The launch copy says "20% cheaper," but the real question isn't the unit price — it's the cost per task. This article settles three accounts: the rate card, the per-task cost by effort tier, and a set of measured numbers that challenge the "cheaper" story. Prices checked Nov 12–15, 2026; the checkout page price is authoritative.
| Item (per 1M tokens) | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Input | $8 | $10 | $4 |
| Output | $40 | $50 | $20 |
| Fast mode (input / output) | $16 / $80 | — | — |
| Batch (input / output) | $4 / $20 | — | — |
| Cache reads | $0.20 | — | $0.20 (same as Fable) |
| Cache writes (5 min / 1 hour) | $10 / $16 | — | — |
$8/$40 is 20% below Fable 5.1's $10/$50 and twice Opus 5.5's $4/$20. OrcaRouter re-checked Anthropic's price list on Nov 12: no change, and not a limited-time introductory price — a rate card you can budget against. Fast mode (up to 2.5× faster generation) is available only on the Messages API, Claude API, Google Cloud, Microsoft Foundry (ID claude-fable-5-2-fast) and Claude Code — not on Bedrock. Cache reads at $0.20 are the one line identical to Opus, so long-context tasks with repeated prefixes may bill below the list price in practice.
Artificial Analysis' cost per index task (Nov 15, 2026 reading) against each tier's score. Scores are the AA Intelligence Index v5.0.1; costs are dollars per index task:
| Effort tier | AA Index | Cost per task | Cost per point | In one line |
|---|---|---|---|---|
| Low | 54 | $3.10 | $0.057 | Cheapest; enough for light tasks |
| Medium | 56 | $3.80 | $0.068 | Claude Code default; the orchestration starting point |
| High | 62 | $9.08 | $0.146 | The sweet spot: 3 points below Max at under 35% of its cost |
| Xhigh | 62 | $14.83 | $0.239 | Ties High at 62, $5.75 dearer; beats Max only on specific tasks |
| Max | 65 | $26.54 | $0.408 | 3 more points for $17.46 more; about 45.3 minutes per task |
This table is the launch week's most-shared framing: "3 more points for $12 more" — Max scores 3 points above Xhigh at about $12 more per task ($26.54 vs $14.83). Community write-ups rank High as the sweet spot and Max as the worst value. Time splits by tier too: Max about 45.3 minutes per task, Xhigh 27.7, High 16.1 (AA model page, Nov 15, 2026 reading, drifting).
The system card's §8.4 FrontierCode 2.0 tier sweep (reported by VentureBeat) is the core evidence that Max is not the best tier: Low 34.2% / $2.10, Medium 37.6% / $2.48, High 45.3% / $6.72, Xhigh 49.1% / $14.19, Max 45.7% / $44.70. Max costs 3.15× Xhigh and scores 3.4 points lower; 22 of Max's 660 runs timed out and were still scored — counted as zero, Max would be 44.4% (below Opus 5.5's 45.5%). High nearly ties Opus Max (45.3% vs 45.5%) at about half the cost per task ($6.72 vs $12.82).
ComputingForGeeks ran three DevOps tasks via OpenRouter (locked to Anthropic), scored on an Ubuntu 26.04 experimental VM: on medium, Fable 5.2 wrote twice the tokens of 5.1, and the three tasks cost $0.446 against 5.1's $0.285 — 56% more per task; on high, three tasks cost $0.838 against 5.1's $0.329; in the same run Opus 5.5 cost $0.208 and Sonnet 5.5 $0.047. Fast mode and security-refusal reruns were also tested. The per-token price fell, but the model got chattier, and the actual bill went up — the strongest counter-evidence in the pricing story. It's a three-task sample that needs more replication, but it points the same way as AA's "$26.54 per task, most expensive on the board."
Casagbic's SQL test is a second footnote to the same conclusion: across 1,280 queries, Fable 5.2 scored 320/320 on the low/medium tiers at $0.029 per query; Haiku 5.5 scored 319/320 at about 54× lower cost. On accuracy-ceiling tasks, Fable's premium is nearly pure waste.
First, the tier: start everyday work on medium or high, and reserve Max for tasks worth $26.54; xhigh ties high at 62, so only pay up on tasks where xhigh beats max, such as FrontierCode and ARC-AGI-2. Second, cache and Batch: cache reads at $0.20 and Batch at $4/$20 are the levers for long-context and deferrable bulk work. Third, split routing: the community consensus of "Fable plans, Opus executes" — Fable Medium orchestrates, Opus 5.5 executes, putting the expensive model on planning and the cheap one on execution.
Reference: ComputingForGeeks — Claude Fable 5.2 released: features, benchmarks (3-task DevOps hands-on test) (rewritten and organized by ModelBrief; facts per official sources and Artificial Analysis)