Skip to content

diagrams.md

Latest commit

 

History

History
322 lines (255 loc) · 15.7 KB

diagrams.md

File metadata and controls

322 lines (255 loc) · 15.7 KB
Raw

Visual Optimization Diagrams

Mermaid flowcharts for key cost optimization decisions. These render natively on GitHub.

Table of Contents

Claude Model Family (July 2026)

The current Claude model lineup, their positioning, and cost tiers. Mythos 5 is the limited-availability sibling of Fable 5 under Project Glasswing -- same specs and price, no safety classifiers.

Updated 2026-07-25 for the Opus 5 launch. Opus 5 (claude-opus-5, GA 2026-07-24) is the new Opus flagship at the same $5 / $25 as Opus 4.8, which moved into the Legacy accordion. Fast Mode narrowed to Opus 5 and Opus 4.8 only, both at a flat 2x. Mythos Preview retired on 2026-06-30.

flowchart TB
    subgraph GA["Generally Available"]
        direction TB
        fable5["Fable 5 (most capable)<br/>$10 / $50 per 1M<br/>1M context · 128K output<br/>Always-on adaptive thinking · No Fast Mode"]
        opus5["Opus 5 (Opus flagship)<br/>$5 / $25 per 1M<br/>1M context · 128K output<br/>Thinking ON by default · Fast Mode (beta)"]
        opus8["Opus 4.8 (legacy)<br/>$5 / $25 per 1M<br/>1M context · 128K output<br/>Adaptive thinking · Fast Mode (beta)"]
        opus7["Opus 4.7 (legacy)<br/>$5 / $25 per 1M<br/>1M context · 128K output<br/>Adaptive thinking · Fast Mode removed"]
        opus6["Opus 4.6 (legacy)<br/>$5 / $25 per 1M<br/>1M context · 128K output<br/>Extended + adaptive · Fast Mode removed"]
        opus45["Opus 4.5<br/>$5 / $25 per 1M<br/>200K context · 64K output<br/>Extended thinking"]
        sonnet5["Sonnet 5 (safe default)<br/>$3 / $15 per 1M<br/>$2 / $10 intro to 2026-08-31<br/>Extended + adaptive thinking"]
        sonnet["Sonnet 4.6<br/>$3 / $15 per 1M<br/>1M context · 64K output<br/>Extended + adaptive thinking"]
        sonnet45["Sonnet 4.5<br/>$3 / $15 per 1M<br/>200K context · 64K output<br/>Extended thinking"]
        haiku["Haiku 4.5 (budget)<br/>$1 / $5 per 1M<br/>200K context · 64K output<br/>Extended thinking"]
    end

    subgraph RP["Limited Availability (Project Glasswing)"]
        direction TB
        mythos5["Mythos 5<br/>$10 / $50 per 1M<br/>Fable 5 without safety classifiers<br/>Approved Glasswing customers only"]
        mythos["Mythos Preview<br/>Retired 2026-06-30<br/>Migrate to Mythos 5"]
    end

    classDef flagship fill:#f4d0e0,stroke:#c94a7a,stroke-width:2px,color:#222
    classDef snapshot fill:#f4e0d0,stroke:#c97a4a,color:#222
    classDef default fill:#d0e8f4,stroke:#4a8ac9,stroke-width:2px,color:#222
    classDef budget fill:#d0f4d5,stroke:#4ac96a,color:#222
    classDef preview fill:#e8d0f4,stroke:#7a4ac9,color:#222

    class fable5,opus5 flagship
    class opus8,opus7,opus6,opus45,sonnet,sonnet45 snapshot
    class sonnet5 default
    class haiku budget
    class mythos5,mythos preview
Loading

Model Positioning

Model Access Best For Why Not
Fable 5 GA on every platform (Anthropic API, Claude Platform on AWS, Bedrock, Vertex AI, Microsoft Foundry) The hardest reasoning and longest agentic runs; Mythos-class capability 2x Opus 5 pricing; always-on thinking; safety classifiers can refuse; no Fast Mode
Opus 5 GA on every platform (Anthropic API, Claude Platform on AWS, Bedrock, Vertex AI) Complex agentic coding, multi-file refactors, long autonomous runs; the current "start here" Opus Thinking is ON by default and bills as output, so the effective cost is higher than 4.8 at the same posted rate
Opus 4.8 GA (legacy) Prompts already tuned to this snapshot; the fallback target for Opus 5 cyber refusals No cost argument to stay; retires no sooner than 2027-05-28
Opus 4.7 GA (legacy) Pinned snapshots tuned to 4.7 Fast Mode was removed: speed: "fast" now returns an error
Opus 4.6 GA (legacy) Workloads tuned to the older tokenizer; stable snapshot Fast Mode was removed silently: the request succeeds at standard speed and standard rates
Opus 4.5 GA Pinned snapshots only 200K context (not 1M); no Fast Mode; retires 2026-11-24
Sonnet 5 GA Everyday development (the safe default) Stretched on complex architecture + long agentic runs
Sonnet 4.6 GA Prompts tuned to this snapshot Sonnet 5 is the current mid-tier; no cost difference at standard rates
Sonnet 4.5 GA Pinned snapshots only 200K context (not 1M); retires 2026-09-29
Haiku 4.5 GA Formatting, renaming, simple edits, file lookups Lacks reasoning depth for multi-file work; 200K context; 4,096-token cache floor
Mythos 5 Glasswing only Fable 5's capabilities without safety classifiers (approved customers) No self-serve access; use Fable 5 instead
Mythos Preview Retired 2026-06-30 (superseded by Mythos 5) No longer callable; migrate to Mythos 5

Model Selection Decision Tree

Use this to pick the right model before starting a task. Starting with Sonnet is always a safe default.

flowchart TD
    A[Start: evaluate task] --> B{"Complex architecture,<br/>long agentic run, or<br/>multi-file refactor?"}
    B -- Yes --> C["Use Opus 5<br/>$5 / $25 per 1M"]
    B -- No --> D{"Standard feature work,<br/>code review, or<br/>writing tests?"}
    D -- Yes --> E["Use Sonnet 5<br/>$3 / $15 per 1M"]
    D -- No --> F{"Simple fix, formatting,<br/>boilerplate, or<br/>file lookup?"}
    F -- Yes --> G["Use Haiku 4.5<br/>$1 / $5 per 1M"]
    F -- No --> H["Not sure?<br/>Start with Sonnet 5"]

    C -. "latency-critical?" .-> C2["Enable Fast Mode<br/>on Opus 5 or 4.8<br/>(2x rate, 2.5x OTPS)"]
    C -. "cost-sensitive?" .-> C3["Lower output_config.effort,<br/>or thinking disabled<br/>(effort high or below only)"]

    classDef flagship fill:#f4d0e0,stroke:#c94a7a,stroke-width:2px,color:#222
    classDef fast fill:#f4e0d0,stroke:#c97a4a,color:#222
    classDef mid fill:#d0e8f4,stroke:#4a8ac9,stroke-width:2px,color:#222
    classDef low fill:#d0f4d5,stroke:#4ac96a,color:#222

    class C flagship
    class C2 fast
    class C3,E,H mid
    class G low
Loading

Quick Reference

Complexity Model Cost (Input/Output per 1M) Examples
Maximum Fable 5 $10 / $50 Hardest reasoning, longest autonomous agentic runs, Mythos-class workloads
High Opus 5 $5 / $25 Architecture design, complex debugging, large refactors, long agentic runs
High (Fast Mode) Opus 5 or Opus 4.8 $10 / $50 Latency-critical urgent work (2x premium, 2.5x output tokens/sec)
High (legacy) Opus 4.7 / 4.6 $5 / $25 Pinned snapshots only. Fast Mode is gone: 4.7 errors, 4.6 silently runs standard
Medium Sonnet 5 $3 / $15 Feature implementation, code review, test writing ($2 / $10 intro through 2026-08-31)
Low Haiku 4.5 $1 / $5 Formatting, renaming, boilerplate, lookups

Opus 5 caveat. The posted rate matches Opus 4.8, but thinking is ON by default and reasoning tokens bill as output. Expect the same task to cost more until you tune output_config.effort (defaults to high) or set thinking: {type: "disabled"}, which is only legal at effort high or below -- pairing it with xhigh or max returns a 400. max_tokens caps thinking plus text combined, so raise it before running at high effort.

Session Cost Optimization Flowchart

Follow this checklist at the start of every Claude Code session to minimize waste.

flowchart TD
    A[Start session] --> B{"CLAUDE.md<br/>over 150 lines?"}
    B -- Yes --> C["Trim CLAUDE.md<br/>under 150 lines"]
    B -- No --> D
    C --> D{".claudeignore<br/>exists?"}
    D -- No --> E["Create .claudeignore<br/>(exclude node_modules,<br/>dist, lock files)"]
    D -- Yes --> F
    E --> F["Choose model<br/>by task complexity"]
    F --> G{"Task is complex?"}
    G -- Yes --> H["Use Plan Mode<br/>before coding"]
    G -- No --> I
    H --> I["Work on task"]
    I --> J["Monitor with /usage"]
    J --> K{"Context<br/>getting large?"}
    K -- Yes --> L["Run /compact"]
    K -- No --> M
    L --> M{"Starting a<br/>new task?"}
    M -- Yes --> N["Start fresh session<br/>to reset context"]
    M -- No --> J

    classDef action fill:#d0f4d5,stroke:#4ac96a,color:#222
    classDef warn fill:#f4ead0,stroke:#c9a84a,color:#222
    classDef check fill:#d0e8f4,stroke:#4a8ac9,color:#222

    class A,F,I,J check
    class C,E,H,L,N action
Loading

Key Checkpoints

  1. CLAUDE.md size -- Every line loads on every turn. Keep it under 150 lines to avoid recurring token waste.
  2. .claudeignore -- Prevents Claude from reading large generated or vendored files.
  3. Model selection -- Match model to task complexity (see decision tree above).
  4. Plan Mode -- For complex tasks, plan first to avoid expensive iterative dead ends.
  5. /compact -- Summarizes conversation history to reduce context size mid-session.
  6. Fresh sessions -- New tasks should get new sessions. Stale context from prior tasks is pure waste.

Cost Tier Strategy Map

Which strategies matter most depends on your monthly spend. Focus on high-impact changes first.

flowchart TD
    A[Monthly Claude Code spend] --> B{"Less than<br/>$50 / month?"}
    B -- Yes --> T1
    B -- No --> D{"$50 to<br/>$200 / month?"}
    D -- Yes --> T2
    D -- No --> T3

    subgraph T1["Tier 1: Basics"]
        direction TB
        C1["Select models by task complexity"]
        C2["Trim CLAUDE.md under 150 lines"]
        C3["Use /compact when context grows"]
    end

    subgraph T2["Tier 2: Intermediate"]
        direction TB
        E1["Everything in Tier 1"]
        E2["Add .claudeignore to all projects"]
        E3["Delegate searches to subagents"]
        E4["Build /compact + Plan Mode habits"]
    end

    subgraph T3["Tier 3: Full Optimization"]
        direction TB
        F1["Everything in Tiers 1 and 2"]
        F2["Set per-developer budgets"]
        F3["Run usage analyzer weekly"]
        F4["Use token estimator on large prompts"]
        F5["Evaluate Batch API for bulk work"]
    end

    classDef tier1 fill:#d0f4d5,stroke:#4ac96a,color:#222
    classDef tier2 fill:#f4ead0,stroke:#c9a84a,color:#222
    classDef tier3 fill:#f4d0d0,stroke:#c94a4a,color:#222

    class T1 tier1
    class T2 tier2
    class T3 tier3
Loading

Strategy Summary by Tier

Tier Monthly Spend Focus Areas Expected Savings
1 - Basics < $50 Model selection, CLAUDE.md trimming, /compact 15-30%
2 - Intermediate $50-200 Add .claudeignore, subagents, Plan Mode habits 30-45%
3 - Full Optimization > $200 Team budgets, usage analyzer, token estimator, Batch API 40-60%

Pricing Modifier Stack

How the various multipliers combine on top of the base $/MTok rate. Each modifier stacks multiplicatively.

flowchart LR
    base["Base rate<br/>Opus 5 $5 / $25"] --> cache{"Cache hit<br/>or write?"}

    cache -- "Cache read hit" --> cacheRead["× 0.1<br/>(90% off input)"]
    cache -- "5-min write" --> write5m["× 1.25"]
    cache -- "1-hour write" --> write1h["× 2.0"]
    cache -- "No cache" --> normal["× 1.0"]

    cacheRead --> batch{"Batch API?"}
    write5m --> batch
    write1h --> batch
    normal --> batch

    batch -- "Yes" --> batchYes["× 0.5<br/>(50% off)"]
    batch -- "No" --> batchNo["× 1.0"]

    batchYes --> region{"Platform /<br/>region?"}
    batchNo --> region

    region -- "Global / API" --> regGlobal["× 1.0"]
    region -- "Regional endpoint" --> regRegional["× 1.1<br/>(+10%)"]
    region -- "US-only data<br/>residency" --> regData["× 1.1<br/>(+10%)"]

    regGlobal --> fast{"Fast Mode?<br/>(Opus 5 or Opus 4.8 only,<br/>beta)"}
    regRegional --> fast
    regData --> fast

    fast -- "Yes (Opus 5 / 4.8)" --> fastYes["× 2<br/>(beta, ~2.5x OTPS)"]
    fast -- "Opus 4.7" --> fastErr["Error<br/>(Fast Mode removed)"]
    fast -- "Opus 4.6" --> fastSilent["× 1.0<br/>(silently standard)"]
    fast -- "No" --> fastNo["× 1.0"]

    fastYes --> final["Final $/MTok"]
    fastSilent --> final
    fastNo --> final

    classDef discount fill:#d0f4d5,stroke:#4ac96a,color:#222
    classDef premium fill:#f4ead0,stroke:#c9a84a,color:#222
    classDef expensive fill:#f4d0d0,stroke:#c94a4a,color:#222
    classDef result fill:#d0e8f4,stroke:#4a8ac9,stroke-width:2px,color:#222

    class cacheRead,batchYes discount
    class write5m,write1h,regRegional,regData premium
    class fastYes,fastErr expensive
    class final result
Loading

Stacking Examples (Opus 5 input at $5/MTok base)

Scenario Calculation Effective rate
Standard API call $5 × 1 $5.00
Cache read hit $5 × 0.1 $0.50
Batch API $5 × 0.5 $2.50
Batch + cache read $5 × 0.5 × 0.1 $0.25
Regional endpoint on Bedrock $5 × 1.1 $5.50
Regional + data residency $5 × 1.1 × 1.1 $6.05
Fast Mode (Opus 5 or 4.8, beta) $5 × 2 $10.00
Fast Mode + cache read $5 × 2 × 0.1 $1.00
Fast Mode + 5m cache write $5 × 2 × 1.25 $12.50
Fast Mode + 1h cache write $5 × 2 × 2 $20.00
Fast Mode + data residency $5 × 2 × 1.1 $11.00

Notes:

  • Fast Mode is now a flat 2x on Opus 5 and Opus 4.8 only. The old 6x tier ($30 / $150 on Opus 4.7 and 4.6) no longer exists: speed: "fast" returns an error on Opus 4.7, and Opus 4.6 accepts it but silently runs at standard speed and standard rates. Check usage.speed in the response -- if it comes back "standard", you did not get Fast Mode.
  • Fast Mode cannot combine with Batch API or Priority Tier.
  • Switching between Fast and Standard speeds invalidates the prompt cache (different speed prefixes don't share cache).
  • Cache-write/hit multipliers DO stack on top of Fast Mode rates.
  • Data residency (inference_geo: "us") only applies to Opus 4.6, Sonnet 4.6, and all later models (Opus 5 included) on Anthropic API and Claude Platform on AWS. Earlier models error if the parameter is set.
  • Not shown on the diagram: extended thinking has no multiplier -- reasoning tokens simply bill as output at the normal rate. On Opus 5 thinking is on by default, so the same request produces more billable output tokens than it did on Opus 4.8 at the identical posted rate.

Minimum Cacheable Prompt Length

A cache_control block below the model's floor is silently ignored. No error, no cache_creation_input_tokens, full input price on every turn.

Model Minimum cacheable prompt
Opus 5 512 tokens
Fable 5 / Mythos 5 512 tokens
Opus 4.8 1,024 tokens
Opus 4.7 2,048 tokens
Opus 4.6 / Opus 4.5 4,096 tokens
Opus 4.1 1,024 tokens
Sonnet 5 / 4.6 / 4.5 1,024 tokens
Haiku 4.5 4,096 tokens
Haiku 3.5 2,048 tokens

Size your cached prefix against the highest floor you route to. A 2,000-token prefix caches on Opus 5 and Sonnet 5 but not on Haiku 4.5, so a Haiku-first routing tier can quietly pay full price while your Opus tier looks fine.

Related Guides