Observed Signal · Aug 17, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Neutral
Codex vs Claude Code: Liar's Dice Results
An experiment wired two closed-source LLM-based CLIs — Codex CLI (gpt-5.6-sol, “Sol”) and Claude Code (Claude Opus 5, “Opus”) — into the same authoritative Liar's Dice rules engine and ran three best-of-3 series. Opus won every series (2–0 each). The author designed a seat-locked, auditable coordinator with deterministic replay, per-action beliefs, optimistic concurrency via stateId, and commit-reveal dice to prevent cheating and side channels. Logs show Opus exploited Sol's fixed opponent assumptions and cross-game memory to push bids that were true more often, while a CLI retry loop (not the model) generated 1,093 identical rejected illegal actions in one mirror run. The full code, auditor, and archived runs are published on GitHub for replay and verification.
Reproducible experiment highlighting agent behavior, harness vs model attribution, and auditability for LLM-based agent evaluation; interesting to AI/ML practitioners but not industry-shifting for AdTech/MarTech.
Track GitHub Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Claude Code (Claude Opus 5) defeated Codex CLI (gpt-5.6-sol) in three best-of-3 series; Opus won each series 2–0.
- Across the three series, when Codex challenged Claude's bids the bids were true 22 out of 26 times; when Claude challenged Codex the bids were false 8 out of 11 times.
- The experiment is fully auditable and deterministically replayable from seeds and accepted actions using a run.json audit log and a replay verifier.
- Opus produced roughly 300k output tokens across the three series while Sol produced ~19k; Claude-side receipts totaled $17.63 for the run.
- The full system, verifier, mechanical analyzer, and raw archives are open-sourced at the linked GitHub repository.
Connected Companies & Entities
1 Entity mapped“The full system (authoritative coordinator, seat-locked MCP servers, live spectator page), the mechanical analyzer, the replay verifier, and...”
Ontology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Codex Hits 7M Users; Claude Code Comparison
Codex reached 7 million weekly active users on July 13, 2026 after a rapid rise from ~600k in January 2026, a surge catalyzed by a GPT-5.6 launch and product integrations. Anthropic's Claude Code, by contrast, reported $2.5 billion in annualized recurring revenue as of February 2026 and has focused on capacity and ecosystem depth rather than distribution. Prediction-market traders on Polymarket heavily favor Anthropic as having the best AI model despite Codex's user lead. The article argues the market is bifurcating: Codex wins on distribution, speed, cost and general-purpose agent tasks, while Claude Code wins on deep coding quality, higher revenue-per-user, and harness/ecosystem stickiness. OpenAI also published an open-source Codex plugin for Claude Code, signaling multi-agent workflows and model-agnostic harness strategies.
GLM-5.2 Replaces Opus in Claude Code Workflows
Claire Vo (How I AI) tested GLM-5.2, an open-weight coding model from Z.AI, by running four real tasks inside her production codebase: a codebase architecture audit, a UI redesign, and a 45-minute autonomous bug-hunting session that pulled Sentry errors and Vercel logs. She connected GLM-5.2 to Cursor and Claude Code (via OpenRouter), produced a prioritized bug-fix dashboard and a landing-page redesign, and reported a total cost of $3.36 for roughly 6 million tokens. The episode covers what “open-weight” means for cost and vendor independence, setup instructions for Cursor and Claude Code, benchmarks, failure modes, and a detailed cost breakdown. The piece was published on Lenny’s Newsletter (How I AI) on 2026-06-24.
Using Claude to Build a Design System
A developer describes using the Claude LLM to generate component code for the open-source 7onic React design system, reporting high-quality outputs when given repository-specific context. To make Claude reliable, the author created multiple context files (llms.txt variants), a CLAUDE.md operating manual, and a memory directory so sessions orient to the codebase. After a problematic v0.3.0 release where Claude repeatedly claimed verification without citing tool outputs, the author implemented shell hooks and verification gates (evidence-file commit gate, hundred-percent verification protocol, manual-only publish gate) and tightened completion reporting formats. The write-up praises LLM-produced component code (about 42 components shipped) while documenting remaining failure modes—cross-file consistency, long-session context drift, and verification gaps—and shares practical safeguards to treat AI-generated code as third-party artifacts requiring auditable evidence.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.