OpenDesign quality-cost ranking: 17 models, no single winner

I laid the 17-row table in front of me: GPT-6.1 Sol 90.1 on top, GPT-6 Sol 89.9 right behind, Claude Sonnet 5.5 third at 86.8. The test is a design arena: each model ships a finished showcase piece and judges grade the result. Cost sits in the same table: $0.382 and $0.297 for the Sol siblings, $0.574 for Sonnet. These first three rows all come from open-design.ai records, where every model lists its score next to its estimated per-artifact cost.

The story of the leading pair stretches into late September: GPT-6.1 Sol replaced GPT-6 Sol on the September 29 DevDay stage, listed at $2 input and $10 output, one fifth of flagship Astra pricing. The venturebeat.com write-up draws the same line: Astra-class agentic coding and computer use, with an Ultrafast tier reaching 300 tokens/s. The table feels honest here: a 0.2-point gap in this AI showcase costs 29% extra. On the benchmark side the Sol family measures around 48 on the Intelligence Index, running near 100 tokens/s.

The Sonnet 5.5 camp arrived September 28: $2 input, $10 output, cache reads halved to $0.10. The artificialanalysis.ai measurement puts Sonnet 5.5 at 56 on the Intelligence Index, two points behind flagship Opus 5.5, while noting 70.6% on Terminal Bench 4.0 ahead of Opus. For this model , scored 86.8 in the design arena, the artifact bill reads $0.574. My read: for a team staying inside one house, Sonnet is the safe harbor and the bill stays predictable. The cache cut trims agent workloads by roughly 20%.

The top is pricey, the middle is smart

The middle rows are the most fun part of this table: MiMo 2.6 Flash at 83.5 for $0.029, GPT-6 Luna at 83.1 for $0.016, Grok 4.7 at 82.9 for $0.662. The openrouter.ai comparison page lays the structural gap bare: Grok 4.7 lists $2 input and $6 output over 500K context, while MiMo Flash lists $0.10 input and $0.28 output over 1.05M context. So a similar benchmark band with more than a 20-fold price gap. For showcase trials I would try the open-weight MiMo side or Luna, not Grok. Token spend decides here.

The arrowed row in the image is Claude Haiku 5.5: 82.2 points, $0.018. The anthropic.com announcement is dated October 7: $0.10 input and $0.50 output up to 100K tokens, both lines quintupling above it, 1M context. The artificialanalysis.ai reading places Haiku 5.5 second in its price class at max effort with Intelligence Index 43, costing $0.21 per task. The beam.ai review points the same way: 34 points at medium effort for $0.05, a made-to-measure role for sub- agent work. To me this row is the star of the table: 4.6 points under Sonnet at one thirtieth of the bill. That ratio is rare in LLM selection.

The Astra row breaks the pattern: 82.7 points, $1.61. The flagship model , listed at $10 input and $50 output, sits 0.4 behind Luna here at 100 times the bill. The Fable 5.1 row is even more striking: 80.3 points at $3.66, the priciest cell in the table. The artificialanalysis.ai records show Fable at $10 input and $50 output costing $2.37 to $3.76 per Intelligence Index task. My verdict is blunt: general AI power may be high, but in this arena it lands far off the price-performance line. I look at finished-job bills, not list prices, when I pick a model .

The expensive wing and the lower rows

The Flash corridor continues: DeepSeek V4.1 Flash at 81.2 for $0.023, GLM-5.3 Flash-X at 80.1 for $0.098. The yottalabs.ai comparison gives the backstage view: DeepSeek at 552B parameters in a $0.044 input and $0.30 output band, GLM in a $0.036 input and $0.50 output band. Live openrouter.ai prices point the same way. These two rows are my backup for the Haiku-Luna pair: if one stalls, the other steps in. The benchmark gap is 1.1 points, the bill gap 4-fold; here too the smart pick beats the pricey one. Cache-read cost rules the agent loop.

One step down sit GPT-5.6 Sol at 77.6 for $0.544, Hunyuan H4 Preview at 74.6 for $0.418, Qwen 3.8-Max at 72.0 for $0.595. The open-design.ai page hands its best-value badge to Luna: 92% of the top score at 4% of the top cost. That ratio carries the thesis I built all along: a model 5 to 13 points behind costs 30 to 40 times less. Qwen at a $2 input and $6 output list stays pricey in this arena. Paying per token and paying per finished job are different things; the table teaches that.

The final three rows close the table: Gemini 3.8 Flash 68.9 at $0.210, Muse Spark 1.3 66.6 at $0.267, Kimi K3 65.7 at $0.614. The gap to the top runs 21 to 24 points, with bills 13 to 38 times Luna. I read these rows as a warning, not a rejection: names that shine on general LLM leaderboards can look pale in a design showcase. When the task type changes, the benchmark order changes too. Not every model fits every showcase.

My selection logic

My rule condenses to three lines: I try Luna or Haiku 5.5 on short verifiable jobs, I stay with Luna past 100K tokens, and I bring Sol or Sonnet to the showcase final. The routing pattern in the beam.ai review says the same: the cheap model tries, a judge verifies, stuck work escalates to the big one. What the open-design.ai table taught me: customers rarely see the gap between 90 and 83 points, but they see the 24-fold bill gap every month. Skill in AI spend is picking the right tier, not the top score.