Appendix B · dated September 2026

State of the field

The facts in this book most likely to change, kept in one place so the chapters can stay timeless.

How to read this appendix

The chapters explain mechanisms, which change slowly. This page lists the specifics that change monthly: model names, sizes, prices, context lengths, hardware. Every row carries a claim marker, and one more distinction matters here. Rows marked public come from a primary source: a technical report, a model card, or official documentation. Rows in the last table come from secondary reporting that was not checked against a primary source when this page was written; treat them as leads, not facts. Last reviewed: 2026-09-27.

The Claude API, as used in the book's code

The book's code uses claude-opus-5 for Dispatch and claude-haiku-4-5 as a cheap judge. Model ids and prices below are from the Claude API documentation bundled with the development tools used to write the book, as cached on 2026-06-24 public. Prices are US dollars per million tokens. Check the provider's pricing page before relying on them.

ModelModel idContextInputOutput
Claude Fable 5.1claude-fable-5-11M$10.00$50.00
Claude Opus 5.5claude-opus-5-51M$4.00$20.00
Claude Opus 5claude-opus-51M$5.00$25.00
Claude Sonnet 5claude-sonnet-51M$2.00$10.00
Claude Haiku 4.5claude-haiku-4-5200K$1.00$5.00

Related rules the chapters rely on, from the same documentation public:

Other closed frontier families

The other large developers of closed frontier models publish product documentation, pricing, context lengths, and system cards, and do not publish parameter counts, training data, or architectures public for what they publish, unknown for the rest. This book therefore uses open models for every internal number. For current model names, prices, and context lengths, use each provider's documentation: OpenAI (GPT family), Google (Gemini family), xAI (Grok family), and others. Common claims across their documentation as of 2026: context windows of hundreds of thousands to millions of tokens, native image and audio input, and reasoning modes with adjustable effort inferred from provider documentation.

Open-weight families the book draws on

These are the models behind the book's public numbers, with the facts the chapters use. Each has newer versions than the ones cited; the cited ones are those with detailed technical reports.

FamilyWhat the book usesChapters
Llama 3 / 3.1 (Meta, 2024)8B, 70B, 405B dense; 128,256-token vocabulary; GQA with 8 KV heads; RoPE; 15–15.6 T training tokens; 3.8 × 10²⁵ FLOPs for 405B; 16k H100s; SFT + rejection sampling + DPO in roundsthroughout
Llama 3.2 (Meta, 2024)1B and 3B pruned from 8B and distilled; 11B and 90B vision models with cross-attention adapters25, 27
Llama 4 (Meta, 2025)mixture-of-experts models with natively multimodal input and very long advertised context12
DeepSeek-V2 / V3 (2024)multi-head latent attention; fine-grained MoE with shared experts; V3 671B total, 37B active; FP8 training11, 12, 27
DeepSeek-R1 (2025)reinforcement learning with verifiable rewards and GRPO; distillation into 1.5B–70B students with ≈ 800k samples16, 27
Qwen2 / 2.5 / 3 (Alibaba)dense and MoE models; Qwen2-VL dynamic resolution and multimodal RoPE12, 25
Mistral / MixtralMixtral 8×7B: 46.7B total, 12.9B active; sliding-window attention in early Mistral models12
Gemma 2 / 3 (Google)interleaved local and global attention; small models trained with distillation12, 27
OLMo / OLMo 2 (AI2)fully open data, code, and training logs, including documented loss spikes and fixes9, 13
Phi (Microsoft)small models trained on filtered and synthetic "textbook-quality" data27

Hardware in use

HardwareWhat the book usesEvidence
NVIDIA H100 (Hopper)80 GB HBM3, ≈ 3.35 TB/s, ≈ 1 PFLOP/s dense bf16; NVLink 900 GB/spublic spec sheet
NVIDIA H200 / B200 (Blackwell)141 GB and ≈ 180–192 GB of HBM respectively; higher bandwidth; FP8 and FP4 supportpublic spec sheets; figures vary by configuration
Google TPU, AMD MI300-seriesused for frontier training and serving by several labspublic announcements
Cluster sizes for frontier runstens of thousands of accelerators; power and data-centre capacity are now planning constraintspublic statements; exact figures mostly unknown

Recent releases reported, not yet verified

Secondary sources in August and September 2026 report the following. None was checked against a primary technical report when this page was written, so none is used in the chapters. Verify before citing inferred from the sources listed.

Sources consulted: taskade.com "10 Best Open-Source LLMs, August 2026"; hidekazu-konishi.com open-weights release timeline; llm-stats.com model updates (September 2026); computingforgeeks.com open-source LLM comparison.

What to re-check next quarter