July 19–20, 2026: @AstraiaAI posted a thread that hit ~197K views in hours:
"I'm surprised this wasn't common knowledge — Sonnet is 1T parameters. Opus is 5T parameters. Fable/Mythos is 10T parameters."
Replies split between "finally someone said it" and "src?" — @Odd_Xignal asked for better sources; @mark_k and @alessandro_a0 called the figures unfounded X/Reddit estimates. Astraia doubled down: "By Anthropic's primary compute partner. Cut the cope."
The numbers are not Anthropic official. They do rhyme with an earlier public anchor almost everyone missed when it first dropped: Elon Musk's Grok sizing comment — the same ratios Astraia now calls "common knowledge."
TL;DR — claims vs evidence tiers
| Claim | Who said it | Evidence tier | explainx.ai read |
|---|---|---|---|
| Sonnet ≈ 1T total | Astraia (Jul 19); implied by Musk (Apr 2026) | Tier 2–3 — hearsay + algebra | Plausible order-of-magnitude; unconfirmed |
| Opus ≈ 5T total | Same | Tier 2–3 | May describe Opus 4/4.1 era or marketing round-number; current 4.5/4.6 serving tier may be smaller |
| Fable/Mythos ≈ 10T total | Astraia; some researcher forums | Tier 3 only | Highest controversy — throughput math argues against 10T for what Vertex/Bedrock serve today |
| Grok ≈ 0.5T total | Musk (direct) | Tier 1 for Grok only | Confirmed by Musk; used as ratio anchor for Claude |
| Anthropic official counts | — | None published | Procurement should not cite X threads |
The Musk anchor everyone keeps re-deriving
In April 2026, @MetaMaster asked whether Grok 4.20 is 500B total or 500B active in a larger MoE. Musk replied (paraphrased from multiple mirrors including Adam Holter's breakdown and 36Kr's coverage):
"0.5T total. Current Grok is half the size of Sonnet and 1/10th the size of Opus. Very strong model for its size."
Back-of-napkin math:
| If Grok = | And Grok is… | Implied size |
|---|---|---|
| 0.5T | ½ × Sonnet | Sonnet ≈ 1T |
| 0.5T | ¹⁄₁₀ × Opus | Opus ≈ 5T |
Musk was defending Colossus 2 training seven models, including a 10T flagship — so the ratios served xAI's "efficient at scale" narrative. Netizens immediately posted "Sonnet 1T, Opus 5T"; Anthropic never commented.
Astraia's July thread adds Fable/Mythos at 10T — aligning Mythos with xAI's largest training target, not Musk's 0.5T shipping Grok. That leap is Astraia's extension, not Musk's quote.
For Grok-vs-Claude product context, see explainx.ai's Grok 4.5 vs Claude Opus comparison and Grok 4.5 private beta at SpaceX/Tesla (V9 1.5T foundation — a different numbering scheme entirely).
What Anthropic actually publishes (and doesn't)
Across Fable 5 / Mythos 5 launch, export-control restores, and July Fable limit extensions, Anthropic discloses:
- Pricing ($10/$50 per M tokens for Fable/Mythos class)
- Benchmarks (SWE-Bench Pro, etc.)
- Safety routing (Opus 4.8 fallbacks under 5% of sessions)
- Context and tool policies
Not disclosed: total parameters, active parameters per token, expert counts, training token volume, or MoE routing ratios.
That silence is strategic — parameter counts are easy to misread (total vs active), fuel competitive intelligence, and age poorly when Anthropic ships distilled tiers under familiar names.
Throughput analysis: why 10T skeptics have a case
Independent analyst unexcitedneurons (March 2026) estimated Claude Opus 4.5/4.6 using OpenRouter token/sec on Google Vertex, calibrated against open MoE models with known active params (DeepSeek V3.1, GLM-4.7, Kimi K2).
Core method: decode throughput ≈ f(active parameters loaded per token). Compare Opus vs Chinese MoE baselines on the same cloud stack — providers are not incentivized to artificially slow one model.
Findings relevant to the July debate:
| Model (served) | Throughput signal | Estimated total params (MoE-dependent) |
|---|---|---|
| Opus 4.5/4.6 | ~40–43 tps on Vertex | ~1T–3T depending on 8:160 vs 8:256 vs 8:384 routing |
| Sonnet 4.5/4.6 | Faster than Opus (~41–51 tps) | Smaller; author guesses ~1.5× ratio, not 5× |
| Casual "10T+" Opus | Would imply ~10× slower serving vs DeepSeek/Kimi | "Inconsistent with observed throughput" |
The author argues Anthropic likely pretrains multi-trillion checkpoints (Opus 4/4.1 era ~5T–6T in their reconstruction) then distills into cheaper "Opus 4.5" tiers that are closer to a heavily post-trained Sonnet — explaining why API pricing dropped 3× from Opus 4.1 ($15/$75) to Opus 4.5/4.6 ($5/$25) while the brand name stayed "Opus."
Implication for Astraia's chart:
- Sonnet 1T / Opus 5T may describe generational peaks or pre-distill checkpoints, not necessarily Fable 5 weights running in Claude Code today.
- Fable/Mythos 10T is harder to reconcile with Vertex throughput unless Fable routes vastly more sparsely than Kimi — possible in theory, not evidenced.
Fable, Mythos, and the 10T line
@Build4mBottom replied "no even mythos is 4t" — Astraia: "You are wrong." No primary document settles it.
What is official: Fable 5 and Mythos 5 share one model — Mythos lifts safeguards for Annex A cyber-defender cohorts. Parameter count, if any, would be identical; only policy differs.
Third-party synthesis (e.g. AIThinkerLab's evidence review) stacks Musk ratios, Mythos ~10T forum lore, and cost reverse-engineering into a Tier-3 bucket — explicitly not Anthropic confirmed.
@crystlizedd captured the benchmark paradox:
"oh, so fable 5 is getting beat by Kimi K3... A 3T param model?"
Moonshot's Kimi K3 platform docs cite 2.8T total, 16/896 experts active — a different point on the total vs active curve than a hypothetical 10T / ~1T-active Mythos MoE. GPT-5.6 vs Fable 5 shows frontier rankings moving on benchmarks and price, not parameter leaderboard bragging.
HN/X thread dynamics — what skeptics get right
Critics in Astraia's replies (@DBrodniak, @altanaldeniz, @virtualcity69) converge on fair points:
- No definitive proof — "primary compute partner" is unverifiable without a name, contract, or Anthropic statement.
- Estimates mutate by generation — 36Kr noted earlier industry guesses for Claude Opus 4 in the 300B–500B range — far below 5T.
- Serving economics contradict huge gaps — Sonnet at 60% of Opus API price suggests nowhere near 5× total size for current tiers (matches throughput write-up).
- Smaller open models win some tasks — parameter count is neither necessary nor sufficient for coding agents.
What skeptics underweight: Musk's ratios have stayed stable for months, repeated by multiple outlets, and match pricing/distillation narratives if "Opus 5T" means legacy pretrain not July 2026 endpoint.
Practical guidance for builders
| If you need… | Do this | Don't do this |
|---|---|---|
| Procurement slides | Cite benchmarks + $/task from Fable vs GPT-5.6 | Paste 10T from X without "unverified" label |
| Capacity planning | Measure your tokens/sec and queue depth on Anthropic API | Assume 10T dense memory footprints |
| Competitive intel | Track distillation + rename patterns (Opus 4.1 → 4.5 pricing cliff) | Treat parameter rumors as Moat fact |
| Research citations | Write: "Anthropic undisclosed; third-party estimates range ~1T–5T served, ~10T attributed to Mythos-class MoE in unverified reports" | Attribute 10T to Anthropic officially |
Summary
July 19, 2026 revived a months-old Musk algebra problem: Grok 0.5T ⇒ Sonnet ~1T, Opus ~5T. @AstraiaAI added Fable/Mythos 10T and claimed compute-partner sourcing — without verifiable primary documents.
Anthropic still publishes zero parameter counts. The strongest public counter to 10T shipping today is inference throughput (unexcitedneurons): Opus 4.5/4.6 behaves like low-trillions MoE, not an order-of-magnitude larger monster. Kimi K3 at 2.8T beating Fable on some tasks is a feature-quality story, not a disproof that Anthropic ever trained larger sparse checkpoints.
explainx.ai's position: repeat Musk's ratios with attribution; label Astraia's 10T and partner claim Tier-3; demand Anthropic disclosure or your own benchmarks before betting architecture.
Related on explainx.ai
- What Are LLM Parameters? Top 10 Model Sizes (July 2026)
- GPT-5.6 Sol, Terra, Luna vs Claude Fable 5
- Grok 4.5 vs Claude Opus 4.7/4.8
- Kimi K3 — 2.8T open frontier
- Claude Fable 5 and Mythos 5 launch
- Fable 5 limits through August 19
- Alpoge Fable 5 Jacobian claim — unverified math
Primary / secondary sources: @AstraiaAI Jul 19 thread · Adam Holter — Musk ratios · unexcitedneurons throughput estimate · PsyArXiv — cognitive surrender study (epistemic humility companion)
Parameter claims, Musk quotes, and throughput estimates reflect public posts through July 20, 2026. Anthropic may ship new tiers without updating any public cardinality. Not investment or procurement advice.