When Is GPT-5.6 Sol Max Worth 2.42× the Cost for a Coding Agent?

Disclosure: I maintain Tura, an open-source coding-agent project. This is a technical analysis of public benchmarks and our published artifacts, not an independent product review.

The decision is task-shaped, not model-shaped

“Use Max because it is stronger” is not a useful operating rule for coding agents. Max buys more search, revision, and agent rounds. That can be valuable when an agent still has to discover architecture and compatibility constraints. It can be expensive slack when a failing test and the likely change surface are already known.

I compared effort levels in two settings: the public DeepSWE v1.1 records and a behavior-compatible Rust-to-Python rewrite of the eza CLI. The split is the point:

The repair group is deliberately small, so it does not prove that Max harms bug fixes. It does disprove the easier claim: that a stronger effort level is automatically the best default.

The broad average: +3.3 points for 2.42× spend

Across 113 DeepSWE tasks in the same mini-swe-agent harness, High reached 69.4% pass@1 at $3.47 per task; Max reached 72.7% at $8.39. That is +3.3 percentage points for 2.42× cost, 2.11× output tokens, 2.91× input tokens, 1.90× elapsed time, and 1.66× steps. In expectation, cost per passing attempt rises from roughly $5.00 to $11.54.

For eza, the extra trajectory had somewhere to go: compatibility work, build failures, behavioral checks, and revised implementation plans. In the Tura Direct harness, the score rose from 79.8% to 92.3%; in Codex CLI, from 78.8% to 92.3%. Those are meaningful gains, but they came with 2.27–3.27× cost.

A practical routing rule

• Bounded bug or scoped change: start at High; escalate when localization remains uncertain or High fails.
• Feature implementation: start at High; pay for Max when the cost of a failed attempt exceeds roughly 2.4× spend.
• Rewrite or migration: Max is often justified when compatibility surface is broad and independently verified.
• Greenfield project: choose by scope. Max is plausible when one agent owns architecture, implementation, and verification, but there is not yet a matched High-to-Max greenfield benchmark.

The useful metric is not visible reasoning volume. It is completed, verified work: task shape, verifier, cost, tokens, duration, and agent rounds together.

The complete analysis includes 19 tables, charts, the mechanical DeepSWE grouping, source links, and public artifacts: https://turaai.net/blog.html#is-gpt-5-6-sol-max-worth-it

Artifacts: https://github.com/Tura-AI/benchmark/tree/main/blog_data/eza-replication-gpt56-max-20260717 Repository: https://github.com/Tura-AI/tura

I would welcome attempts to reproduce, refine the task taxonomy, or challenge the routing rule.