feat: add Claude Fable 5.1 as a benchmark pricing model

[Mapleeeeeeeeeee]

Why

Claude Fable 5.1 shipped today (2026-09-02) with a different cache price structure: cache reads cost 2.5% of base input instead of the 10% every other tier uses. The benchmark's cost model needs its own pricing row to say anything about it.

Changes

• internal/benchmark/model.go: PricingFable (base input $10, 5-minute cache write $12.50, cache read $0.25 per MTok), TokenCountModelFable = "claude-fable-5-1", aliases fable / fable-5-1 in ResolveModel, error text lists them.
• cmd/cc-session/benchmark.go: -model help text.
• docs/benchmark.md: flag table, alias mapping, and a note that Fable 5.1 rows are not comparable to Opus/Sonnet rows because of the cache-read rate.
• internal/benchmark/model_test.go: table-driven ResolveModel coverage for every alias, a guard that Fable cache read stays at 2.5% of base input, and the unknown-model error text.

Prices quoted from https://platform.claude.com/docs/en/about-claude/pricing: "On Claude Fable 5.1 and Claude Mythos 5.1, a cache hit costs 2.5% of the standard input price ($0.25 USD per million tokens)."

Verification

• go build ./..., go vet ./..., go test ./... green.
• count_tokens accepts claude-fable-5-1 (HTTP 200) with this account, so the real model id is used for token counting.
• Benchmark run with -model fable on the same 20 sessions as the v0.1.77 Opus report (-n 20 -min-kb 100 -days 90 -overhead 40731), results below.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TcvWErnKnuADJsomnZfwFr

Benchmark: Fable 5.1 vs Opus 4.8 pricing, same 20 sessions

NewCtx differs by at most 2 tokens per session (shared tokenizer), so the rows differ only by price.

Median	Opus 4.8	Fable 5.1
Cold break-even	turn 1	turn 1
Cold 10-turn saving	64%	64%
Cold 100-turn saving	35%	34%
Warm break-even	turn 1	turn 5
Warm 10-turn saving	57%	35%
Warm 100-turn saving	35%	33%

Cold numbers are unchanged because they are dominated by cache writes, and the write multiplier (1.25× base input) is the same on both models. Warm numbers drop because re-reading the full uncompressed history costs 2.5% of base on Fable 5.1 instead of 10%, so the one-time write of the compressed context takes longer to pay back. Two sessions (9506f750, 960210d5) only break even at turn 10–11 under warm cache.