Model cheat sheet

As of August 2026. Verify against the models overview — this changes.

The table

* Sonnet 5 introductory pricing: $2 / $10 through 31 August 2026.

Claude Mythos 5 (claude-mythos-5) — same specs and pricing as Fable 5. Invitation-only for defensive cybersecurity via Project Glasswing. No self-serve access.

Cloud IDs: Bedrock uses anthropic.claude-opus-5 style; Google Cloud uses claude-opus-5; Claude Platform on AWS uses the same IDs as the Claude API.

Batch API extended output: Opus 5, Opus 4.8/4.7/4.6, Sonnet 5 and Sonnet 4.6 support up to 300k output tokens with the output-300k-2026-03-24 beta header.

Which model

The composite pattern: Haiku classifies → Sonnet handles → Opus takes only the hard cases. Typically 60–80% cheaper with equal or better quality.

Thinking defaults — memorise this

Effort defaults: high on the Claude API and Claude Code for Opus 5 and Sonnet 5. high everywhere including claude.ai for Opus 4.8.

Request shapes

# Current
client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    thinking={"type": "adaptive"},
    output_config={"effort": "high"},
    messages=[...],
)

# Legacy — 400 error on Claude 4.7+
client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=16000,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[...],
)

Model-specific behaviour

Opus 5

• Longer default responses. Effort does not reliably change visible length — prompt for conciseness.
• Self-verifies well without instruction. Remove verification instructions carried from older prompts; they cause over-verification.
• Delegates to subagents readily; may need damping.

Sonnet 5

• Very literal instruction following.
• Watch response length and thinking-depth calibration.

Opus 4.6 / 4.8

• More upfront exploration, especially at high effort. Prompts saying “if in doubt, use [tool]” will over-trigger.
• Tendency to over-engineer. Counter with an explicit scope-limiting instruction.

Opus 4.5

• When thinking is disabled, unusually sensitive to the word “think”. Use “consider”, “evaluate”, “reason through”.

All current models

• Run independent tool calls in parallel by default
• More concise and less self-congratulatory than earlier generations
• Default to LaTeX for maths
• No prefill support from 4.6 onward
• Over-trigger on aggressive prompt language (CRITICAL: YOU MUST)

Versioning

From the 4.6 generation, model IDs use a dateless format (claude-opus-5) but are still pinned snapshots, not evergreen pointers. Earlier models used dated IDs with an alias pointing at them.

Pin explicitly. Migrate deliberately. Check model deprecations.

Query capabilities programmatically via the Models API — the response includes max_input_tokens, max_tokens and a capabilities object.

Cost arithmetic

Per 1,000 requests of 3,000 input + 500 output tokens:

Add prompt caching (80% of input cached) and Sonnet 5 drops to roughly $9. Add Haiku routing and it drops further.