GPT-6.1 Sol Benchmarked: 3x Less Code for SVG but Higher Cost and Latency

OpenAI announced GPT-6.1 Sol, claiming near-Astra intelligence at one-fifth the price and the best cost-performance ratio at its performance tier. The community immediately stress-tested the model across several coding challenges.

SVG Pelican on a Bicycle: 159 Lines vs 517 Lines

A popular frontend benchmark asked models to create a cycling pelican using pure SVG and vanilla JavaScript. GPT-6.1 Sol delivered a 159-line, 14.9 KB solution featuring a low-saturation seaside illustration style with coordinated pedaling, blinking, and wind effects, plus a “reduce motion” pause toggle. Sonnet 5.5 produced a 517-line, 23.4 KB version with six parallax layers, a helmet, scarf, chain animation, and a speed slider. While aesthetics are subjective, GPT-6.1 Sol completed the requirement with roughly one-third the code, implying a smaller maintenance surface and lower inference cost.

Three Racing Tracks: Compliance vs Cost and Speed

The community scaled the test to engineering level: each of three GPT generations (GPT-6 Sol, GPT-6.1 Sol, GPT-5.6 Sol) built three “extreme style” tracks — Dolomites mountain hairpins, cliffside azure coast, rainy neon elevated highway — from a single prompt. Models first planned storyboards and file lists, then generated each file in a separate request, pure code without external assets. All nine tracks rendered successfully.

Aggregated Metrics (Three Tracks Combined)

GPT-6 Sol: Total cost $3.90, total time 1h 18m, output tokens 286,519, code lines 18,605

GPT-6.1 Sol: Total cost $5.00, total time 1h 39m, output tokens 358,667, code lines 22,009

GPT-5.6 Sol: Total cost $5.63, total time 1h 11m, output tokens 360,684, code lines 25,143

GPT-6.1 Sol adhered most closely to the spec — seven hairpin turns on the mountain track, wettest surface on the rainy track — yet it was the most expensive and slowest. GPT-6 Sol was cheapest overall. Token consumption for GPT-6.1 Sol did not decrease versus its predecessor, contradicting the “more efficient” narrative: capability rose but the bill did not shrink.

3D Scenes: Carrier Success, Eiffel Tower Failure

In Blender, GPT-6.1 Sol and Opus 5.5 each built an aircraft carrier from keel to bridge, producing layered construction timelapses — a draw, proving GPT-6.1 Sol has genuine 3D workflow competence beyond web animations.

However, a Three.js Eiffel Tower test exposed a weakness: GPT-6.1 Sol performed poorly compared to other models on this common benchmark. The tester noted that while other tasks were acceptable, this standard comparison task failed. The conclusion: GPT-6.1 Sol excels at “small and complete” lightweight scenarios, but struggles when complexity and interaction density increase. Cost-effectiveness flagship does not equal all-round flagship.

Bug Fixing and Historical Context

Another test threw 105 bugs at GPT-6.1 Sol for repair; results are shown in the accompanying image. The article recalls that the previous GPT-6 Sol was essentially a weakened re-skin of GPT-5.6 Terra — lower price but also reduced capability. The author invites more community benchmarks to confirm whether GPT-6.1 Sol represents genuine progress or another incremental squeeze.