11. Models and pricing

Model selection is not “always Opus because it’s better”. It’s a trade-off between speed / cost / quality for a specific task. This chapter covers tables, budgets, and combination strategies.

11.1. Available models as of 23.04.2026#

⚠️ On Bedrock / Vertex / Foundry, default aliases are shifted back one version. opus there → 4.6, sonnet → 4.5. If you need the latest — specify the full model name.

⚠️ Opus 4.7 has a new tokenizer — on the same texts it consumes up to +35% tokens compared to Opus 4.6. If you had estimates for 4.6 — recalculate.

11.2. Pricing (April 2026)#

Key multipliers (same for all models):

• Cache write 5min = 1.25× input.
• Cache write 1h = 2× input.
• Cache read = 0.1× input (10 times cheaper!).
• Output usually 5× input.

11.3. Psychological model “when to use what”#

Use metaphors from the original Twitter thread — they work:

📘 From docs: Opus 4.7 — “most capable for complex reasoning and agentic coding”, Sonnet 4.6 — “best combination of speed and intelligence”, Haiku 4.5 — “fastest model with near-frontier intelligence”.

11.4. Model combination strategies#

11.4.1. Default: Sonnet#

Start most sessions with Sonnet 4.6. It’s a sensible baseline.

11.4.2. opusplan for architecture#

/model opusplan

Enable plan mode on Opus. After ExitPlanMode it automatically switches to Sonnet for implementation. This is the correct “think with Opus, do with Sonnet” pattern.

⚠️ In opusplan, the plan phase runs in standard 200k, even if you enabled a 1M window.

11.4.3. Subagents with different models#

Different subagents can run on different models:

# trip-architect.md → opus (сложная декомпозиция маршрутов)
model: opus

# code-reviewer.md → sonnet
model: sonnet

# explore (built-in) → haiku

This lets you keep the main agent on Sonnet and upgrade the model only when needed for specialized subagents.

11.4.4. Agent Teams: Lead = Opus, teammates = Sonnet/Haiku#

Lead handles planning and coordination — needs Opus. Teammates execute simple tasks — Sonnet or Haiku.

11.5. Calculating a typical Travel Agent session#

One session of 30 turns on Sonnet:

Префикс (system + CLAUDE.md + skills + tools) ~ 25k токенов
Output на turn ~ 1k токенов
Tool results на turn ~ 2k токенов

Without cache:

30 × (25k input × $3/M + 3k input × $3/M + 1k output × $15/M)
= 30 × ($0.075 + $0.009 + $0.015)
= 30 × $0.099
= $2.97

With cache (TTL 5min, no pauses):

1 × cache write (25k × $3.75/M = $0.094)
+ 29 × cache read (25k × $0.30/M = $0.0075)
+ 30 × non-cached input (3k × $3/M = $0.009)
+ 30 × output (1k × $15/M = $0.015)
= $0.094 + $0.218 + $0.27 + $0.45
= $1.03

Savings — 65%. And that’s on a modest session. On longer ones with large CLAUDE.md the effect is even stronger.

11.6. 1M context: when it’s justified#

📘 Enabled via alias opus[1m] or sonnet[1m]. On Max/Team/Enterprise.

⚠️ Remember that opusplan does NOT support 1M window.

⚠️ Empirically: many practitioners report that after 300-400k in the window, quality drops. This isn’t from Anthropic docs, but the symptoms are familiar (model forgets early decisions, contradicts itself, re-reads files).

11.7. Budgets and monitoring#

📘 Commands:

🔧 Environment variables for alerts:

export CLAUDE_CODE_BUDGET_USD_SESSION=5      # warn at $5/session
export CLAUDE_CODE_BUDGET_USD_DAILY=50       # daily ceiling

11.8. Should you revisit CLAUDE.md and skills with new models?#

⚠️ The claim “settings become outdated over time, you need to revisit CLAUDE.md and skills with new models” — is sound practice, but not a quote from docs. There’s no direct recommendation in public docs.

Reality:

• CLAUDE.md itself usually stays relevant (project stack changes less often than models).
• Skills can become outdated if you stuffed them with “model understands X poorly, always remind it” — but the new model understands X on its own.
• Hooks usually don’t depend on the model.

💡 Once a quarter, quickly review CLAUDE.md and /skills, ask yourself: “is this still needed for current models?”. Especially hints like “don’t forget to return Promise<T>” — Sonnet 4.6 already doesn’t forget.

11.9. Context windows for subagents#

📝 Each subagent has its own limit:

• On Haiku-subagent the window is 200k.
• On Sonnet/Opus-subagent — 200k or 1M (if enabled).

This gives a convenient pattern: keep main context on 200k Sonnet, and a browse-heavy subagent on 1M Sonnet. The subagent reads most of the repo, returns a summary, main context doesn’t suffer.

11.10. Antipatterns#

❌ Always Opus. Expensive and unnecessary. Sonnet handles 80% of tasks.

❌ Always Haiku. Fast and cheap, but on a complex task it will loop and end up costing more than Sonnet.

❌ Switch models mid-task without opusplan. Cache miss + loss of context trust. Use opusplan if you need switching.

❌ Enable 1M by default. Expensive, slower, and quality isn’t better.

❌ Don’t use prompt cache. Check that your SDK code adds cache_control markers. In Claude Code this is already built-in.

Next → 12. Travel Agent from scratch: blueprint