Appendix A — What’s Changed Since Publication

Keywords

updates, changelog, errata, model updates, version changes

Appendix N: What’s Changed Since Publication

The AI field evolves rapidly. This appendix tracks significant changes since the book’s publication, helping you stay current without re-reading entire chapters. We commit to quarterly updates.

How to Use This Appendix

Model Updates: New model releases and their impact on book content
Code Updates: Changes to libraries, frameworks, and example code
Errata: Corrections to errors in the text
Chapter Annotations: Specific updates organized by chapter

Check the publication date of your edition and read updates from that point forward.

Q1 2026 Updates (Publication Edition)

This is the initial publication. No updates yet.

Model Landscape at Publication

Provider	Flagship Model	Context	Notes
OpenAI	GPT-5.2	128K	Primary examples in book
Anthropic	Claude Opus 4.5	200K	Used for complex reasoning examples
Google	Gemini 2.0 Ultra	1M	Long-context examples
Meta	Llama 3.1 405B	128K	Open-weights reference
Mistral	Mixtral-Next	32K	Efficient MoE examples

Framework Versions at Publication

Framework	Version	Notes
LangChain	0.3.x	Post-expression language refactor
LlamaIndex	0.11.x	Workflows API
vLLM	0.6.x	PagedAttention v2
DSPy	2.5.x	Stable release

Q2 2026 Updates

Currency refresh completed May 2026. Model lineups, pricing, and tool versions across the book were updated to reflect the state of the field as of late May 2026. Because frontier models now ship on a roughly monthly cadence, treat specific version numbers and prices as a snapshot; the methodology in each chapter is the durable part.

Model Changes

Anthropic — Claude Opus 4.8 / Sonnet 4.6 (replacing Opus 4.5 / Sonnet 4.5) - Opus advanced 4.5 → 4.8; Sonnet 4.5 → 4.6. Haiku remains 4.5. - Pricing: Opus input dropped sharply from $15/M to $5/M (output $75 → $25/M); Sonnet holds at $3/$15; Haiku $1/$5. Opus and Sonnet now offer a 1M-token context window. - Affected: Ch01, Ch04, Ch06, Ch11, Ch12, Ch14, Ch32, Appendices B/E/K.

OpenAI — GPT-5.5 / GPT-5.4 (replacing GPT-5.2) - GPT-5.5 (April 2026) is the flagship at $5/$30 per 1M with a ~1.05M context window; GPT-5.4 is the lower-cost frontier tier at $2.50/$15, with mini/nano variants for routed traffic. - Affected: Ch01, Ch09, Ch12, Ch13, Ch17, Ch32, Appendices A/B/K.

Google — Gemini 3.1 Pro / 3.5 Flash (replacing Gemini 3 Pro / Flash) - Gemini 3.1 Pro reaches a 2M-token context window ($2/$12 per 1M ≤200K, $4/$18 above); Gemini 3.5 Flash (May 19, 2026) is $1.50/$9 per 1M. - Affected: Ch01, Ch12, Ch13, Ch17, Appendices A/K.

Held intentionally: Meta’s open-weight line is still cited as Llama 4 Scout/Maverick — successor reporting (Llama 5 vs. a proprietary pivot) was contradictory at the time of writing, so no successor is asserted. DeepSeek references remain V3/R1 (V4 was preview-only). xAI updated Grok 4.1 → 4.3.

Code Updates

vLLM 0.15 → 0.19.x — version pins and prose updated. (v0.21.0 shipped May 15 but was yanked; 0.19.x is the safe stable line.)

PyTorch 2.9 → 2.12 and HuggingFace Transformers 4.45 → 5.9 — note Transformers v5 is a major version. Version-pin blocks in Appendix B updated accordingly.

Errata

Ch14 (Backend Engineering): the cost-aware-routing example’s pricing dict had an internally inconsistent GPT-5.5 rate; corrected to $5/$30 per 1M and the legacy gpt-5 key relabeled gpt-5.4.
Worked cost calculations in Ch01, Ch09, Ch32, and Appendix K were re-derived after price changes so the arithmetic remains correct.

Q3 2026 Updates

Planned for September 2026

Q4 2026 Updates

Planned for December 2026

Reporting Errata

Found an error? Help improve future editions:

GitHub Issues: Submit at [repository URL]/issues with the label errata
Email: [contact email]

Please include: - Chapter and page number (or section heading) - The error you found - Suggested correction (if applicable) - Your name (for acknowledgment, optional)

All verified corrections will be credited in future editions.

Version History

Edition	Publication Date	Major Changes
1.0	Q1 2026	Initial publication

Acknowledgments

Thanks to readers who reported issues and suggested improvements:

[To be populated as corrections are submitted]

A Note on Keeping Current

This appendix captures point-in-time changes, but AI engineering requires continuous learning. Recommended practices:

Follow key researchers on Twitter/X and read their papers
Subscribe to newsletters like The Batch, Import AI, and Last Week in AI
Join communities like the LangChain Discord, LocalLLaMA subreddit, and AI engineering Slack groups
Build regularly - nothing teaches like shipping systems to production

The fundamentals in this book (attention, embeddings, evaluation, architecture) change slowly. The implementations evolve rapidly. Focus your update energy on the latter.

--- title: "What's Changed Since Publication" keywords: [updates, changelog, errata, model updates, version changes] --- # Appendix N: What's Changed Since Publication {.unnumbered} The AI field evolves rapidly. This appendix tracks significant changes since the book's publication, helping you stay current without re-reading entire chapters. We commit to quarterly updates. --- ## How to Use This Appendix - **Model Updates**: New model releases and their impact on book content - **Code Updates**: Changes to libraries, frameworks, and example code - **Errata**: Corrections to errors in the text - **Chapter Annotations**: Specific updates organized by chapter Check the publication date of your edition and read updates from that point forward. --- ## Q1 2026 Updates (Publication Edition) *This is the initial publication. No updates yet.* ### Model Landscape at Publication | Provider | Flagship Model | Context | Notes | |----------|---------------|---------|-------| | OpenAI | GPT-5.2 | 128K | Primary examples in book | | Anthropic | Claude Opus 4.5 | 200K | Used for complex reasoning examples | | Google | Gemini 2.0 Ultra | 1M | Long-context examples | | Meta | Llama 3.1 405B | 128K | Open-weights reference | | Mistral | Mixtral-Next | 32K | Efficient MoE examples | ### Framework Versions at Publication | Framework | Version | Notes | |-----------|---------|-------| | LangChain | 0.3.x | Post-expression language refactor | | LlamaIndex | 0.11.x | Workflows API | | vLLM | 0.6.x | PagedAttention v2 | | DSPy | 2.5.x | Stable release | --- ## Q2 2026 Updates *Currency refresh completed May 2026. Model lineups, pricing, and tool versions across the book were updated to reflect the state of the field as of late May 2026. Because frontier models now ship on a roughly monthly cadence, treat specific version numbers and prices as a snapshot; the methodology in each chapter is the durable part.* ### Model Changes **Anthropic — Claude Opus 4.8 / Sonnet 4.6 (replacing Opus 4.5 / Sonnet 4.5)** - Opus advanced 4.5 → 4.8; Sonnet 4.5 → 4.6. Haiku remains 4.5. - **Pricing:** Opus input dropped sharply from $15/M to **$5/M** (output $75 → $25/M); Sonnet holds at $3/$15; Haiku $1/$5. Opus and Sonnet now offer a 1M-token context window. - Affected: Ch01, Ch04, Ch06, Ch11, Ch12, Ch14, Ch32, Appendices B/E/K. **OpenAI — GPT-5.5 / GPT-5.4 (replacing GPT-5.2)** - GPT-5.5 (April 2026) is the flagship at $5/$30 per 1M with a ~1.05M context window; GPT-5.4 is the lower-cost frontier tier at $2.50/$15, with mini/nano variants for routed traffic. - Affected: Ch01, Ch09, Ch12, Ch13, Ch17, Ch32, Appendices A/B/K. **Google — Gemini 3.1 Pro / 3.5 Flash (replacing Gemini 3 Pro / Flash)** - Gemini 3.1 Pro reaches a 2M-token context window ($2/$12 per 1M ≤200K, $4/$18 above); Gemini 3.5 Flash (May 19, 2026) is $1.50/$9 per 1M. - Affected: Ch01, Ch12, Ch13, Ch17, Appendices A/K. **Held intentionally:** Meta's open-weight line is still cited as Llama 4 Scout/Maverick — successor reporting (Llama 5 vs. a proprietary pivot) was contradictory at the time of writing, so no successor is asserted. DeepSeek references remain V3/R1 (V4 was preview-only). xAI updated Grok 4.1 → 4.3. ### Code Updates **vLLM 0.15 → 0.19.x** — version pins and prose updated. (v0.21.0 shipped May 15 but was yanked; 0.19.x is the safe stable line.) **PyTorch 2.9 → 2.12** and **HuggingFace Transformers 4.45 → 5.9** — note Transformers v5 is a major version. Version-pin blocks in Appendix B updated accordingly. ### Errata - **Ch14 (Backend Engineering):** the cost-aware-routing example's pricing dict had an internally inconsistent GPT-5.5 rate; corrected to $5/$30 per 1M and the legacy `gpt-5` key relabeled `gpt-5.4`. - Worked cost calculations in Ch01, Ch09, Ch32, and Appendix K were re-derived after price changes so the arithmetic remains correct. --- ## Q3 2026 Updates *Planned for September 2026* --- ## Q4 2026 Updates *Planned for December 2026* --- ## Reporting Errata Found an error? Help improve future editions: 1. **GitHub Issues**: Submit at [repository URL]/issues with the label `errata` 2. **Email**: [contact email] Please include: - Chapter and page number (or section heading) - The error you found - Suggested correction (if applicable) - Your name (for acknowledgment, optional) All verified corrections will be credited in future editions. --- ## Version History | Edition | Publication Date | Major Changes | |---------|-----------------|---------------| | 1.0 | Q1 2026 | Initial publication | --- ## Acknowledgments Thanks to readers who reported issues and suggested improvements: *[To be populated as corrections are submitted]* --- ## A Note on Keeping Current This appendix captures point-in-time changes, but AI engineering requires continuous learning. Recommended practices: 1. **Follow key researchers** on Twitter/X and read their papers 2. **Subscribe to newsletters** like The Batch, Import AI, and Last Week in AI 3. **Join communities** like the LangChain Discord, LocalLLaMA subreddit, and AI engineering Slack groups 4. **Build regularly** - nothing teaches like shipping systems to production The fundamentals in this book (attention, embeddings, evaluation, architecture) change slowly. The implementations evolve rapidly. Focus your update energy on the latter.