Ideas

GPT-6 Sol on Bidlo Bench

Matt Wolfe2 min de lectura

GPT-6 Sol passed all 90 attempts on Bidlo Bench's 30 heavy civil estimating tasks, compared with 89 for GPT-5.6 Sol. The tasks cover supplied specification changes, bid-item extraction, quantities, delivered prices, and complete record proposals. The cost to run the model per successful result was $0.00146, versus $0.00251 for the earlier model.

Compare the models

Text v0.1 · 30 estimating tasks · 3 attempts each · September 26, 2026

80%85%90%95%100%$0.0000$0.003$0.006$0.009$0.012GPT-6 Astra: 100.0% passed; $0.0053 cost / attemptGPT-6 Astra (Fast): 100.0% passed; $0.0106 cost / attemptGPT-6 Luna: 92.2% passed; $0.000076 cost / attemptGPT 5.6 Terra: 95.6% passed; $0.0013 cost / attemptGPT 5.6 Luna: 96.7% passed; $0.000168 cost / attemptGPT 5.6 Sol: 98.9% passed; $0.0025 cost / attemptClaude Sonnet 5: 96.7% passed; $0.0027 cost / attemptClaude Fable 5.1: 100.0% passed; $0.0088 cost / attemptClaude Opus 5.5: 100.0% passed; $0.0039 cost / attemptClaude Opus 5: 98.9% passed; $0.0065 cost / attemptClaude Haiku 4.5: 98.9% passed; $0.0037 cost / attemptGrok Build 0.1: 96.7% passed; $0.0032 cost / attemptGrok 4.20 Reasoning: 98.9% passed; $0.0033 cost / attemptGemini 3.5 Flash Lite: 98.9% passed; $0.0018 cost / attemptGemini 3.1 Pro Preview: 97.8% passed; $0.0068 cost / attemptGPT-6 Sol: 100.0% passed; $0.0015 cost / attemptGPT-6 SolTask success · 80–100% scaleAverage inference cost per attempt (USD) · lower is better

Select a point or model row for its measured result and uncertainty.

Each point is one tested model configuration. The line joins observed best trade-offs; it is not a curve of untested reasoning settings.

Rankings by observed contractor-task success
1Adaptive thinking; requested effort: medium100.0%90/90$0.008890/90 billed47590/90 reported4.46 sn=90 passed
1Adaptive thinking; requested effort: medium100.0%90/90$0.003990/90 billed49290/90 reported3.66 sn=90 passed
1Reasoning effort: medium100.0%90/90$0.005390/90 billed30890/90 reported2.22 sn=90 passed
1Reasoning effort: medium100.0%90/90$0.010690/90 billed30890/90 reported1.71 sn=90 passed
1Reasoning effort: high100.0%90/90$0.001590/90 billed34890/90 reported2.25 sn=90 passed
6Thinking budget: 8,000 tokens98.9%89/90$0.003790/90 billed98290/90 reported4.83 sn=89 passed
6Adaptive thinking; requested effort: medium98.9%89/90$0.006590/90 billed55890/90 reported3.32 sn=89 passed
6Requested thinking level: medium98.9%89/90$0.001890/90 billed93590/90 reported4.95 sn=89 passed
6Reasoning effort: high98.9%89/90$0.002590/90 billed32690/90 reported2.16 sn=89 passed
6Provider default; no explicit reasoning setting sent98.9%89/90$0.003390/90 billed1,63690/90 reported7.03 sn=89 passed
11Requested thinking level: medium97.8%88/90$0.006890/90 billed78490/90 reported7.15 sn=88 passed
11Requested thinking level: medium97.8%88/90Unknown88/90 billedUnknown88/90 reported7.27 sn=88 passed
13Adaptive thinking; requested effort: medium96.7%87/90$0.002790/90 billed56690/90 reported3.83 sn=87 passed
13Reasoning effort: medium96.7%87/90$0.00016890/90 billed35090/90 reported2.45 sn=87 passed
13Provider default; no explicit reasoning setting sent96.7%87/90$0.003290/90 billed1,93290/90 reported8.39 sn=87 passed
16Requested reasoning effort: high95.6%86/90Unknown89/90 billedUnknown89/90 reported2.17 sn=86 passed
16Reasoning effort: medium95.6%86/90$0.001390/90 billed32290/90 reported2.04 sn=86 passed
16Provider default; no explicit reasoning setting sent95.6%86/90Unknown88/90 billedUnknown88/90 reported5.92 sn=86 passed
19Reasoning effort: medium92.2%83/90$0.00007690/90 billed35490/90 reported2.02 sn=83 passed

Equal scores share a rank. Small differences on 30 cases do not establish broad model superiority. Costs include failed attempts; incomplete billing or token totals remain unknown. Time is conditional on success. These scores cover supplied-text estimating tasks, not general intelligence or completed plan-to-workbook jobs.

Correctness and output compliance

The earlier Sol model's single unsuccessful answer returned the correct quote total but omitted a fixed charge from the effective unit price. Both values matter when comparing offers. GPT-6 Sol passed that check on every repeat, along with the rest of the evaluated outcomes.

GPT-6 Sol passed all three repeats on 30 of 30 tasks; GPT-5.6 Sol did so on 29. This one-attempt difference does not establish a broad accuracy advantage. The observed success-rate gain was 1.1 percentage points, with a 95% interval of 0.0–3.3 points.

Compliance with the requested machine-readable format was 87/90 for GPT-6 Sol and 86/90 for GPT-5.6 Sol. Equivalent answer formats still passed the correctness checks.

Completion time and cost

All 30 tasks had a successful response from both models. The median paired time ratio was 1.12×, with a 95% interval of 0.9995–1.4883×. The interval includes 1, so the higher point estimate does not establish a slowdown.

All 90 attempts had billing records for each model. The run cost $0.13113 for GPT-6 Sol and $0.22354 for GPT-5.6 Sol. Cost per successful result includes the earlier model's failed attempt as well as the successful ones.

For this task set, GPT-6 Sol delivered every required outcome at lower inference cost. The evidence does not support choosing it because it is faster. Reading a complete plan set or delivering a checked Excel workbook will require separate tests.

Method

Text v0.1 contains 30 synthetic tasks, repeated three times across 19 models: 1,710 attempts. Models received the same supplied text in fresh contexts, using recorded model settings. No original plan files were parsed, tools used, or Excel workbooks generated. The cases are not a representative sample of customer projects.

Grading v1.3 checks required outcomes and critical errors. After inspecting outputs, we revised grading to accept narrowly defined equivalent representations, applied uniformly to every model. Compliance with the requested answer format remains separate. Task definitions, arithmetic, graders, and results received independent AI review.

Elapsed time spans client request through response parsing. Paired ratios use task-level median successful times; 95% intervals use 2,000 source-group bootstrap resamples. The 180-second execution safeguard bounded requests. Errors retain observed time; timing creates no quality tiers.

Gateway inference cost is the charge to run the model, in USD. It includes unsuccessful attempts with known billing and is not a Bidlo customer charge. See aggregate results and settings, the 19-model comparison table (CSV), and the benchmark introduction.

Archivado en: Ideas

Autor:
Matt Wolfe

Explora el mercado de licitaciones con Licitar.

Elige lo que recibes.

Elige lo que te llega.

  • Producto
  • Investigación
  • Historias