How does the FrankX swarm pick a model?

By job and architecture, not by a single leaderboard. A recommendation names the primary model, an alternate, what not to do, and the evidence kind: first-party, vendor-claim, independent composite, or not-run.

Did you run Grok 4.6 against Claude in Model Arena?

No. Model Arena only publishes SIS JSON receipts. The current harness dispatches Claude Code Agent models. Vendor and Artificial Analysis scores for Grok 4.6 stay labeled and live on /llm-hub/grok-4-6.

Is Grok Imagine the same as Grok 4.6?

No. Grok 4.6 is the reasoning model. Still-lifes came from Grok Imagine, Codex image_gen, and Antigravity generate_image.

Who wrote the Grok 4.6 brief and this hub update?

Hermes default on Grok 4.6 drafted the packet and this board. Claude Code ran a read-only review of the publish path (no n8n, arena stays receipt-gated). Frank remains the human publish gate.

Which flagship should I start with?

Match the job. Long-running mid-price agents: Grok 4.6. Receipt-gated battles: Claude Fable or Opus via Model Arena. Cheap closed chat: Grok 4.3. Google 3.5 that shipped: Gemini 3.5 Flash. GPT-5.5 stays vendor-claim until we run it.

Research hub · architecture routing

Which model for which architecture —with sources, tests, and a named author.

The swarm recommends a job and a system shape first, then a model. Vendor scores, independent composites, and first-party receipts stay labeled. Frank still publishes.

Swarm recommendations Model Arena Methodology

104

Research domains

104

Domains with sources

653+

Source references

Specialist agent roles

Swarm recommendation · as of 2026-08-16

Architecture first. Then a model.

These are routing cards for how a system should be built, not slogans about FrankX. Each row names the job, the architecture, a primary, an alternate, and the evidence kind. No SIS battle score is implied.

FrankX agent swarm model and architecture recommendations
Job	Architecture	Primary	Evidence	Read
Long-running coding and knowledge agents	One reasoning model with tools, verification loops, and a human publish gate. Do not swap the model mid-task. Do not: Do not treat Artificial Analysis Index ties as a “win.”	Grok 4.6 Alt: Claude Opus / Fable 5 when the task is long-horizon and cost is secondary	Independent composite AA Intelligence Index 61 for Grok 4.6 (vendor + AA, 12 Aug 2026). No SIS arena receipt for Grok.	Open
Receipt-gated model battles	Dispatch only models the Starlight arena harness can pin. Publish JSON or do not publish a winner. Do not: Do not invent a Grok 4.6 arena winner from vendor benches.	Claude Fable / Opus / Sonnet / Haiku Alt: Hold Grok, GPT, and Gemini until the harness can write a receipt	First-party SIS tools/arena is Claude Code Agent-native. Existing /research/model-arena cards are those receipts.	Open
Quiet product still-life	Use a native image tool already in the session. Inspect the pixels. Show only frames that pass QA. Do not: Do not route this work through FAL. Do not publish missed briefs.	Grok Imagine, Codex image_gen, or Antigravity generate_image Alt: Pick the CLI you are already in	First-party Same prompt, 16 Aug 2026. All three returned a publishable still-life. Quality id still did not appear in the Grok tool result. Video not run.	Open
Catalog, price, and compare closed models	Keep a registry with live pricing and compare pages. Routing rows must name evidence kind. Do not: Do not dump vendor tables onto Arcanea or starlightintelligence.org.	LLM Hub / AI Ops models catalog Alt: Dated analysis posts when a flagship ships	Vendor-claim Hub registry plus /llm-hub/compare/grok-4-6-vs-grok-4-3. Prices are vendor list.	Open
Lowest-cost closed frontier chat	Prefer the cheaper prior Grok when 500k context and agent stamina are not required. Do not: Do not retire 4.3 solely because 4.6 exists.	Grok 4.3 Alt: Grok 4.6 when the task is long-running	Vendor-claim Existing hub routing: 4.3 remains the cheap closed-frontier row.	Open

Flagship models

Architecture and job, not a crown. Evidence kind is on every card. Registry notes that were not re-run this week stay vendor-claim.

ga · Independent composite

Grok 4.6

Long-running agents at mid price

Post-training / agent RL refresh. 500k context.

AA Index 61 (vendor + AA, 12 Aug 2026). No SIS arena receipt.

ga · First-party

Claude Fable 5

Receipt-gated arena + long-horizon Claude work

1M context. SIS arena can pin Fable. Use when the harness must write a JSON receipt.

Model Arena cards are Claude Code Agent receipts. Launch benches remain vendor-claim.

ga · Vendor-claim

Claude Opus 4.8

Cost-secondary coding and knowledge work

1M context. Point release after 4.7. Registry notes $5/$25 unchanged.

From the in-repo model registry. Not re-benchmarked in this wave.

ga · Vendor-claim

GPT-5.5

Long-context OpenAI agents

1M context. Registry: agentic/long-context step; narrow terminal-agent edge claimed by vendor.

No first-party SIS receipt in this wave.

ga · Vendor-claim

Gemini 3.5 Flash

Shipped Google 3.5 line

1M context. Flash is the GA 3.5 model. Pro remains preview in the registry.

Antigravity generate_image is a separate native image path, not this text model.

previous · Vendor-claim

Grok 4.3

Cheaper closed-frontier chat

1M context. Keep when 4.6 agent stamina is not required.

Previous xAI flagship. Do not retire it only because 4.6 exists.

Executed tests

Only runs that produced an artifact are listed. Holds stay holds.

First-party · 16 Aug 2026

One still-life, three native generators

Grok Imagine, Codex image_gen, and Antigravity generate_image. Only QA-pass frames. No FAL.

SIS receipts · Claude-native

Model Arena

Receipt-gated battles only. No Grok 4.6 card until the harness can pin it.

Vendor + AA · 12 Aug 2026

Grok 4.6 brief

Same-scale post-training refresh. AA Index 61. Labeled scores. Not an arena winner.

Who did what

Model, role, skill, and date for this wave. Human publish still sits last.

01 · 2026-08-14 → 2026-08-16
Grok 4.6 · Hermes default / authoring
Wrote the sourced Grok 4.6 brief, hub registry row, 4.6 vs 4.3 compare page, and this recommendation board.
Skills / tools: content-swarm-production, seo-geo-public-contract-audit, estate-design-excellence
02 · 2026-08-14
Claude Code · Independent read-only review
Confirmed: do not add n8n or Railway MCP for research publishing; Model Arena stays receipt-gated; do not invent a Grok winner.
Skills / tools: claude-code -p, print mode
03 · 2026-08-15
Grok Imagine · Image backend (not Grok 4.6 text)
Returned a curated still-life. Runtime model grok-imagine-image. Quality id still not in the tool result. Video not run.
Skills / tools: hermes image_generate, image-workflow-orchestrator
04 · 2026-08-16
Codex + Antigravity · Native image peers
Same still-life brief. Codex image_gen and agy generate_image both wrote real files that passed QA.
Skills / tools: codex exec image_gen.imagegen · agy generate_image
05 · 2026-08-17
Frank Riemer · Human publish gate
Human publish gate. This wave ships only after PR 483 merges to main and Vercel deploys.
Skills / tools: GitHub PR + Vercel git integration

Questions

How does the FrankX swarm pick a model?: By job and architecture, not by a single leaderboard. A recommendation names the primary model, an alternate, what not to do, and the evidence kind: first-party, vendor-claim, independent composite, or not-run.
Did you run Grok 4.6 against Claude in Model Arena?: No. Model Arena only publishes SIS JSON receipts. The current harness dispatches Claude Code Agent models. Vendor and Artificial Analysis scores for Grok 4.6 stay labeled and live on /llm-hub/grok-4-6.
Is Grok Imagine the same as Grok 4.6?: No. Grok 4.6 is the reasoning model. Still-lifes came from Grok Imagine, Codex image_gen, and Antigravity generate_image.
Who wrote the Grok 4.6 brief and this hub update?: Hermes default on Grok 4.6 drafted the packet and this board. Claude Code ran a read-only review of the publish path (no n8n, arena stays receipt-gated). Frank remains the human publish gate.
Which flagship should I start with?: Match the job. Long-running mid-price agents: Grok 4.6. Receipt-gated battles: Claude Fable or Opus via Model Arena. Cheap closed chat: Grok 4.3. Google 3.5 that shipped: Gemini 3.5 Flash. GPT-5.5 stays vendor-claim until we run it.

Flagship Articles

Long-form investigations that preserve sources, questions, and the distinction between reported evidence and my interpretation.

meaning · longevity · ai era

Blue Zones, Ikigai, and the AI Era

Longevity research from Okinawa and Sardinia, read as a systems question: what keeps work meaningful when AI does more of it.

12 min read

New foundational program

Four personal qualities. Four research lenses.

Freedom, Mastery, Meaning, and Connection begin as autobiographical commitments. This research program asks what autonomy, expertise, meaning, belonging, and collective intelligence can responsibly add — without turning a personal constitution into a universal personality theory.

Freedom

Autonomy lens

Mastery

Expertise lens

Meaning

Coherence lens

Connection

Belonging lens

Inspect sources and limits Read the personal constitution

Recently refreshed

The domains I have most recently revisited, with source dates and unresolved questions kept close to the synthesis.

2026-08-18

Frontier Reasoning Models & Test-Time Compute

System 2 deliberate reasoning, process reward models, and inference-time search scaling

Empirical research and architectural synthesis on test-time compute scaling, reinforcement learning from verifiable rewards (RLVR), Process Reward Models (PRMs), and dynamic search mechanisms in models like OpenAI o1/o3-mini, Claude 3.7 Sonnet with extended thinking, DeepSeek R1, and Gemini 2.0 Thinking.

6 source referencesRead Brief

2026-08-18

Mixture-of-Experts (MoE) & Multi-Head Latent Attention

Sparse activation scaling, auxiliary-loss-free routing, and memory bandwidth optimization

Architectural investigation into next-generation Sparse Mixture-of-Experts (MoE) architectures, Multi-Head Latent Attention (MLA), dynamic expert routing, and hardware-aligned memory compression in frontier models like DeepSeek-V3, Mixtral, and Qwen 2.5-Max.

6 source referencesRead Brief

2026-08-18

Context Engineering & Long-Context Architecture

Million-token context windows, needle-in-a-haystack retrieval, prompt caching, and attention sink dynamics

Systematic study of context window architectures (1M to 10M+ tokens), RoPE frequency scaling, KV-cache eviction policies, prompt caching economics, and structured context curation for enterprise AI systems.

6 source referencesRead Brief

All research domains

104 research areas organized by topic. Specialist agents map the evidence and contradictions; I review what the page can responsibly conclude.

Which model for which architecture —with sources, tests, and a named author.

Architecture first. Then a model.

Flagship models

Grok 4.6

Claude Fable 5

Claude Opus 4.8

GPT-5.5

Gemini 3.5 Flash

Grok 4.3

Executed tests

One still-life, three native generators

Model Arena

Grok 4.6 brief

Who did what

Grok 4.6 · Hermes default / authoring

Claude Code · Independent read-only review

Grok Imagine · Image backend (not Grok 4.6 text)

Codex + Antigravity · Native image peers

Frank Riemer · Human publish gate

Questions

Flagship Articles

Blue Zones, Ikigai, and the AI Era

Four personal qualities. Four research lenses.

Recently refreshed

Frontier Reasoning Models & Test-Time Compute

Mixture-of-Experts (MoE) & Multi-Head Latent Attention

Context Engineering & Long-Context Architecture

All research domains

Frontier Reasoning Models & Test-Time Compute

Mixture-of-Experts (MoE) & Multi-Head Latent Attention

Context Engineering & Long-Context Architecture

Reinforcement Learning from Verifiable Rewards (RLVR)

Multimodal Reasoning & Vision-Language-Action Models

Post-Training Distillation & Speculative Decoding

Synthetic Data Curation & Automated Curricula

Diffusion Transformers & Generative Neural Video

Neural Audio Synthesis & Conversational Speech Models

Sparse Attention, State Space Models & Linear Transformers

Representation Engineering & Mechanistic Interpretability

Adversarial Robustness & Jailbreak Defense Architectures

Multilingual Frontier Intelligence & Cross-Lingual Transfer

Mathematical Theorem Proving & Formal Verification AI

Embodied Physical AI & Spatial World Models

Enterprise AI Model Strategy: Build, Fine-Tune, or Buy?

Multi-Agent Swarm Orchestration & Consensus Protocols

Agentic Memory Architectures: Episodic, Semantic & Procedural

Model Context Protocol (MCP) & Universal Tool Ecosystems

Self-Correction, Reflexion Loops & Tree-of-Thoughts

Goal-Oriented Action Planning (GOAP) & Dynamic AI Planners

Agentic Evals, SWE-bench & Trajectory Benchmarking

Agent Sovereignty, Sandboxing & Security Boundaries

Agent-to-Agent Protocols (A2A) & Interoperability Standards

Intent Architecture & Deterministic Semantic Compilers

GraphRAG: Knowledge Graphs & Relational Reasoning

Agent Skills Frameworks: Modular Packaging & L0–L5 Progression

Swarm Telemetry, OpenTelemetry & Distributed Agent Tracing

Human-in-the-Loop (HITL) Governance & Verification Gates

Autonomous Coding Agents & Full-Stack Software Engineering

Computer Use, GUI Agents & Visual Grounding

NVIDIA Blackwell & Rubin GPU Architecture

Language Processing Units (LPUs) & SRAM Silicon

Wafer-Scale Engines & Cerebras CS-3 Systems

AI Factories, Megawatt Datacenters & Grid Infrastructure

High-Speed AI Network Fabrics: InfiniBand, RoCEv2 & Optical Switching

Oracle Cloud (OCI) Superclusters & Sovereign Cloud Architecture

AI Energy Economics, Nuclear Power & SMR Micro-Grids

On-Device Edge AI Silicon & Neural Processing Units (NPUs)

Open Silicon & RISC-V AI Accelerators (Tenstorrent)

Application-Specific Transformer ASICs (Etched Sohu)

Vector Database Infrastructure & Distributed Indexing

AI Inference Optimization Runtimes & Serving Engines

Confidential Computing & Hardware-Attested GPU Security

High-Throughput AI Storage & Distributed Parallel Filesystems

Neutral Atom Quantum Computing & Rydberg Arrays

Superconducting Qubit Systems & Transmon Physics

Topological Qubits & Majorana Zero Modes

Quantum Error Correction (QEC) & Fault-Tolerant Thresholds

Quantum Machine Learning (QML) & Variational Circuits

Quantum Sensing, Atomic Metrology & Gravimetry

Post-Quantum Cryptography (PQC) & NIST Standards