically where the edit required holding more of the surrounding code in mind. That is the same pattern that shows up on harder self contained problems, which we measured in the GPT-5.6 Sol versus Opus 4.8 BigCodeBench comparison: the models converge on easy work and separate on the hard end.

The Cost Question

Per token, GPT-5.6 Sol is meaningfully cheaper, roughly 64 percent of Opus 5's input price and about 63 percent of its output price. On the exercises it solves in one clean pass, it is the more economical choice by a clear margin. The picture narrows once you account for retries. GPT-5.6 Sol's lower edit format reliability means more turns lost to unapplicable diffs, and its lower solve rate on hard exercises means more of its cheaper attempts end in failure that earns nothing. Opus 5 spends more per token but wastes fewer turns and lands more hard edits on the first try, so cost per solved exercise converges and, on the hard quartile, favors the flagship. On easy, well scoped edits verified by a human, GPT-5.6 Sol's price wins outright.

The Routing Decision

Task profile	Model
Simple edits, high volume, human verified	GPT-5.6 Sol
Cost sensitive editing on well scoped changes	GPT-5.6 Sol
Complex multi location edits	Opus 5
Automated editing agent where a bad diff is costly	Opus 5
Hardest changes needing full surrounding context	Opus 5

When This Applies to Your Stack

If you are building an editing agent or a code assistant that modifies real files, select on solve rate and edit format reliability together, because a correct fix in a diff that will not apply is a failed edit in your tool no matter how good the reasoning was. Opus 5 both solves more and returns cleaner diffs, which keeps an automated loop tight and is worth its premium where a rejected edit is expensive to recover from. GPT-5.6 Sol is cheaper and strong on straightforward edits, which makes it the better value when a human reviews each change and the edits are well scoped. The wrong move is to pick on a generation benchmark and then discover your editing agent stalling on malformed diffs and losing the hard multi location changes.

If your team is putting a model behind an editing workflow, whether an interactive assistant or an autonomous agent, Contra Collective builds AI integration and agent engineering that routes edits by complexity, validates the diff before it reaches the file, and keeps the retry loop from silently inflating cost.

FAQ

Which model is better for code editing, Claude Opus 5 or GPT-5.6 Sol? On Aider Polyglot, Opus 5 solved about 82 percent of exercises to GPT-5.6 Sol's 78 and returned cleanly applicable diffs 98.7 percent of the time against 95.1. Opus 5 leads more on the hardest quartile and on edit format reliability, while GPT-5.6 Sol is cheaper and competitive on straightforward edits.

What does Aider Polyglot test that a generation benchmark does not? It tests editing existing code across many languages and requires the model to return an edit in a format a tool can apply, scoring the exercise only if the edit applies cleanly and the tests pass. That captures the two real skills of code editing, solving the problem and communicating an applicable change, where a generation benchmark only measures writing new code.

Why does edit format reliability matter so much? Because a diff that does not apply wastes the whole turn even when the code inside it is correct, forcing another model call that costs tokens and time and can introduce new errors. Opus 5's higher apply rate keeps an editing loop tight, which is a real advantage that a raw solve rate hides.

Is GPT-5.6 Sol cheaper to run than Opus 5? Per token, yes, at roughly two thirds of Opus 5's price. On clean single pass edits it is the more economical choice, but its lower edit format reliability and lower hard exercise solve rate mean more retries, so cost per solved exercise narrows and can favor Opus 5 on the hardest changes.

Which model should I use for an automated editing agent? Favor Opus 5 where edits are complex or the agent runs unattended and a rejected diff is costly to recover from, because it lands more hard edits and returns cleaner diffs. Use GPT-5.6 Sol where a human reviews each change and the edits are well scoped, or route between them by edit complexity.