Claude Opus 5 vs GPT-5.6 Sol: Frontier Analysis

Claude Opus 5 and GPT-5.6 Sol now sit in the same hard coding-agent decision: when the job is not merely to produce a patch, but to produce a change a maintainer would accept, which route should receive the task?

The current FrontierCode 1.1 leaderboard gives a narrow but useful answer. Opus 5 scores 53.4% on mergeability versus Sol's 47.5%. It also costs$4.30 per rollout, compared with Sol's $6.30. Output volume is nearly equal, 33.6K tokens for Opus and 33.2K for Sol.

The Verdict

Route FrontierCode-like coding tasks to Opus 5 first. On this leaderboard configuration, it has a 5.9-point mergeability advantage and a 32% lower cost per rollout. The combination matters more than either number alone. A cheaper run that produces less mergeable work simply shifts cost into retries and review. Here, Opus has both the better result and the lower observed task cost.

That is not a license to call Sol weak. A 47.5% score on a benchmark designed around maintainer-grade changes is competitive. The important limitation is configuration: Opus ran at medium effort, Sol at max. The result tells an operator which published configuration is more efficient today. It does not cleanly measure the underlying models at identical compute budgets.

What FrontierCode Measures

FrontierCode is not a patch-execution leaderboard. Cognition asks whether a change would be merged, combining unit tests, verifiers, and rubrics defined by the maintainers who wrote the tasks. The benchmark also detects solution-bearing internet use and assigns zero to flagged runs. That makes it a better fit for agent routing than benchmarks that only ask whether a test suite eventually passes.

The distinction explains why the score and pass rate are separate columns. Opus and Sol both show a higher pass rate than their final mergeability score. A successful execution is necessary, but it is not enough if the change has poor scope discipline, misses project conventions, or fails the maintainer-defined review criteria.

The Current Result

Cost is not just list price

Both APIs list the same $5 per million input tokens. Opus's output price is $25 per million, while Sol's is $30. That small list-price gap cannot by itself explain a $2 task-cost difference because output totals are almost identical. The leaderboard result reflects the full rollout, including the agent's model behavior under its selected effort setting.

The Effort Caveat

This is the caveat that should accompany every headline. Opus's entry is marked medium. Sol's is marked max. Higher effort normally gives a model more room to reason, revise, and consume tokens. It can raise quality, but it also changes latency and cost. The leaderboard does not provide an equal-effort, equal-token counterfactual.

That asymmetry makes the practical finding more interesting, not less: the published Opus configuration reaches a stronger result without requiring its maximum setting. But it also prevents a stronger claim about intrinsic intelligence. Do not extrapolate this one result to every coding environment, or assume Sol max is the best OpenAI setting for your tasks.

Price and API Surface

Context limits are not the differentiator for ordinary repository tasks. Both models can carry far more context than most agent loops should send blindly. Retrieval quality, tool state, test feedback, and iteration policy are more likely to determine task success than an extra 50K tokens of available input capacity.

Routing Decision

• Start with Opus 5 for mergeability-sensitive coding. The current FrontierCode entry gives it the better observed quality and lower task cost.
• Keep Sol in the candidate set.Its result is close enough that a codebase-specific evaluation can reverse the order, especially if your workflow benefits from OpenAI's tools or existing infrastructure.
• Match effort before making a permanent choice. Test medium, high, and max configurations against the same task set and record completion, review acceptance, cost, latency, and retries.
• Use the final review outcome as the north-star metric. A one-point benchmark gain is not worth much if it creates noisy diffs or increases human review time.

The central FrontierCode finding is operational: a published medium-effort Opus 5 run outperformed a max-effort Sol run while costing less and emitting almost the same number of tokens. That is a strong default-route signal. It is not the final answer for a production system, which should use the same measurement discipline on its own repositories and reviewers.