Claude Fable 5
A new tier above Opus — Anthropic's most capable model yet
Anthropic's most capable tier above Opus — native 1M context, 128K max output, $10/$50 per MTok. After a month-long regulatory suspension, access was restored on July 15, 2026 and it is live on QCode for all plans.
On June 12, 2026, Fable 5 (and Mythos 5) were suspended worldwide under a US export-control directive. Access was restored in mid-July and QCode re-enabled Fable 5 for all plans on July 15, 2026 — selecting claude-fable-5 works, no action needed. Separately, Anthropic has tiered Fable 5 on its own subscription plans since July 19; that does not touch the API. See the availability tracker for the full dated history.
Key Highlights
Fable 5 is Anthropic's most powerful, most intelligent model to date — a brand-new flagship tier above the Opus line
Same-class 1M context window as Opus 4.8 — whole-repo analysis and long documents in one shot
Up to 128K output tokens per response (streaming required for large outputs) — big refactors and long reports in one pass
$10 input / $50 output per MTok — 2× Opus 4.8, matching the stronger capability tier
Positioning: Built for the Hardest Tasks
Anthropic positions Fable 5 as the flagship for the most demanding work — long-horizon agentic execution, deep reasoning and knowledge work beyond the Opus line
Stronger long-horizon agentic execution
More reliable than Opus 4.8 on long autonomous runs — hand it complex multi-step engineering tasks end to end
Deeper reasoning
A new ceiling for reasoning depth, tunable via effort levels (high / xhigh / max) to balance depth and cost
Code generation & refactoring
Carries forward Opus 4.8's high-reliability codegen, with stronger results on hard cross-file refactors and architecture-level tasks
Knowledge work & long documents
The flagship pick for research, analysis and writing — stays coherent across long chains within the 1M context
Multimodal understanding
Image input supported — screenshots, charts and document parsing on par with the Opus line
⭐ Benchmarks: official numbers only
Anthropic positions Fable 5 as its 'most powerful, most intelligent model'; per-benchmark scores will follow official publication — this page does not invent numbers
Specs & API Surface
Fable 5 shares the same API surface as Opus 4.7/4.8 — existing code migrates at near-zero cost
1M context / 128K output
Native 1M context plus 128K max output (streaming required for large outputs) — specs aligned with Opus 4.8
Adaptive thinking
Adaptive thinking only: thinking={type:"adaptive"}; note an explicit disabled returns 400 — simply omit the param when thinking isn't needed
Full effort range
All effort levels supported — low / medium / high / xhigh / max; high / xhigh recommended for coding and agentic work
Caching & tooling unchanged
Prompt caching, structured outputs and tool use behave as on Opus 4.8 — no changes to your cache or tool chain
Ecosystem & Tooling
xhigh effort level
Full support for the xhigh level between high and max — the recommended setting for coding and agentic tasks
/ultrareview deep code review
Claude Code ecosystem commands carry over — an independent review session with a stronger model pushes production risk even lower
Task Budgets
task_budget (beta) supports Fable 5 — long agentic loops see their token budget and self-moderate
Available in Claude Code
Select claude-fable-5 directly in Claude Code and other tools — seamless with existing workflows
Migration Guide (⭐ key section)
Opus 4.8 → Fable 5 is just a model-ID swap — same API surface, with two things to watch (the thinking param and budget).
1. One-line upgrade
Change the model ID to claude-fable-5 — same API surface as Opus 4.7/4.8, a drop-in replacement
2. Re-budget at 2×
$10 input / $50 output per MTok is 2× Opus 4.8 — run high-value tasks on Fable 5 and keep routine work on Opus 4.8
3. No explicit thinking-disabled
thinking={type:"disabled"} returns 400 on Fable 5 (still accepted on Opus 4.8); omit the thinking param instead, or just use adaptive
4. Sampling params removed
As on Opus 4.7/4.8: temperature / top_p / top_k are removed and return 400 — steer with prompting and effort
client.messages.create(
model="claude-fable-5",
thinking={"type": "disabled"} # 400 on Fable 5
)
client.messages.create(
model="claude-fable-5",
thinking={"type": "adaptive"},
effort="xhigh" # low / medium / high / xhigh / max
)
Fable 5 vs Opus 4.8 vs Sonnet 4.6
Three Claude tiers side by side: when to reach for Fable 5, and when Opus 4.8 / Sonnet 4.6 is the better deal
| Metric | Fable 5 | Opus 4.8 | Sonnet 4.6 |
|---|---|---|---|
| Positioning | Top flagship (above Opus) | Opus flagship, value workhorse | Speed-intelligence balance |
| Input $ / MTok | $10 | $5 | $3 |
| Output $ / MTok | $50 | $25 | $15 |
| Context window | Native 1M | Native 1M | 1M |
| Max output | 128K | 128K | 64K |
Using Fable 5 on QCode
Fable 5 is available now, sharing one balance and one API with Opus 4.8 and Sonnet 5 — no migration required.
Billed at official-price ratio
Charged proportionally to Anthropic's official $10/$50 pricing — same rate system as every other Claude model on the platform
effort / thinking passthrough
Full passthrough of xhigh / max effort, adaptive thinking and other Fable 5 parameters
Swap between Opus 4.8 and Fable 5 at will
Switching is a model-id change: send the hard problems to claude-fable-5 and leave the rest on Opus 4.8. Same key, same quota, no migration.
Direct China access, low latency
Multi-node smart routing with circuit-breaker fallback — avoids the instability of reaching the official API from China
Run Opus 4.8 and Fable 5 side by side on QCode
Run Opus 4.8 and Fable 5 side by side on QCode — one balance, one API. Plans from $8.57/mo; the key goes live the moment payment clears.
Related Articles
Claude Opus 4.8 Complete Guide
Capabilities and migration guide for the previous flagship, Opus 4.8
Claude Agent Teams Guide
The complete playbook for parallel multi-Claude collaboration on complex tasks
Agentic Coding Trends 2026
The fundamental shift from conversational assistance to autonomous execution