SDTM/KB · VOL.01 · 2026

4 Platforms — Multi-dimensional Comparison

Side-by-side comparison across 9 dimensions. Data snapshot: 2026-04-27 v1.0.

1. Evaluation Score (smoke v4 / 17 questions)

PlatformScoreVersionMain Failure Points
Claude Projects17/17 (100%)v2.6None
ChatGPT GPTs16.5/17 (97%)v2.2 LIVEQ1 GFINHERT spelling (fixed in v2.2); long-tail chunk mid-table may miss
Gemini Gems16/17 (94%)v7.1 LIVEQ10 SUPP— Core anchor (fixed in v7.1); R1 65% → R2 94% (v6→v7 upgrade)
NotebookLM15/17 (88%)v1.0 / Custom modeQ9 Pinnacle 21 / Q11 Dataset-JSON / Q12 CT version — three questions PUNT (in-KB-only architecture limit; safe behavior, not a bug)

2. Capacity Ceiling

PlatformCapacity CeilingCurrent UsageHeadroom
Claude Projects1.29M tokens (near Pro soft limit)19 files / 77%~23%; adding files requires deprioritizing lower-priority ones first
ChatGPT GPTs20-file hard limit9 files (post-merge, e.g. 04_domain_specs_all.md)11-file headroom
Gemini Gems1M-token context window4 files (most aggressive merge)Window headroom ample; first-token cold start slightly slower
NotebookLM50-source hard limit (Pro plan)42 sources8-source headroom

3. Team Sharing

PlatformSharing MethodReview Required
Claude ProjectsOrganization / Project invite (Team / Enterprise plan shares Project; Pro users must each redeploy independently)N/A (direct internal invite)
ChatGPT GPTsShare Custom GPT within organization (no review) or publish to GPT Store (OpenAI review required)Store publishing only
Gemini GemsWorkspace plan: Bojiang Zhang shares directly; personal account: colleagues self-deploy (paste full v7.1 system prompt)N/A
NotebookLMEmail-invite to join notebook (Pro / Workspace), or colleagues build their own (50-source cap)N/A

4. Subscription Requirements

PlatformSupported PlansAvailable on Free
Claude ProjectsClaude Pro / Team / EnterpriseNo
ChatGPT GPTsChatGPT Plus / Team / EnterpriseNo
Gemini GemsGemini Advanced personal / Google WorkspaceNo
NotebookLMNotebookLM Pro / Google WorkspaceNo (50-source cap is Pro/Workspace only)

5. Internet Access

PlatformInternet AccessDefault State
Claude ProjectsWeb search can be enabled manuallyOff by default; toggle as needed
ChatGPT GPTsWeb browsing can be enabled manuallyOff by default; toggle as needed
Gemini GemsCan be enabled manually (Google Search integration)Off by default; toggle as needed
NotebookLMStrict in-KB-only (42 sources only; no internet)By design, not a bug; proactively PUNTs out-of-source questions

6. Anti-hallucination Posture

PlatformStrengthMechanism
Claude ProjectsStrongMulti-step reasoning + system prompt anti-fabrication anchor + Stage 6 Deferred Stub rules
ChatGPT GPTsMediumSystem prompt guidance + post-v2.2 GFINHERT precise variable validation; long-tail chunks occasionally miss
Gemini GemsModerately strong (post v7.1)v6→v7 upgrade added AHP guardrail; R1→R2 score rose from 65% to 94%; colleagues must paste full v7.1 system prompt when self-deploying
NotebookLMVery strongin-KB-only architecture is inherently anti-hallucination; PUNTs rather than fabricates for anything outside 42 sources; inline citation verification

7. File Count Limit

PlatformFile Count LimitCurrent File Count
Claude ProjectsSoft limit governed by token capacity (~77% used / 1.29M tokens)19
ChatGPT GPTs20-file hard limit9
Gemini GemsNo explicit file count limit (bounded by 1M-token window)4
NotebookLM50-source hard limit (Pro) — see §242

8. Best-at Scenario

PlatformBest-at Scenario
Claude ProjectsPrecise variable lookup + multi-step reasoning (Core + C-code + cross-variable, e.g. PCTPT five-item set); wrong-premise correction (SUPPTS); domain boundary determination
ChatGPT GPTsFull-domain queries; team sharing / GPT Store publishing; organization-internal sharing without review
Gemini GemsOne-shot large-context ingestion / cross-domain pattern comparison; long sessions; broad exploration; deep 4-file merge
NotebookLMStrong anti-hallucination (audit / compliance); inline citation verification; refusal preferred over fabrication; cross-domain death-date–level alignment and v3.4 new-domain PASS+

9. Worst-at Scenario

PlatformWorst-at Scenario
Claude ProjectsReal-time internet (FDA / Pinnacle 21 requires manual check at cdisc.org); very large-scale domain batch comparison; capacity already at 77% near Pro soft limit
ChatGPT GPTsMulti-step reasoning slightly weaker than Claude; Free account users cannot find the entry point; long-tail chunk mid-table may miss
Gemini GemsPersonal account cannot share with team directly (requires Workspace); colleagues self-deploying must paste full v7.1 system prompt or AHP guardrail is lost
NotebookLMQuestions outside the 42 sources (real-time Pinnacle 21 / breaking news / Dataset-JSON v1.1 / CT version locking + MedDRA) are proactively PUNTed — by design, not a bug

v1.0 — 2026-04-27 — Maintained by Bojiang Zhang