Cheatsheet: Choosing your model and the effort dial

Current Claude families

Family	Model ID	Input / Output (per MTok)	Context	Max output
Opus 4.8 (current flagship)	`claude-opus-4-8`	$5 / $25	1M tok	128k tok
Sonnet 4.6	`claude-sonnet-4-6`	$3 / $15	1M tok	64k tok
Haiku 4.5	`claude-haiku-4-5` (alias for `claude-haiku-4-5-20251001`)	$1 / $5	200k tok	64k tok
Opus 4.7 (legacy, same posture as 4.8)	`claude-opus-4-7`	$5 / $25	1M tok	128k tok

Ratios: Sonnet about 1.7x cheaper per output than Opus. Haiku 5x cheaper than Opus.

Opus 4.8 description (verbatim): Anthropic’s most capable model for complex reasoning and agentic coding. Per the docs at drafting; check the Models overview for current.

Default-pick rule

Sonnet by default (most production workloads).
Opus when the task is genuinely hard.
Haiku when volume matters and the task is light.
Mix-and-match across models inside one app.

Model ID convention

Generation	Form	What it is
4.6 and later (Opus 4.8, Opus 4.7, Sonnet 4.6, Opus 4.6)	Dateless	IS the pinned snapshot
Pre-4.6 (Haiku 4.5, Sonnet 4.5, Opus 4.5, Opus 4.1)	Date-suffixed (canonical)	Pinned snapshot
Pre-4.6	Dateless (alias)	Convenience pointer; can re-resolve

Production: use the pinned form. For 4.6+ that is the dateless ID; for pre-4.6 that is the date-suffixed ID.

Effort parameter

client.messages.create(
    model="claude-opus-4-8",
    max_tokens=4096,
    messages=[...],
    output_config={"effort": "medium"},  # <-- goes here, not top-level
)

Value	Use case
`max`	Frontier problems; absolute highest capability. Opus 4.8 / Mythos / Opus 4.7 / Opus 4.6 / Sonnet 4.6
`xhigh`	Long-horizon coding and agentic work. Opus 4.8 and Opus 4.7
`high` (default)	Complex reasoning, difficult coding. All supported models
`medium`	Balanced; the Sonnet 4.6 production default; the Opus 4.8 / 4.7 cost-sensitive default
`low`	Speed and lowest cost; simple tasks, subagents

Affects ALL tokens: text response, tool calls, extended thinking. At lower effort, fewer tool calls + terser preambles + combined operations.

Supported on: Opus 4.8, Mythos Preview, Opus 4.7, Opus 4.6, Sonnet 4.6, Opus 4.5. NOT Haiku 4.5.

high and not setting the parameter are equivalent.

Sane starting effort by model

Model	Production default	Notes
Sonnet 4.6	`medium`	Use `low` for chat / latency-sensitive; `high` for complex reasoning
Opus 4.8	`xhigh` (coding / agentic) or `high` (general)	API default is `high` on all surfaces; `medium`/`low` only when evals show quality holds. Docs: “the guidance for Opus 4.7 also applies to Opus 4.8”
Opus 4.7 (legacy, same posture as 4.8)	`xhigh` (coding / agentic) or `high` (general)	`medium` for cost-sensitive; `max` only if evals show headroom
Opus 4.6	`high`	`medium` for cost-sensitive
Haiku 4.5	N/A (no effort support)	What you get is what the model decides

Thinking modes

Model	Adaptive thinking	Manual extended thinking
Opus 4.8	Recommended (default)	NOT supported (returns 400)
Opus 4.7	Recommended (default)	NOT supported (returns 400)
Sonnet 4.6	Recommended	Deprecated but functional
Opus 4.6	Recommended	Deprecated but functional
Haiku 4.5	NOT supported	Supported (the only mode)

Adaptive: thinking: {type: "adaptive"} (model decides; effort controls depth). Manual: thinking: {type: "enabled", budget_tokens: N, display: "summarized" | "omitted"}.

Response includes thinking content blocks (with a signature field) before text blocks when thinking fires.

Cost shape (worked example, 100k calls/day)

Feature: 100k calls/day; classifier (500 in + 20 out tokens),
then answer for 70% (500 in + 300 out).

All-Opus:    classifier $300 + answer $700 = $1,000/day
All-Sonnet:  $180 + $420 = $600/day  (~40% cheaper)
Mix (Haiku classifier + Sonnet answer):
             $60 + $420 = $480/day  (~52% cheaper than all-Opus)

Model + parameter choice both move the dial materially. Lesson 7 (caching) adds the third lever; lesson 12 turns usage logs into dashboards.

Common pitfalls

Failure	Recognize by	Fix
Defaulting to Opus everywhere	High bill, similar quality on most calls	Eval Sonnet against Opus on real prompts; swap where they tie
Using alias in production	Surprise behavior change after Anthropic patches	Pin to date-suffixed form on pre-4.6 models
Forgetting effort affects tool calls	Loop costs compound mysteriously	Lower effort in inner loop steps; eval the quality trade
Manual thinking on Opus 4.7	400 error	Switch to adaptive thinking + effort
Mixing models without eval	Silent quality regression on Haiku step	Eval routing decisions; pin per-step model to capability need

What this lesson does NOT cover (and where to find it)

Topic	Lands at
Tool use (define + handle)	Lesson 4
Server-side tools (web search, code execution)	Lesson 5
Model Context Protocol	Lesson 6
Prompt caching (the 3rd cost lever)	Lesson 7
Agent loop (model + effort compound across steps)	Lesson 8 onward
Admin / Usage / Cost APIs (org-level dashboards)	Lesson 12

Source

Anthropic public Claude docs, Models overview, Extended thinking, Adaptive thinking, Effort, at https://platform.claude.com/docs/. Structural-mirror citation; see references.