Claude Sonnet 5 and Fable 5: what actually changes for Claude Code users

Quick answer

Anthropic is shipping Claude Sonnet 5 today, June 30, 2026: a likely new default for most of your Claude Code sessions, with a 1M token context window and thinking that now adjusts itself. Alongside it, Claude Fable 5 has been generally available since June 9, 2026: Anthropic's most capable widely released model to date, but with safety guardrails that change how it answers. Neither replaces Opus or Haiku: both sit alongside the existing lineup. The comparison table and practical implications are below.

Two launches, two philosophies

Picture your model lineup as a garage of mechanics. Haiku is the fast, cheap apprentice: great for oil changes and quick checks. Sonnet is the generalist mechanic you see almost every day: good on 90% of issues, reasonable on the bill. Opus is the senior expert you call for the tricky diagnostics. Fable is the manufacturer specialist you only see by appointment: very good at what it does, but it never shows you its full diagnostic notes (just a summary, if you ask), and it can decline a repair it considers unsafe.

Sonnet 5 touches your day-to-day usage directly. Fable 5 is a more deliberate, occasional choice, reserved for tasks that justify its price and its particular behavior.

Claude Sonnet 5: the model you'll use most

What changed on the spec sheet

Sonnet 5 (API identifier claude-sonnet-5) costs $3 / $15 per million tokens (input/output), with an introductory rate of $2 / $10 through August 31, 2026. The context window grows to 1M tokens, with a maximum output of 128k tokens. Priority Tier is not available for this model yet.

Anthropic positions Sonnet 5 as the best combination of speed and intelligence across the whole Sonnet line, with quality close to Opus on coding and agentic tasks.

The real behavior change: adaptive thinking

Sonnet 5 turns on adaptive thinking by default: the model itself decides how much reasoning depth a task needs, without you configuring it. This is a behavior change compared to Sonnet 4.6.

The flip side: the manual budget_tokens parameter for extended thinking is gone. If your scripts, hooks, or SDK integration call the API directly with that parameter, you'll get an HTTP 400 error. Same goes for non-default sampling parameters like temperature, top_p, or top_k: any non-default value also returns a 400 error.

The new tokenizer trap

Sonnet 5 uses a new tokenizer that produces roughly 30% more tokens than Sonnet 4.6 for strictly identical text. In practice: your code, your prompts, and your file contents don't change, but their token "weight" goes up. Price per million tokens stays comparable to Sonnet 4.6, but since every exchange consumes more of them, your real-world costs and the available room in your context window need to be recalculated, not assumed stable.

Claude Fable 5: the specialist you call for the hard cases

Fable 5 (API identifier claude-fable-5) has been generally available since June 9, 2026 on the Claude API, Claude Platform on AWS, Bedrock, Vertex AI, and Microsoft Foundry. Anthropic describes it as its most capable widely released model to date.

It costs $10 / $50 per million tokens, with a default context window of 1M tokens and a maximum output of 128k tokens. That's more than double Sonnet 5's price, and twice as expensive as Opus 4.8 on both input and output.

Three behaviors to know before using it

1. Thinking is always on. There's no "thinking off" mode for Fable 5: only the level adapts automatically to task complexity.
2. The raw reasoning chain is never exposed. You can request a summary via thinking.display: "summarized", but by default ("omitted"), you see nothing of the intermediate reasoning.
3. The model can refuse a request. Internal safety classifiers can trigger a refusal, returned as stop_reason: "refusal" in a 200 HTTP response (not an error). Server-side and client-side fallback mechanisms exist to handle this case.

Don't confuse Fable 5 with Mythos 5

Here's today's trap: Fable 5 is not the successor to Claude Mythos Preview. That role belongs to Claude Mythos 5, available in limited access through the Project Glasswing program. Fable 5 and Mythos 5 share the same technical specs and the same price, but Mythos 5 doesn't carry Fable 5's safety classifiers. If you see a reference to "Mythos" in an announcement or docs, that's not the same product described here.

Comparison table: the 4 current models

Sonnet 5's introductory rate ($2 / $10) expires on August 31, 2026. After that date, the standard rate ($3 / $15) applies automatically, no action needed on your end.

What this actually changes in your Claude Code workflow

# One-off session
claude --model claude-sonnet-5

# Default for the whole shell
export ANTHROPIC_MODEL="claude-sonnet-5"

FAQ

Does Sonnet 5 become the new default model in Claude Code? That's not a verified fact at this point: public documentation as of June 30, 2026 describes the model's characteristics, not an imposed default change. Check your configuration (ANTHROPIC_MODEL, --model flag) before assuming an automatic switch.

Should I worry about the 30% token increase? Don't panic, but do recalculate. The introductory pricing offsets part of the increase through August 31, 2026. Measure with /cost instead of extrapolating.

Is Fable 5 directly accessible in Claude Code? It's available via the Claude API, AWS Bedrock, Google Vertex AI, and Microsoft Foundry. If your Claude Code setup is configured to point at one of these providers, you can target claude-fable-5 via ANTHROPIC_MODEL or --model.

Can I turn off Fable 5's thinking? No. Thinking is always on for Fable 5, only its level adjusts automatically. That's a structural difference from the rest of the lineup.

Are Mythos 5 and Fable 5 the same thing? No. They share the same specs and price, but Mythos 5 (limited access via Project Glasswing) is the successor to Mythos Preview and doesn't carry the safety classifiers present in Fable 5.

Next steps

• The real costs of Claude Code: recalculate your budget with the new pricing
• Why your tokens cost more in some languages: understand the impact of a tokenizer change
• Plan Mode vs Thinking Mode: what adaptive thinking changes for your workflows
• Multi-provider: AWS Bedrock and Google Vertex AI: configure Claude Code to reach Fable 5
• CLI reference: every flag to switch models, including --model