Fable 5 vs Sonnet 5 — token cost (Njenga)

A member-only Medium post by Joe Njenga (July 2026) comparing what Claude Fable 5 and Claude Sonnet 5 cost to run inside Claude Code, framed as a lesson in "how to stop burning tokens." Only the baseline test is readable without a Medium subscription; the coding comparison that follows is behind the paywall.

The baseline anecdote

Njenga typed a single "Hello" into Claude Code on each model and read off the cost stats the CLI reports. His numbers for that one prompt:

His read: Sonnet 5 pulled more from cache (24.8k vs 15.6k), and since cache reads are billed below fresh input, the model that caches better ends up cheaper — even while generating more output. He calls the gap a "3.2× pricing difference" on a single-word prompt and says it widened his curiosity about how the two scale on real coding tasks (that section is paywalled).

Why "Hello" costs 47 cents

The eye-catching $0.4745 is not the price of one word. In Claude Code every turn ships the whole system prompt, tool definitions, and accumulated session context, so a trivial "Hello" still carries ~13–25k tokens of input + cache. The post's own token counts make that plain; the headline is the harness overhead, not the greeting.

How much to trust this

T4, and treat the numbers as one anecdote. It's a paywalled personal blog, a single run per model, no repetition or averaging, and the striking claim — that Fable 5, the Claude 5-family model built for fast Claude Code work, is ~3× more expensive than Sonnet 5 on the same prompt — runs against the intuition that a lightweight model should be the cheap option. It may be real (caching behaviour and per-token rates differ by model), but nothing here is authoritative pricing. The open question is flagged in synthesis; it wants a T1 Anthropic price sheet to settle. See llm-api-pricing for the caching lever the post leans on.

Cross-spoke context

The setting is Claude Code (an agentic coding harness → agentic-tooling-wiki) and the mechanism is prompt caching (an inference-serving detail → llm-inference-wiki, kv-cache). The in-scope substance here is the model cost comparison, so it lives in this spoke; those two facets are noted, not re-paged.

Re-tiered T4 → T2 (curator's call, 2026-08-30)

The tier now grades the evidence, and there is some. Njenga ran both models on the same prompt and published what came back: input tokens, cache reads, output tokens, wall-clock, cost. The 2026-08-04 rule sends "someone tested, measured, compared or used the thing" to T2, and the 2026-08-30 redefinition made T4 mean unevidenced — blogspam and the unevidenced practitioner argument being its two kinds. This page is neither. Weak evidence and absent evidence are different things, and the ladder now has a tier that means the second.

The corpus had already made this error in the opposite direction: on 2026-08-25 three MachineLearningMastery walk-throughs were graded as blogspam, and one of them held the only third-party timing of continuous batching in the whole corpus.

T2 does not mean trust the number. Everything in ## How much to trust this above stands and is the reason this is T2 rather than T1: one run per model, no repetition, no averaging, a one-word prompt whose cost is dominated by harness overhead rather than by the model, and the coding comparison that would actually matter sitting behind a paywall. Read it as a single observation from outside Anthropic, which is what it is and what the corpus was short of.

Related

claude-fable-5 · claude-sonnet-5 · llm-api-pricing · joe-njenga · kv-cache · synthesis