Less compute.
Better software.

Less work in every comparison.

MagLev streamed less decode KV traffic in every comparison below — 16 of 16. The median reduction was 63.7%; pooling all traffic on both sides gives 3.61×. This surface charges context size against every decoded token, so no cache policy can move it.

Multiples are baseline total divided by MagLev total. Above 1.00× means MagLev used less. Compute values are token-derived work proxies, not measured hardware FLOPs. The uncached scenario is shown separately, not presented as actual work.

Lower energy requirements. Explicit boundaries.

Both modeled energy surfaces — KV-cache and total memory traffic — were lower for MagLev in all 10 comparisons listed here. Pairings where MagLev drew more energy on either surface are withheld from this table and named in Scope & reporting.

Modeled, not wall-metered. Total here includes KV-cache and decode weight-traffic estimates; it is not total datacenter electricity. Total modeled energy is the strictest surface in the study: it charges MagLev for the extra output it delivers. 8 of the 18 metered pairings draw more energy on one of the two surfaces — concentrated in the earlier Benchmark 4 runs, where MagLev produced substantially more product — and are withheld from the table above rather than shown as energy results. Across the 10 pairings that do qualify, pooling both sides gives 1.63×. Medians are reported instead of arithmetic means because a small number of pairings are decisive outliers in both directions. Baselines are the same as in the compute table.

And the software scored higher.

In the Benchmark 1 and Benchmark 4 same-model quality comparisons below, MagLev scored higher seven times and tied once. In Benchmark 4, MagLev (Opus 5) earned 97.0, versus 88.5 for Claude Code (Opus 5).

Quality and compute datasets have different coverage. The B4 GPT quality comparator is Codex; its metered compute comparator is Claude Code (Opus 5). The B1 Gemini result includes a non-completion, not two completed artifacts. Scores are compared within each study, not averaged across rubrics.

Coverage note: the Benchmark 3 pairing is metered for compute and energy but was not externally graded, so it does not appear in the table above.

Four studies. Different demands.

A progression from a compact programming assignment to extended software-building work.

Divide by the work actually delivered.

Every figure above is a session total, and a total punishes the system that produced more software. These surfaces divide the same metered work by the output tokens that work produced. Same logs, same formulas, corrected units — and the gap widens.

The capacity statement.

A model call is bounded by how much KV state has to be streamed from memory for every token it writes. Context size therefore converts directly into how many of these sessions fit on one card — here, at each arm's own mean context. The strongest terminal result on this page is MagLev (Grok 4.6) vs Grok Build Terminal (Grok 4.6 Build) · V1 at 509×.

Ratios are MagLev advantage: how many times more compute the native harness spent to deliver the same unit of output. Sessions per card and decode rate assume one 80 GB H100 SXM at 3.35 TB/s with decode bound by KV traffic; the constants are fixed and identical on both sides. Terminal rate uses the final decile of model calls, with a minimum window so a short arm reports no rate rather than a single-sample one. This section publishes only pairings where MagLev finished ahead on every surface shown. Held back and named here instead: MagLev (GPT-5.6-SOL) vs Codex (GPT-5.6-SOL) · V2 (context escalation at 0.56×).