A living version of Appendices A and B — model tier pricing, cache discounts, and provider differences — plus a cross-provider table covering Anthropic, OpenAI, Google, and xAI that the book doesn't print. Updated as providers change their pricing; the book's printed tables are a snapshot, this page isn't.
Representative rates by capability tier. All figures per million tokens.
| Tier | Input $/M | Output $/M | Cached $/M | Typical Latency (500 tok) |
|---|---|---|---|---|
| Fast (Haiku-class) | $1.00 | $5.00 | $0.10 | ~300ms |
| Mid (Sonnet-class) | $3.00 | $15.00 | $0.30 | ~800ms |
| Premium (Opus-class) | $5.00 | $25.00 | $0.50 | ~1,500ms |
Anthropic pricing, current as of this edition. Substitute the equivalent tier from your provider — the ratios (output ≈5x input, cached ≈0.1x input) are the durable part.
The book's tables use Anthropic's Fast/Mid/Premium framework throughout — it doesn't print competitor $/M rates, since they'd be stale before the ink dried. This table does, and gets updated when providers change theirs.
| Provider | Model | Tier | Input $/M | Cached $/M | Output $/M |
|---|---|---|---|---|---|
| Anthropic | Claude Haiku 4.5 | Fast | $1.00 | $0.10 | $5.00 |
| OpenAI | GPT-5.6-Luna | Fast | $1.00 | $0.10 | $6.00 |
| Gemini 3.1 Flash-Lite | Fast | $0.25 | $0.025 | $1.50 | |
| xAI | Grok 4.1 Fast | Fast | $0.20 | — | $0.50 |
| Anthropic | Claude Sonnet 5 (intro, thru Aug 31 '26) | Mid | $2.00 | $0.20 | $10.00 |
| OpenAI | GPT-5.6-Terra | Mid | $2.50 | $0.25 | $15.00 |
| Gemini 3.5 Flash | Mid | $1.50 | $0.15 | $9.00 | |
| xAI | Grok 4.3 | Mid | $1.25 | — | $2.50 |
| Anthropic | Claude Opus 4.8 | Premium | $5.00 | $0.50 | $25.00 |
| OpenAI | GPT-5.6-Sol | Premium | $5.00 | $0.50 | $30.00 |
| Gemini 3.1 Pro (Preview) | Premium | $2.00 | $0.20 | $12.00 | |
| xAI | Grok 4.5 | Premium | $2.00 | $0.50 | $6.00 |
| Anthropic | Claude Fable 5 (NEW) | Frontier | $10.00 | $1.00 | $50.00 |
"—" means the provider hadn't published a cached-input rate for that model at verification time. Tier assignment (Fast/Mid/Premium) is this page's mapping onto Anthropic's framework, not an official designation. Frontier is a new tier as of July 2026 — Anthropic shipped Claude Fable 5 (and a limited-availability Mythos 5) at roughly 2x Opus pricing; OpenAI, Google, and xAI haven't released a comparable above-flagship tier yet. Claude Sonnet 5's introductory pricing ($2/$10) reverts to standard pricing ($3/$15, cached $0.30) on September 1, 2026.
Pricing, caching behavior, and model capabilities vary across providers. These differences affect architecture decisions.
| Feature | Anthropic (Claude) | OpenAI (GPT) | Google (Gemini) |
|---|---|---|---|
| Prompt caching | Automatic with cache breakpoints | Manual with cached_content | Context caching via API |
| Cache pricing | 90% discount on cached input | 50-90% discount (90% on newest models) | 75% discount on cached input |
| Cache TTL | 5 min default (auto-extends on use); 1 hr optional | Automatic; retained up to 24 hr on current models | Up to 1 hour |
| Output cost ratio | 5x input cost, uniform across tiers | ~8x input cost on flagship models | 2-4x input cost |
| Streaming | SSE with tool use | SSE with function calling | SSE with grounding |
| Max context | 1M (200K on fast tier) | 1M+ across current model family | 1M+ tokens |
| Fast tier model | Haiku | GPT mini / nano class | Flash |
Verify output cost ratio and cache discount against the provider's current price sheet before relying on it for planning.
Prompt caching mechanics differ enough to change how you structure requests, not just what they cost.
| Feature | Anthropic (Claude) | OpenAI (GPT-5 series) | AWS Bedrock |
|---|---|---|---|
| Minimum cache size | Model-dependent, ~1K–4K tokens | 1,024 tokens | Model-dependent, ~1K–4K tokens (mirrors Anthropic) |
| Default TTL | 5 min (1-hour extended tier available) | Up to 24 hours (default since mid-2026) | 5 min (1-hour extended tier available) |
| Max breakpoints | 4 | Automatic | Varies by model |
| Cache marker syntax | "cache_control": {"type": "ephemeral"} | Automatic prefix caching | "cache_control": {"type": "ephemeral"} |
| Pricing: cache write | 1.25x input price (5-min TTL); 2x (1-hour TTL) | Included | 1.25x input price (5-min TTL); 2x (1-hour TTL) |
| Pricing: cache read | 0.1x input price | 0.1x input price (~0.25x on GPT-4.1-era models) | 0.1x input price |
Pricing and features change. Verify with your provider's current documentation before implementation.
We refresh this page whenever a provider changes pricing or cache behavior in a way that matters for the book's numbers — and we'll let you know when the book itself is live.
These three tables are Appendices A and B, plus one table from Chapter 5. The book has 59 more — including the reference architecture that turns these numbers into a deployable cost model.
Get notified when it's live →