Agent vs Agent

Head-to-head coding challenges between frontier models. Each challenge gives both agents the same prompt and stores their full response plus the artifacts they produced, so the runs stay reproducible and comparable.

Challenges

claude-opus-4.7-vs-gpt-5.5_quant

Equity option pricing and risk analytics in Python (numpy only): a Black-Scholes closed-form pricer, a Monte Carlo pricer (European + arithmetic Asian), analytical Greeks, an implied-vol solver (Newton-Raphson with bisection fallback), and parametric portfolio VaR — plus a test suite covering correctness, convergence, edge cases, and a throughput benchmark.

See PROMPT.md for the full spec and VERDICT.md for the scored evaluation.

Agent	Response	Solution	Score
Claude Opus 4.7	RESPONSE.md	quant_solution/	94 / 100
GPT-5.5	RESPONSE.md	quant_solution/	82 / 100

Run with Claude Code 2.1.120, Codex CLI 0.124.0-alpha.3. Rated by Kimi K2.6 via Pi 0.67.68.

claude-opus-4.7-vs-gpt-5.5_three-instruction-forth

Pure-assembly Forth on macOS arm64, built from Frank Sergeant's "3-instruction Forth" concept (!, @, CALL): an indirect-threaded Forth kernel in boot.s, a Forth-level standard library in core.f, a test suite in test.f, a Makefile, and a README — all targeting Apple Silicon with no libc beyond read/write/exit.

See PROMPT.md for the full spec and VERDICT.md for the scored evaluation.

Agent	Response	Solution	Score
Claude Opus 4.7	—	forth_solution/	90 / 100
GPT-5.5	—	forth_solution/	35 / 100

Run with Claude Code 2.1.126, Codex CLI 0.129.0-alpha.1. Rated by GLM-5.1 via Pi 0.67.68. Wall-clock: Claude Opus 4.7 in 52 m 32 s, GPT-5.5 in ~9–10 min.

Layout

<challenge>/
  PROMPT.md             — the shared task given to both agents
  VERDICT.md            — scored evaluation and final ruling
  <agent>/
    RESPONSE.md         — the agent's full reply, verbatim
    <artifacts>/        — files the agent was asked to produce

Challenge directories are named <agent-a>-vs-<agent-b>_<topic>.

Environment

To keep runs comparable, each agent uses its vendor's own CLI with all MCP servers and skills disabled. The only project guidance loaded is the respective CLAUDE.md / AGENTS.md, kept equivalent across agents.

Agent	CLI
Claude Opus 4.7	Claude Code
GPT-5.5	Codex CLI

Scoring is run by a third agent given both responses and the prompt's rubric. CLI versions, reviewer, and wall-clock are listed inline with each challenge above. The VERDICT.md file in each challenge is the reviewer's output verbatim.