Claude Fable 5

What it is

Claude Fable 5 is the top of Anthropic's widely released line, dated June 2026 here. Anthropic describes it as their most capable generally available model, aimed at next-generation intelligence for long-running agents, while pointing at Opus 5 for complex agentic coding. Read plainly: Fable is the tier above on raw capability, but it is not the tier Anthropic recommends by default, and a system tuned for Opus or Sonnet will not automatically get the best out of it.

The defining difference is that adaptive thinking is always on and cannot be switched off. Everything else on this sheet follows from that.

How the composite is built

Overall 90 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. Fable 5 tops the board on the axis it is built for and gives the points back on the two that reward getting things done quickly.

Coding 94

Joint second with Mythos 5, behind Opus 5's 96. Strong by any measure, but the gap runs the wrong way for the price: the cheaper model writes better code.

Terminal 90

Five points behind Opus 5. Shell work rewards a fast, decisive loop over deep deliberation, and always-on thinking is a handicap here rather than an advantage.

Reasoning 97

The highest reasoning score anywhere on this board. This is the entire argument for Fable 5 — ambiguous specifications, contradictory requirements, problems where the first plausible answer is usually wrong.

Tool use 92

Very good, four points off Opus 5 and Mythos 5. Reliable in long chains, but not the reason to pick it.

Context 95

Near the top, behind Gemini 3.6 Pro's 98. Nominally 1M tokens, but the Opus 4.7-generation tokeniser means the same text produces roughly 30% more tokens than earlier Claude models — the window holds meaningfully less prose than the number suggests.

Speed 48

Joint slowest on the board with Mythos 5. Anthropic classes it as slower outright. At 8% weight the composite barely notices; your users will.

Where it fits

As an escalation, not a default. Build the loop on Sonnet 5 or Opus 5, measure where it actually fails, and route only those cases up. Fable earns its keep on hard architectural decisions, subtle debugging, and review passes over work another model produced — batch and background jobs where latency is a scheduling detail.

Limits

• You cannot buy a cheap answer. On a trivial classification the model still reasons. Routing small work here wastes money in a way that only shows up on the bill.
• Latency has a floor. Anything interactive is a product decision, not just an infrastructure one.
• Prompts that manufacture reasoning are counterproductive. "Think step by step" is redundant at best and duplicates a process already running at worst.
• Its knowledge cutoff is January 2026 — earlier than Opus 5's May 2026. The top tier is not the best-informed one. Give it a search tool rather than trusting parametric recall.
• Worst value on the rated board. 49.6 index points per $20.00/M blended is 2.5.

Price and access

$10 per million input tokens and $50 per million output — twice Opus 5 on both sides, blending to $20.00/M. Available on the Claude API, OpenRouter and OpenCode Zen. Anthropic has since listed Claude Fable 5.1 at the same $10/$50; it appears on this board unscored. Last scored 15 Sep 2026.

Alternatives on this board

• Claude Opus 5 — 92 overall at half the price, better at coding and terminal work, quicker. The default this model should be escalated from.
• GPT-5.6 Sol — 90 overall, Reasoning 95, at $2/$10. The direct competitor on hard reasoning, at a fifth of the blended cost.
• DeepSeek R2 — 83 overall but Reasoning 93, a deliberate chain-of-thought model worth testing before paying frontier rates for thinking.
• Claude Mythos 5 — the same specs and price without consumer safeguards, invitation-only for defensive security work.

Sources