20 with Hard Intelligence data
Unified ranking · lane-aware explanation
Why the ranking looks like this.
The public ranking is not a single vibe score. It orders measured entrants by a transparent overall formula while keeping Full / Agentic, SWE MVP, Hard Intelligence, cost, and reliability visible.
Overall 88.22
#1 to #20
Full + SWE + published Hard Intelligence when measured
Static HTML plus public JSON
Chart
Overall ladder
Every ranked entrant ordered by public overall score.
Overall is a lane mean, not a hidden replacement for source measurements.
Chart
Tradeoff scatter maps
Each point is one tested model at the intersection of two public telemetry axes.
Use these maps to read quality versus cost, speed, and recorded token use. Runtime and token axes are normalized per item and show coverage in hover cards; they are telemetry, not current overall score inputs.
Chart
Lane contrast
Top eight entrants with Full, SWE, and Hard Intelligence shown side by side.
Hard Intelligence is shown as its own lane so cross-lane strengths and weaknesses stay visible.
Chart
Measured cost context
Cost is shown because deployment economics matter, but it does not secretly rewrite capability scores.
Very expensive rows are not punished twice; cost is visible telemetry and part of the public interpretation.
Chart
Lane balance pressure
Largest gap between each entrant’s strongest and weakest measured major lane.
Lower pressure means a more even profile; higher pressure explains why one strong lane may not lift the overall rank by itself.
Breadth wins the top spot
GPT‑5.5 leads because its measured lanes stay high together: overall 88.22, Full 85.76, SWE 85.51, and Hard Intelligence 93.40.
Full / Agentic alone does not decide
DeepSeek V4 Flash owns Full rank #1 at 95.80, but the overall formula still checks SWE and Hard Intelligence before ordering the table.
SWE is a separate capability signal
GLM‑5.2 owns SWE rank #1 at 89.58. That lane rewards practical implementation and review behavior rather than only general prompt competence.
Hard Intelligence reshapes the table
GPT‑5.5 owns Hard Intelligence rank #1 at 93.40. That lane tests active inquiry, adaptation, repair, and authority integrity separately from Full and SWE.
The clearest drag is visible
Step 3.7 Flash has a Full/SWE average near 85.09, but Hard Intelligence is 33.99, so the blended overall lands at 68.06.
Full ranking table
Table with reasons, not just numbers.
Each row states the score formula, lane ranks, cost context, and the main reason the entrant lands at its current position.
| Rank | Model | Overall | Full | SWE | Hard IQ | Formula | Cost + telemetry | Why here |
|---|---|---|---|---|---|---|---|---|
| #1 | GPT‑5.5 | 88.22 | 85.76 #10 | 85.51 #7 | 93.40 #1 | mean(Full, SWE, Hard Intelligence) | $4.071 19.57s/item · tokens complete | Overall 88.22 uses mean(Full, SWE, Hard Intelligence). Strength signal: Hard Intelligence rank #1. Main limiter: SWE MVP at 85.51. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #2 | GPT‑5.6 Terra | 87.70 | 88.85 #5 | 86.08 #6 | 88.18 #4 | mean(Full, SWE, Hard Intelligence) | $1.867 5.75s/item · tokens complete | Overall 87.70 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 88.85. Main limiter: SWE MVP at 86.08. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #3 | DeepSeek V4 Flash | 86.08 | 95.80 #1 | 86.95 #4 | 75.48 #10 | mean(Full, SWE, Hard Intelligence) | $0.290 22.93s/item · tokens complete | Overall 86.08 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full rank #1. Main limiter: Hard Intelligence at 75.48. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #4 | GPT‑5.6 Sol | 84.57 | 86.27 #9 | 76.12 #12 | 91.31 #3 | mean(Full, SWE, Hard Intelligence) | $3.858 13.30s/item · tokens complete | Overall 84.57 uses mean(Full, SWE, Hard Intelligence). Strength signal: Hard Intelligence rank #3. Main limiter: SWE MVP at 76.12. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #5 | Claude Opus 4.8 | 84.45 | 83.32 #13 | 88.67 #3 | 81.37 #7 | mean(Full, SWE, Hard Intelligence) | $6.115 10.43s/item · tokens complete | Overall 84.45 uses mean(Full, SWE, Hard Intelligence). Strength signal: SWE rank #3. Main limiter: Hard Intelligence at 81.37. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #6 | Claude Fable 5 | 83.74 | 77.02 #19 | 81.42 #8 | 92.78 #2 | mean(Full, SWE, Hard Intelligence) | $12.285 12.37s/item · tokens complete | Overall 83.74 uses mean(Full, SWE, Hard Intelligence). Strength signal: Hard Intelligence rank #2. Main limiter: Full / Agentic at 77.02. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #7 | GLM‑5.2 | 83.68 | 86.87 #8 | 89.58 #1 | 74.58 #11 | mean(Full, SWE, Hard Intelligence) | $2.768 91.19s/item · tokens complete | Overall 83.68 uses mean(Full, SWE, Hard Intelligence). Strength signal: SWE rank #1. Main limiter: Hard Intelligence at 74.58. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #8 | Gemini 3.5 Flash | 83.38 | 88.79 #6 | 73.59 #14 | 87.75 #5 | mean(Full, SWE, Hard Intelligence) | $1.646 9.13s/item · tokens complete | Overall 83.38 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 88.79. Main limiter: SWE MVP at 73.59. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #9 | DeepSeek V4 Pro | 83.36 | 92.04 #2 | 79.68 #10 | 78.38 #9 | mean(Full, SWE, Hard Intelligence) | $0.335 23.25s/item · tokens complete | Overall 83.36 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full rank #2. Main limiter: Hard Intelligence at 78.38. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #10 | GPT‑5.6 Luna | 82.13 | 89.90 #3 | 75.84 #13 | 80.66 #8 | mean(Full, SWE, Hard Intelligence) | $1.000 8.02s/item · tokens complete | Overall 82.13 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full rank #3. Main limiter: SWE MVP at 75.84. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #11 | Claude Sonnet 5 | 82.08 | 81.64 #15 | 78.08 #11 | 86.53 #6 | mean(Full, SWE, Hard Intelligence) | $2.812 12.51s/item · tokens complete | Overall 82.08 uses mean(Full, SWE, Hard Intelligence). Strength signal: Hard Intelligence at 86.53. Main limiter: SWE MVP at 78.08. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #12 | Qwen3.7 Max | 80.71 | 83.24 #14 | 88.99 #2 | 69.90 #13 | mean(Full, SWE, Hard Intelligence) | $0.906 22.71s/item · tokens complete | Overall 80.71 uses mean(Full, SWE, Hard Intelligence). Strength signal: SWE rank #2. Main limiter: Hard Intelligence at 69.90. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #13 | MiniMax M3 | 79.66 | 85.33 #11 | 86.88 #5 | 66.77 #14 | mean(Full, SWE, Hard Intelligence) | $0.182 27.52s/item · tokens complete | Overall 79.66 uses mean(Full, SWE, Hard Intelligence). Strength signal: SWE MVP at 86.88. Main limiter: Hard Intelligence at 66.77. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #14 | MiniMax M3 Direct Plus | 74.63 | 84.32 #12 | 67.77 #15 | 71.80 #12 | mean(Full, SWE, Hard Intelligence) | $0.0041 21.94s/item · tokens complete | Overall 74.63 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 84.32. Main limiter: SWE MVP at 67.77. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #15 | Kimi K2.7 Code | 71.01 | 87.95 #7 | 58.61 #17 | 66.46 #15 | mean(Full, SWE, Hard Intelligence) | $0.732 28.60s/item · tokens complete | Overall 71.01 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 87.95. Main limiter: SWE MVP at 58.61. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #16 | Step 3.7 Flash | 68.06 | 89.79 #4 | 80.39 #9 | 33.99 #20 | mean(Full, SWE, Hard Intelligence) | $0.494 27.77s/item · tokens complete | Overall 68.06 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 89.79. Main limiter: Hard Intelligence at 33.99. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #17 | NVIDIA Nemotron 3 Ultra | 67.58 | 79.54 #17 | 58.63 #16 | 64.56 #16 | mean(Full, SWE, Hard Intelligence) | $0.489 70.71s/item · tokens complete | Overall 67.58 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 79.54. Main limiter: SWE MVP at 58.63. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #18 | Gemma4‑12B‑Coder Fable5/Composer2.5 Q4_K_MLocal model | 61.19 | 80.60 #16 | 54.01 #18 | 48.96 #18 | mean(Full, SWE, Hard Intelligence) | $0 21.52s/item · tokens complete | Overall 61.19 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 80.60. Main limiter: Hard Intelligence at 48.96. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #19 | Ornith‑1.0‑35B Q4_K_MLocal model | 55.57 | 77.96 #18 | 39.16 #20 | 49.58 #17 | mean(Full, SWE, Hard Intelligence) | $0 117.66s/item · tokens complete | Overall 55.57 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 77.96. Main limiter: SWE MVP at 39.16. Hard Intelligence contributes to the ranking as a separate measured lane. |
| #20 | Qwythos‑9B Claude Mythos Q8_0Local model | 55.13 | 74.98 #20 | 46.51 #19 | 43.91 #19 | mean(Full, SWE, Hard Intelligence) | $0 21.13s/item · tokens complete | Overall 55.13 uses mean(Full, SWE, Hard Intelligence). Strength signal: Full / Agentic at 74.98. Main limiter: Hard Intelligence at 43.91. Hard Intelligence contributes to the ranking as a separate measured lane. |
Why the leader leads
LeaderGPT‑5.5
Overall88.22
Full85.76
SWE85.51
Hard IQ93.40
The top rank belongs to the entrant with the strongest cross-lane balance under the current formula, not simply the best isolated lane score.
How Hard Intelligence is handled
Scopeactive inquiry + adaptation + repair
Formula roleincluded when measured
Blank cellsnot yet measured
Interpretationseparate from Full and SWE
When a Hard Intelligence score is published, it becomes the third major lane in the overall mean. Otherwise the row remains ranked by the measured lanes it has.
How to compare close rows
Overallfirst glance
Lane ranksdiagnosis
Costruntime context
Reliabilityoperational risk
Close overall scores should be read through the lane breakdown. A model can be strong for building software while weaker at active inquiry, or the reverse.