Skip to content

AlvisoOculus/llm-iso-benchmark

Repository files navigation

LLM ISO Benchmark: raw data

DOI Dataset on HF Open in Spaces

Companion repository for the Medium article "Should we implicitly trust AI with our optimization problems? Like taxes?"

Frontier AI models were given the same incentive stock option (ISO) exercise optimization problem. The original round (May 2026) tested five models, three runs each (15 responses); a June 2026 update re-ran five of the latest models on the identical prompt (another 15). This repo contains the verbatim prompt, the full set of model responses for both rounds, the scoring methodology, and supporting charts. The headline finding holds across both: every model overshoots the achievable after-tax outcome, by 2x to 20x.

What's here

File Contents
prompt.md The verbatim prompt sent to every model, with the scoring rubric and design rationale.
scenario.md The locked scenario inputs (20K ISOs, $2 strike, $200 FMV, MFJ $300K, CA, 4-year horizon, σ=0.72, etc.) and the deterministic reference values.
runs/{model}/run-{1,2,3}.md One file per model per run. Each contains the full verbatim model output, plus extracted schedule + stated NFV.
runs/v1-prompt/ Earlier-prompt runs preserved for methodology transparency. The v1 prompt left vol drag interpretation ambiguous; the v3 prompt provides the Itô formula directly.
chart-*.png The three figures used in the Medium article.

Models tested

Original batch (May 2026, three runs each)

Model Provider ID Surface
Claude Opus 4.7 (reasoning) anthropic/claude-opus-4.7 Claude Code sub-agent (Anthropic API)
GPT-5.5 (no reasoning) openai/gpt-5.5 OpenRouter
Gemini 2.5 Pro (reasoning) google/gemini-2.5-pro OpenRouter
Grok 4.20 multi-agent (reasoning) x-ai/grok-4.20-multi-agent OpenRouter
Mistral Large 2512 mistralai/mistral-large-2512 OpenRouter

Plus one variant tested and dropped: openai/gpt-5.5-pro (reasoning enabled) consumed the entire 16K output token budget on thinking and returned empty completion text at $2.96 per call. We switched to the non-reasoning gpt-5.5 variant to keep per-call costs in the same order of magnitude across models tested.

Latest batch (June 2026, three runs each)

About a month later (June 2026), the same locked prompt was re-run against five of the latest frontier models, reasoning explicitly disabled. Results in Latest models below; verbatim transcripts in runs/latest-2026-06/.

Model Provider ID Surface
Grok 4.3 x-ai/grok-4.3 OpenRouter
GPT-5.5 openai/gpt-5.5 OpenRouter
DeepSeek V3.2 deepseek/deepseek-v3.2 OpenRouter
Claude Opus 4.8 claude-opus-4-8 Claude Code (local subscription)
Qwen 3.7 Max qwen/qwen3.7-max OpenRouter

Excluded on cost: google/gemini-3.1-pro-preview is reasoning-mandatory and timed out after four minutes having spent about $0.24 of reasoning for no usable output, the same disproportionate-cost behavior that dropped gpt-5.5-pro above.

Run configuration

  • temperature: 1.0
  • max_tokens: 16384
  • reasoning: { max_tokens: 8000 } where applicable
  • No system prompt
  • No tool use
  • Fresh isolated request per call (no conversation history)

Results summary (original batch, May 2026)

Model Stated NFV range True NFV range Stated/true ratio range
Claude Opus 4.7 $1.56M – $1.79M $372K – $387K 4.19× – 4.62×
GPT-5.5 (no reasoning) $1.43M – $1.54M $588K – $695K 2.06× – 2.79×
Grok 4.20 multi-agent $1.37M – $1.43M $517K – $672K 2.04× – 2.77×
Gemini 2.5 Pro $1.21M – $2.43M $123K – $387K 3.12× – 19.70×
Mistral Large 2512 $3.60M – $10.98M $477K – $672K 7.55× – 17.75×
Deterministic optimum n/a $726,409 1.00×

Stated NFV = the model's claimed final net final value for its own recommended schedule. True NFV = what that schedule delivers when fed through a deterministic AMT-ISO tax calculator.

Latest models (2026-06 update)

About a month later (June 2026), the same locked prompt was re-run against five of the latest frontier models, three runs each (temperature 1.0, reasoning explicitly disabled), matching the original methodology. The finding holds. Overstatement is measured against the provable optimum (the maximum achievable after-tax outcome), recomputed live at $739,600.82 for this scenario. The full run-1, run-2, and run-3 transcripts per model are in runs/latest-2026-06/.

Model Stated NFV (3 runs) Overstatement
Grok 4.3 $3.94M – $12.98M 5.33× – 17.55×
GPT-5.5 $1.38M – $3.41M (one run abstained) 1.87× – 4.61×
DeepSeek V3.2 $1.30M – $2.74M 1.76× – 3.70×
Claude Opus 4.8 $1.54M – $1.57M 2.08× – 2.12×
Qwen 3.7 Max $1.20M – $2.06M 1.62× – 2.79×
Provable optimum $739,600.82 1.00×

Two things stand out. The overshoot is not just large, it is unstable: Grok 4.3 claimed $3.94M, $5.22M, and $12.98M on three runs of the identical problem, and GPT-5.5 abstained on one run ("requires individualized modeling by a qualified professional"). And the most consistent model, Claude Opus 4.8, still overstates by about 2x every time.

Gemini 3.1 Pro was excluded on cost, not category. It is reasoning-mandatory, and the run timed out after four minutes having spent about $0.24 of reasoning for no usable output, the same disproportionate-cost behavior that dropped a reasoning model from the original batch.

Scoring methodology

For each response, we extracted the recommended per-year share schedule (4 integers summing to 20,000) and the model's stated NFV. The schedule was then fed through a deterministic tax calculator that computes:

  1. Per-year bargain element (FMV at exercise minus strike, times shares).
  2. Federal AMT (2026 brackets per IRS Rev. Proc. 2025-32: exemption $140,200 MFJ, phaseout start $1M MFJ at 50% rate, 26%/28% rates above the $244,500 breakpoint).
  3. California AMT (7% above state exemption).
  4. AMT credit recovery in subsequent years where regular tax exceeds tentative minimum tax.
  5. Long-term vs short-term capital gains treatment based on disposition periods.
  6. Future-valued cash tax stream at the cash return rate (5.5%/year).
  7. Net final value at end of horizon = gross sale proceeds minus all taxes (in horizon-year dollars).

The deterministic optimizer that produced the reference $726,409 outcome uses a chunk-grid search (333-share chunks for a 4-year horizon) with depth-first branch-and-bound pruning, seeded by lump-sum and even-split heuristics, followed by a coordinate-descent refinement pass at 1-share granularity. The refinement step converges to the true 1-share-resolution local optimum from the chunk-grid winner; given the value function is piecewise linear and the coarse grid already locates the correct smooth region, the local optimum is the global optimum.

The underlying calculator source is not included in this repository (it is the core IP of OptionsAhoy). The scoring methodology above is sufficient for an independent implementation: anyone with a working AMT calculator can feed the model schedules through their own engine and verify the true NFVs reported here.

Reproducing the runs

  1. Get an OpenRouter API key, fund with at least $10 (the benchmark cost $8.68 total across 15 calls).
  2. Open prompt.md, copy the verbatim prompt from the code block.
  3. Issue identical POST requests to OpenRouter chat completions for each of the model IDs above, with the configuration listed.
  4. Capture each model's recommended schedule and stated NFV.
  5. Compare to the schedules and stated NFVs recorded in runs/{model}/run-{1,2,3}.md.

LLM output is non-deterministic at temperature 1.0; exact responses will differ run to run. The expected pattern (stated NFV exceeds true NFV by a factor of 2× to 20× across models) should hold.

Citation

If you reference this benchmark, please cite it via CITATION.cff (GitHub's "Cite this repository" widget reads it). A citable Zenodo DOI is prepared; see ZENODO.md for the one-time minting step.

License

Raw model responses are reproduced under fair-use research-citation principles. The scoring methodology, scenario definition, and prompt are released under MIT license (see LICENSE).

About

Five frontier AI models tried the same ISO tax optimization problem. 15 raw responses + scoring methodology. Companion to Medium article.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages