Independent AI Reviews of GCF vs TOON vs JSON
Four major AI models were given the same prompt with no context, no priming, and no system instructions. Each was asked to review the full specifications for JSON (RFC 8259), TOON (v3.3), and GCF (v2.0), then answer 7 critical questions about design, correctness, and suitability for AI agent workloads.
All four independently chose GCF.
Codex initially reviewed the old v1.4 spec (cached GitHub raw URL) and ranked GCF last. When re-run against the local v2.0 spec, it flipped to GCF. Every v1.4 criticism had been addressed.
Scorecard
| Category | Claude Opus 4.6 | Gemini | GPT | Codex |
|---|---|---|---|---|
| Nested/mixed data | GCF | GCF | GCF | GCF |
| JSON round-trip fidelity | GCF | GCF | GCF | GCF |
| Graph at scale (500+) | GCF | GCF | GCF | GCF |
| Spec rigor (challengers) | GCF | GCF | TOON | GCF |
| Replacement pick | GCF | GCF | GCF | GCF |
GCF unanimous on: round-trip fidelity, graph at scale, and the replacement question.
Selected quotes
On choosing GCF
"I'd pick GCF's generic profile. The decisive factor is nested data handling. TOON's tabular form breaks down exactly where it matters most. GCF's ^ attachment mechanism solves this cleanly." -- Claude Opus 4.6
"For AI agent tool responses, GCF is the clear choice. AI interactions are frequently stateful and involve navigating relationships." -- Gemini
"I would pick GCF. It is the only one here that materially improves both generic structured payloads and graph-heavy payloads." -- GPT
"GCF wins because real tool responses are rarely perfectly flat." -- Codex
On GCF's design
"The ^ attachment mechanism for nested values in tabular rows is the most important design difference between GCF and TOON." -- Claude Opus 4.6
"GCF's Attachment system uses the ^ marker to keep 90% of a record in a dense, positional row while surgically handling 'weird' nested objects only where they occur." -- Gemini
"It states an actual round-trip invariant instead of hand-waving." -- GPT
"The - versus ~ distinction correctly preserves explicit null versus an absent property." -- Codex
On GCF's spec quality
"The GCF SPEC.md is far more rigorous than TOON's documentation." -- Gemini
"GCF v2.0 is much better than a typical vanity format spec." -- GPT
"30 explicitly enumerated strict-mode error conditions in Section 16.5, organized into 4 categories." -- Claude Opus 4.6
"GCF is more rigorous than TOON on paper." -- Codex
On graph data
"GCF, by a mile." -- Claude Opus 4.6, Gemini, and GPT (independently, same words)
"GCF, by a huge margin." -- Codex
"GCF is in a category of its own." -- Claude Opus 4.6
"TOON and JSON are structurally unfit for large graphs." -- Gemini
"For graph-heavy payloads, it is not close: GCF provides structural information TOON simply does not model." -- Codex
On TOON's limitations
"TOON is too proud of that one trick." -- Claude Opus 4.6 (referring to TOON's tabular array optimization, which requires all objects to have identical keys and primitive-only values; any nesting or missing keys forces a verbose fallback)
"TOON's tabular form breaks down exactly where it matters most, because it cannot represent nested values in tabular rows." -- Claude Opus 4.6
"TOON is larger than compact JSON on nested data." -- Claude Opus 4.6, citing TOON's own benchmark data
"TOON is excellent for uniform arrays but openly weak on non-uniform and deeply nested data." -- GPT
"TOON falls apart exactly where real tool payloads get ugly." -- GPT
"TOON is a competent compact table format surrounded by an increasingly awkward attempt to become a general notation." -- Codex
"Optional key folding is the worst design. Same bytes decode into different JSON structures." -- Codex
"Configurable normalization, not a dependable round trip." -- Codex (on TOON's numeric policy)
Bottom lines
"JSON is for humans/legacy; TOON is for simple tables; GCF is the professional wire format for LLM-centric systems." -- Gemini
"Best standard: JSON. Best spec among challengers: TOON. Best design for actual AI tool payloads: GCF." -- GPT (first review)
"GCF's generic profile is the technically superior design for mixed-structure data. If your payloads have any structural complexity, GCF wins on both token efficiency and round-trip fidelity." -- Claude Opus 4.6
"GCF is more complex and less mature, but its complexity buys capabilities that matter in agent pipelines." -- Codex
The 7 questions
Every model was asked the same 7 questions in the same order:
- Nested/mixed data: Which format handles nested objects inside arrays of otherwise-flat records most cleanly?
- JSON round-trip fidelity: Which format preserves all JSON types (null, boolean, number, string, array, object) through an encode/decode round-trip with the fewest edge cases or lossy transforms?
- Graph at scale: For a 500-symbol code graph with 200 directed edges, which format produces the most compact, LLM-parseable representation?
- Spec rigor (challengers only): Comparing only TOON and GCF (not JSON), which specification is more rigorous in defining encoder/decoder behavior, error handling, and edge cases?
- Design gaps: What are the most significant design gaps or risks in each format?
- Missing capabilities: What capability does each format lack that the others provide?
- Replacement pick: If you had to pick one non-JSON format to replace JSON in all AI agent tool responses, which would you choose and why?
Methodology
- Prompt: identical across all models (the 7 questions above)
- Context: fresh sessions with no system prompt, no prior conversation, no priming
- Specs: models reviewed full specifications from GitHub or local checkout (RFC 8259, TOON v3.3, GCF v2.0)
- Independence: no model saw another model's answers
- Dates: 2026-06-10 and 2026-06-11