Claude Fable 5 Pricing & Limits: 2026 Cost Breakdown

Last updated: June 2026 · By Ignacy Kwiecień, founder & editor-in-chief, DecodeTheFuture.org

How much does Claude Fable 5 cost per token in 2026?

Claude Fable 5 costs $10 per million input tokens (MTok) and $50 per million output tokens on the standard API, according to Anthropic’s official pricing documentation as of June 23, 2026. That is the base rate before any caching or batch discount, and it applies across the full 1M-token context window with no long-context surcharge. Output is metered at 5× the input rate, which is the usual Claude shape but at a higher absolute level than the rest of the lineup.

Two things make this number worth a dedicated breakdown rather than a one-liner. First, the headline $10/$50 is the least interesting row — the caching and batch rows below change a real production bill far more. Second, the plan-side mechanics shifted on June 23, 2026: Fable 5 stopped being a free inclusion on Pro, Max and Team and now meters against usage credits. This page is strictly about price and limits. For why Fable 5 and Mythos 5 access changed, see our separate write-up on the Fable 5 and Mythos 5 availability change; for where Fable 5 sits among Anthropic’s models, see the Claude models map for 2026.

What is the full Claude Fable 5 API rate card?

The full Fable 5 rate card has six rows that most pricing summaries skip. Competitors quote $10/$50 and stop; the rows that actually decide your bill are the cache writes, the cache hit, and the batch discount. Here is the canonical table, sourced to Anthropic’s platform pricing docs.

Note that Mythos 5 shares this identical rate card — $10/$50 base, same cache and batch rows — but Mythos 5 is limited-availability (early-access / Glasswing) and is not generally purchasable on the standard API the way Fable 5 is. Do not assume you can simply call Mythos 5 at these rates today; the rate card matches, the access does not.

How do prompt caching and batch discounts change the real bill?

Prompt caching and batch processing are the two levers that move a Fable 5 bill the most — caching can cut repeated-context input cost by 90%, and batch halves everything for work you don’t need in real time. The trick is knowing when each one actually pays for itself, because both caching writes cost more than a plain input token up front.

When does prompt caching pay off on Fable 5?

Caching pays off after a single read with the 5-minute cache, or after two reads with the 1-hour cache. The math is direct: a 5-minute cache write costs 1.25× base input ($12.50/MTok), and every subsequent hit costs 0.1× ($1/MTok). So caching a 1M-token prefix once and reading it once costs $12.50 + $1.00 = $13.50, versus $20.00 for two uncached passes ($10 each) — already cheaper on the first reuse. The 1-hour cache writes at 2× ($20/MTok), so it breaks even on the second read and wins clearly from the third onward. For long system prompts, large document context, or agent loops that re-send the same tool definitions every turn, caching is close to mandatory on a $10-input model.

How much does the Batch API save?

The Batch API applies a flat 50% discount in both directions, taking Fable 5 to $5/MTok input and $25/MTok output. There is no caching-versus-batch trade-off to overthink at the rate-card level — batch is for asynchronous work (evals, bulk classification, offline summarization, content pipelines) where you can tolerate the queued turnaround. If your workload is offline and high-volume, batch is the single biggest lever, bigger than caching, because it discounts output too, and output is where Fable 5’s cost concentrates at 5× input.

What does a real Claude Fable 5 bill look like?

A concrete job — 1M input tokens plus 200k output tokens — costs about $20 uncached, drops to roughly $11 if most of the input is a cached prefix you reuse, and falls to $10 on the Batch API. Worked numbers make the levers obvious. The example assumes a single request shape; your real distribution of cached vs fresh input will differ.

The honest read: output dominates Fable 5 economics. In the uncached case, 200k output tokens cost the same as 1M input tokens. Caching attacks input, so it only halves a bill where input is large and reused; if your workload is output-heavy (long generations, verbose agents), the Batch API — which discounts output too — or downtiering to Opus 4.8 saves more. There is one more wrinkle that quietly inflates every Fable 5 estimate, covered next.

Is Claude Fable 5 cheaper than Opus 4.8? (side-by-side cost)

No — Fable 5 is priced at exactly double Opus 4.8 on both input and output, with no exception in the cache or batch rows. Opus 4.8 is the cheaper fallback for the same Anthropic stack, and on identical token volumes it costs half as much across the board. This is the comparison most pricing posts skip, so here it is dollar-for-dollar.

The decision is clean precisely because the premium is uniform: there is no token-mix where Fable 5 becomes relatively cheaper. You are paying a flat 2× for whatever capability advantage Fable 5 gives you on your specific task. Anthropic positions Fable 5 as its most powerful generally available model (a vendor/press framing, not an audited fact — attribute it to Anthropic and outlets like VentureBeat). The practical rule: if Opus 4.8 clears your quality bar on a task, it halves the bill. Reserve Fable 5 for work where the extra capability measurably changes the outcome, and route the rest to Opus 4.8 or a cheaper provider entirely — see our best inference APIs comparison for 2026 for non-Anthropic options.

How does Fable 5 draw down a subscription plan now?

As of June 23, 2026, Fable 5 is no longer included in Pro, Max, Team and seat-based Enterprise plans; continued use draws from usage credits billed at standard API rates ($10/$50). During the free window (June 9–22, 2026) Fable 5 was bundled into those plans at no extra cost. Anthropic frames the June 23 removal as a capacity decision, not a permanent new pricing tier, and has stated intent to restore Fable 5 to standard plans when capacity allows — with no committed date.

Below is the decision tree for what happens when you select Fable 5 on a Claude.ai plan today. The verified parts (the plan window, the API-rate metering, the Opus 4.8 fallback) are labeled as such; the burn-multiplier and credit mechanics are reported from third-party developer guides citing Claude’s in-app messaging, not published on Anthropic’s public docs as of June 23, 2026.

One more reported figure circulates: a daily usage-credit redemption ceiling said to sit around $2,000/day. It is single-sourced and we treat it as uncertain — mentioned here only so you recognize it, not as a confirmed limit. The redemption mechanics also reportedly span chat, Claude Code, Research mode and project files, so credit burn is account-wide, not per-surface.

Is Fable 5 worth it on a plan versus the API?

For most subscribers, the answer post-June-23 is: use the plan for everyday Claude and reach for Fable 5 selectively, because it now burns credits at API rates instead of being free. Anthropic’s plan list prices give the context: Pro is $20/month (or $200/year), Max starts at $100/month, Team is roughly $20–25 per seat per month, and Enterprise is a seat price plus usage at API rates. During the free window, Fable 5 on a $20 Pro plan was an obvious yes. After June 23, the reported ~2× burn rate means Fable 5 chews through a plan’s bundled allowance about twice as fast as Opus 4.8, and overflow is metered at $10/$50.

The practical framing without retelling any availability drama: if your Fable 5 use is occasional and high-value, a plan plus credits is fine. If you run Fable 5 heavily and predictably, model your monthly token volume at the API rate card above and compare it to staying on Opus 4.8 at half the cost. For agent and Claude Code workloads specifically, where token volume balloons with tool loops and long context, the caching discipline from the rate-card section matters more than the plan choice — and if the cost pushes you to weigh alternatives, see our 2026 comparison of AI coding assistants.

What are the API-side rate limits, separate from plan allowances?

API rate limits are a different ceiling from subscription plan allowances and are organized into usage tiers (Tier 1–4 plus Enterprise) that raise your requests-per-minute and tokens-per-minute as your spend history grows. This is the limit that bites production developers, not Claude.ai subscribers: it caps throughput, not monthly dollars. The structure mirrors what we documented for other inference providers in the NVIDIA NIM pricing and limits guide — rate limits, not just per-token price, often decide whether a workflow is actually shippable.

Because tier ceilings depend on your account history and can change, treat them as account-specific and confirm yours in the console rather than hard-coding a number. The engineering takeaway is the same as for any high-volume API: design for backoff, cap concurrency, and keep a cheaper fallback model (Opus 4.8) wired in so a throughput cap doesn’t stall a run mid-workflow.

FAQ