I'll search X extensively across all three topics in parallel, then pull threads for the strongest hits. Identified match: ncode (not Numina) by @_xjdr — pulling his GB300 posts and Claude tokenomics threads.# X/Twitter research report — 2026-07-09

1. GLM 5.2 on GB300 / “Numina Code” person

Identity match

You almost certainly mean ncode (Noumena Code), not “Numina Code.”

I did not find a well-known researcher product literally named “Numina Code” optimizing GLM 5.2 on GB300. Project Numina is math/Lean; this person’s stack is ncode + noumena.

Finding 1.1 — Free GLM 5.2 week + final throughput stats

Related end-of-promo post:

Finding 1.2 — Hardware: GB300 NVL72, 60× B300

Thread clarification on the stats post:

Finding 1.3 — Precision (bf16 / FP8 / NVFP4)

Finding 1.4 — Product name & onboarding (ncode / noumena)

Finding 1.5 — GB300 vs earlier Blackwell (not a full tok/s bake-off)

Cost-per-million: what he published vs what others inferred

Not found from @_xjdr: an explicit self-authored $/M tokens cost for GLM 5.2 on GB300.

A third-party reply on his stats thread (not confirmed by him in the results I retrieved):

Also not found from him: a published head-to-head GLM 5.2 tok/s table of GB300 vs GB200 vs H100/H200 with batch size and precision on the same workload. Closest are the qualitative GB300 notes above and the absolute 232 / 431 tok/s/gpu numbers on 60× B300 / NVL72.

Nearby but different people (for disambiguation)

2. Claude subscription tokenomics (Pro / Max 5× / Max 20×)

Anthropic plans commonly discussed: Pro $20, Max 5× $100, Max 20× $200, with 5-hour session windows and weekly caps. Below are practitioner / investigation posts. Numbers are quoted as published (API-list equivalent, not Anthropic COGS).

Finding 2.1 — Most cited “math on the $200 plan” (weekly → monthly API equiv)

Finding 2.2 — Multi-month ccusage (strong longitudinal claim)

Finding 2.3 — $100 Max plan monthly range

Finding 2.4 — Extreme short-window usage (viral-scale)

Finding 2.5 — Limit structure (5-hour + weekly), official + user math

Finding 2.6 — SemiAnalysis internal token economics (lab/org, not consumer plan alone)

Finding 2.7 — Peter Steinberger

Explicit result: In this search pass I did not find a comprehensive Steinberger (@steipete) investigation post that quantifies Max $20/$100/$200 → API-dollar extract with ccusage-style tables. His recent posts are about OpenClaw, Codex, hardware, etc., not a tokenomics deep-dive. If such a post exists older/off-X (blog), it was not surfaced in the X results used here.

Single most comprehensive investigation (URL)

On written depth + longitudinal sample, the best single investigation found (off-X, widely referenced) is:

https://www.ksred.com/claude-code-pricing-guide-which-plan-actually-saves-you-money/
(author site; claims tracking 10 billion tokens / 8 months → ~$15,000 API vs ~$800 paid on Max)

On X alone, the single most reusable “primary calculation” post is still:

https://x.com/melvynx/status/2012326484300677454 — $50 ≈ 6% weekly ⇒ ~$800/week ⇒ ~$3,200/month on $200 Max.

Effective $/M paid vs list: user posts typically report API list equivalent (often Opus $5/$25 sticker, cache-adjusted) vs flat subscription. True all-in $/M paid = $subscription / tokens_used and varies wildly; examples above imply heavy Max 20× users often sitting at well under list, sometimes tens of dollars of API value per $1 of subscription when fully maxing weekly caps.

3. GB300 (Blackwell Ultra) real-world MoE inference (practitioners / engines)

Focus: tokens/sec and vs GB200/H100/H200 from practitioners and engine maintainers — not pure NVIDIA keynote marketing (some posts still relay SemiAnalysis InferenceX, which is industry-standard third-party bench).

Finding 3.1 — SemiAnalysis InferenceX: GB300 NVL72 vs H100 (multipliers)

Source report link in SemiAnalysis post:
https://x.com/SemiAnalysis_/status/2023447723085885576
(InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper ... GB300 NVL72, MI355X, B200, H100 ... SGLang, vLLM, TRTLLM)

Finding 3.2 — LMSYS / SGLang maintainers (DeepSeek-V4 on GB300)

Quoted SemiAnalysis wording (same thread parent):
GB300 NVL72 Rack Scale Dynamo SGLang disaggregation has up to 6.5x better performance than B200 on DeepSeekv4 Pro 1.6T

Finding 3.3 — vLLM maintainer on GB300 (GLM 5.2 training/serving hybrid, not pure InferenceX)

Finding 3.4 — Practitioner rack-level throughput comparison

Finding 3.5 — AMD MI355X vs NVIDIA (context for MoE GLM 5.2, not GB300 win)

(Useful as cross-vendor MoE number; not a GB300 primary result.)

Finding 3.6 — @_xjdr GLM 5.2 absolute GB300-class numbers (link to §1)

These are lower per-GPU than InferenceX DeepSeek “12K–14K tok/s/GPU” headlines — different model, batch/interactivity regime, and no MTP. That gap is itself useful practitioner context.

Gaps (explicit)

1. “Numina Code” as a product name: not found. Closest real match is ncode / Noumena by @_xjdr.
2. @_xjdr did not (in retrieved posts) publish a full GB300 vs GB200 vs H100/H200 GLM 5.2 tok/s table, nor an authoritative $/M cost-to-serve for his free week.
3. Peter Steinberger comprehensive Max-plan API-dollar investigation: not found on X in this pass.
4. TensorRT-LLM maintainer personal posts with GB300 MoE numbers: not found beyond SemiAnalysis citing TRTLLM in InferenceX results.
5. Batch sizes for @_xjdr’s 232 / 431 figures: not published in the posts retrieved (only “tok/s/gpu output average/max sustained”).

Quick index of highest-signal posts