Fable 5's Included Window Isn't Free, and the Meter Starts June 23

TL;DR

• Included isn’t free. Fable 5 is bundled with Claude Pro/Max plans through June 22, but it draws down your normal plan quota at roughly double the Opus rate. The multiplier appears in a Claude Code UI string and nowhere in Anthropic’s public web docs.
• The real price is exactly 2x Opus 4.8 ($10/$50 vs $5/$25 per million tokens, same tokenizer). The “~2.7x” figure circulating is only valid against pre-4.7 Opus budgets, where a tokenizer change adds up to 35% more tokens on the same text.
• June 23 is the cliff: Anthropic removes Fable 5 from subscription plans and bills it through usage credits at standard API rates. No subscriber discount has been published.
• The #1 benchmark ranking has an asterisk: Artificial Analysis measured Fable 5 handing 8 to 9% of hard benchmark tasks to Opus 4.8 via silent safety fallback.

I turned Fable 5 on the day it shipped and watched my Claude Max usage bar move about twice as fast as it does on Opus. Nothing in the launch post prepared me for that. The post says the new model is “included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost,” and that sentence is true.¹ It is also close to useless as a description of what actually happens to your plan when you use the thing. The window is real, the model is real, and so is a meter that Anthropic never quite shows you. On June 23 that meter converts to money.

This article does the billing math the launch copy skips: what Fable 5 actually costs against the model you were running last week, what “included” does to a plan quota, what changes on June 23, and what the benchmark crown sitting on top of all this is actually made of.

Background

Anthropic shipped Claude Fable 5 on June 9 as a tier above Opus 4.8, not a replacement for it. The launch post positions it for “the most demanding reasoning and long-horizon agentic work,” and it shares its weights with Mythos 5, a classifier-free variant restricted to government use.¹ Opus 4.8 stays in the lineup at its current price.

The launch mechanics are the story. Three sentences from the announcement set up everything that follows:

“From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.” “On June 23, we’ll remove Fable 5 from those plans. Using it after that will require usage credits.” “If capacity allows, we’ll extend the included window.”¹

So: thirteen included days, then a hard switch to metered billing. That pattern should sound familiar. GitHub moved Copilot to metered AI credits this month, and I covered what usage-based billing does to developer behavior when the meter is opaque. Fable 5 runs the same play with a sharper deadline, and GitHub itself is a data point here: Copilot offers Fable 5 GA, but bills it “at provider list pricing under Usage Based Billing,” outside the premium-request allowance entirely.² Even the company that invented the flat-rate AI subscription won’t flat-rate this model.

One more piece of context matters before the math. Subscription quota on Claude plans is not a token meter you can read. It’s a usage bar with weekly and five-hour windows, weighted differently per model, and the weights aren’t published. Keep that in mind; it’s where the most interesting finding lives.

Methodology

Every pricing, date, and multiplier claim in this article was verified against a primary source on June 11, 2026: Anthropic’s launch post, the platform pricing docs, the model migration guide, the rate-limits docs, and Anthropic’s help center, plus Artificial Analysis’s published evaluation notes. Community numbers come from named, linkable primaries (Simon Willison’s blog, specific Hacker News comments) and are labeled as anecdotes, not measurements. Where a claim exists only in third-party writeups or in product UI strings rather than official documentation, I say so explicitly, because that gap turned out to be one of the findings. All 33 source URLs returned live pages at time of writing.

The worked cost example uses the session shape Anthropic’s own pricing page uses (50,000 input tokens, 15,000 output) so anyone can recompute it from the public rate table.

Findings

The price is exactly 2x Opus 4.8, and the “2.7x” number making the rounds uses the wrong baseline

The rate table is clean. Fable 5 costs $10 per million input tokens and $50 per million output tokens. Opus 4.8 costs $5 and $25. The 2x ratio holds across every billing variant: cache writes ($12.50 vs $6.25), cache reads ($1 vs $0.50), and batch ($5/$25 vs $2.50/$12.50).³

Run the standard session through it: 50k input plus 15k output costs $0.625 on Opus 4.8 and $1.25 on Fable 5. Cached and batched variants scale identically. Two times, everywhere, full stop.

So where does the higher figure circulating in community posts come from? A tokenizer change, and it’s being counted against the wrong model. Anthropic’s pricing docs note that Opus 4.7 and later use a new tokenizer that “may use up to 35% more tokens for the same fixed text” compared to earlier models.³ The migration guide is explicit that Fable 5 and Opus 4.8 sit on the same side of that change: “Token counts are roughly unchanged because both models use the same tokenizer.”⁴

That means the multiplier depends entirely on what you’re comparing against:

Baseline	Price ratio	Token inflation	Effective cost multiple
Opus 4.8 (current)	2.0x	none (same tokenizer)	2.0x
Opus 4.6-era budget (pre-tokenizer-change)	2.0x	+30 to 35%	~2.6 to 2.7x

If your cost dashboard was calibrated in the Opus 4.5/4.6 era and never re-baselined, the same fixed workload on Fable 5 lands at roughly 2.6 to 2.7 times your old spend. Against the Opus 4.8 you ran last week, it’s exactly 2x. Writing “effectively 2.7x over Opus” without naming the baseline double-counts a tokenizer change that Opus 4.8 already absorbed. I haven’t found a single community source that derives this correctly, so consider this the corrected math.

Two pricing symmetries fall out of the table and are worth knowing. Fable 5 on the batch API costs exactly what Opus 4.8 costs on the standard API. And Opus 4.8’s fast mode is priced at $10/$50, identical to Fable 5 standard, while fast mode is not available on Fable 5 at all.³ Anthropic’s price sheet quietly tells you what it thinks the premium tier is worth: a doubling, whichever axis you take it on.

”Included” means it eats your plan quota at double weight, and the multiplier lives in a UI string

Here’s the part the launch post doesn’t address at all. Included models draw on your plan’s usage limits, and the launch language (“at no extra cost”) never says what the draw rate is. I noticed my Max usage bar emptying at roughly twice the Opus rate during normal work. It turns out the closest thing to an official statement of that weighting is an interface string in Claude Code:

“Uses your limits ~2x faster than Opus.”

That string, surfaced in product UI rather than documentation, is the multiplier.⁵ No page on Anthropic’s public web documents it. The pricing docs don’t mention plan weighting. The launch post doesn’t either. The only official web text confirming Fable 5 touches plan limits at all sits in a help-center article about a different feature, model fallback, which notes in passing that a fallback response “counts toward your usage limit or consumption.”⁶

Community experience says the lived burn can run steeper than the nominal weighting, because Fable 5 ships with adaptive thinking always on and agentic workflows multiply token volume through subagents. One Hacker News user on the $100 Max tier reported burning an entire five-hour usage window in under eight minutes of max-effort agentic use.⁷ Another reported spending 45% of a $200/month plan’s weekly quota in a single day of code review, against roughly 80% per week doing the same work on Opus 4.8.⁸ These are anecdotes, not controlled measurements; the thinking-token volume and the quota weighting are confounded in both. But the direction is consistent across every report I could verify.

Simon Willison published the cleanest primary numbers from day one: $110.42 of equivalent API usage in a single day, $99.26 of it from one 78.2-million-token session. His verdict: “something of a beast. It’s slow, expensive.”⁹

June 23: the meter converts to money, with no subscriber discount published

When the included window closes, Fable 5 doesn’t move to a higher subscription tier. It leaves subscriptions entirely. Continued use requires usage credits billed at standard API rates, per the launch post.¹ The general usage-credits help article documents the mechanics that exist today: prepaid credits, a $2,000 per day redemption cap, billing at API list price. It contains no Fable-specific section.¹⁰ More detailed descriptions of the post-cliff flow (credits activating after plan usage exhausts, mobile subscribers needing to enable on web) circulate in third-party guides, and I could not verify them against any official page; treat those details as unconfirmed until Anthropic documents them.¹¹

What I can state from primary sources: no included credit allowance for subscribers has been announced, no discounted subscriber rate has been announced, and the only softening language is the launch post’s “if capacity allows, we’ll extend the included window.” On the trust-the-hedge question, it’s worth knowing that Anthropic has already walked back one launch-week claim, quietly relaxing the recursive-self-improvement suppression language it shipped with.¹² The operative planning date is June 23.

Scale Willison’s day-one numbers against the cliff and the shape gets vivid. A heavy agentic day that cost $110 in notional API value during the included window becomes $110 of actual credits afterward. A Max 5x subscription runs $100 a month. One enthusiastic day on the wrong side of June 23 can exceed a month of plan cost. That is not an argument that the price is wrong. It’s an argument that the included window is teaching usage patterns that the post-cliff economics won’t support, and the people most likely to get caught are the ones who never saw the meter running.

The #1 ranking was earned by a composite: 8 to 9% of hard tasks were answered by Opus 4.8

Fable 5 debuted at the top of Artificial Analysis’s Intelligence Index at 64.9, about five points clear of GPT-5.5. The asterisk is in AA’s own naming of what they tested: “Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback).” Per AA’s published notes, the model’s safety classifiers refused and fell back to Opus 4.8 on roughly 8% of Index tasks, 9% of Humanity’s Last Exam, and 9% of AA-Omniscience, and the HLE run cost about $2,200, “the highest of any model we have evaluated.”¹³

The mechanics make the substitution easy to miss. A classifier refusal comes back as HTTP 200 with stop_reason: "refusal" and a category field, not as an error, and an opt-in beta fallbacks parameter retries the request server-side on another model.¹⁴ Configured for fallback, a Fable 5 integration silently becomes a Fable-plus-Opus blend, with no signal in the response stream a casual consumer would notice. Anthropic, for its part, says classifier refusals affect “fewer than 5% of sessions on average” in real use.⁶ The gap between that figure and AA’s 8 to 9% on hard benchmark tasks is unexplained; harder prompts plausibly trip classifiers more often, but nobody has published the breakdown.

I can’t find any precedent for scoring this. Benchmark methodology debates cover contamination, selective disclosure, and private variants (The Leaderboard Illusion preprint catalogued several for Chatbot Arena, though its specific statistics are disputed by LMArena).¹⁵ Silently serving refused tasks with a different model is a new entry in that genre, and as of this writing there is no literature on how a leaderboard should handle it. The honest summary: Fable 5 is very likely the strongest public model right now, and its headline score is partly a measurement of a two-model system.

Discussion

The strongest counterargument to everything above comes from Willison, in a Hacker News thread full of quota complaints: the included window is the subsidy, not the trap. Fable 5 is “underpriced right now,” and the real sticker shock arrives when everyone pays full API rates.⁸ On this reading, complaining about quota burn during a free window gets the direction of the cliff backwards, and GitHub’s refusal to flat-rate the model is evidence that frontier inference at this tier genuinely costs what Anthropic is charging for it. I think that reading is mostly right, which is why this article’s complaint is aimed at the disclosure, not the price.

The economics literature predicts both halves of the reaction. Flat-rate bias research going back to Lambrecht and Skiera’s 2006 work shows consumers systematically overpay for flat plans to avoid “taxi-meter” anxiety, which is exactly the anxiety a hidden quota multiplier plus a hard metering date produces.¹⁶ Later work documents a countervailing flexibility effect: a real segment rationally prefers metered pricing, and some of Fable 5’s eventual credit-paying users will be in it, happily paying $1.25 per heavy session for the best model available. Research disagrees on how large each segment is; the existence of both is not in dispute.

Two practical traps deserve flagging before the cliff. First, local cost-tracking tools haven’t caught up: ccusage, the most popular Claude Code spend tracker, had no claude-fable-5 pricing entry as of June 11, and at least one cost-tracking integration falls through to Sonnet-tier defaults, under-costing Fable sessions by roughly 3.3x.¹⁷ Dashboards will be wrong in the optimistic direction during the exact window when subscribers are deciding whether the model is worth metered money. Second, the capability premium is contested at the margin: several practitioners report Fable 5 as only incrementally better than Opus 4.8 on their workloads,⁸ and at 2x price and roughly 2x quota weight, incremental wins don’t clear the bar for routine work. The boring conclusion survives: route the expensive model to the tasks that earn it.

Limitations of this analysis: the ~2x quota weighting is corroborated by a UI string and multiple independent community reports, not by documentation, and could change without notice. The post-cliff credit mechanics are partly third-party-sourced. And no controlled same-task token-count comparison between Fable 5 and Opus 4.8 has been published by anyone; the “burn is steeper than 2x” reports conflate the weighting with thinking-token volume. That measurement is the obvious follow-up experiment, and I intend to run it.

Conclusion

• Fable 5 costs exactly 2x Opus 4.8 across every billing variant. The ~2.7x figure is only correct against pre-4.7 Opus budgets, where the tokenizer change adds up to 35% token inflation. Name your baseline before you quote a multiplier.
• “Included at no extra cost” means metered against your plan quota at roughly double Opus weight, per a Claude Code UI string that no public Anthropic doc backs up. The opacity, not the weighting, is the problem.
• June 23 converts the meter to money: usage credits at standard API rates, no published subscriber discount, $2,000/day redemption cap. Plan for the date, not the “if capacity allows” hedge.
• The #1 benchmark ranking measures a composite system: 8 to 9% of hard tasks were answered by Opus 4.8 through silent classifier fallback. Probably still the strongest model; not a clean single-model score.
• Before June 23: check whether your cost tracker prices Fable 5 at all, and decide which tasks actually earn a 2x model. The included window is free habit formation for a metered product.

Open questions:

• Will Anthropic publish the plan-quota weighting, or extend the included window? Both would change the calculus; neither has happened as of June 11.
• Is the observed burn genuinely steeper than 2x, or is that thinking-token volume? Only a controlled same-task measurement settles it.
• Why do hard benchmark tasks trip the safety fallback at nearly double Anthropic’s claimed real-world session rate?

Sources

1. Anthropic. “Claude Fable 5 and Claude Mythos 5.” June 9, 2026. https://www.anthropic.com/news/claude-fable-5-mythos-5
2. GitHub Changelog. “Claude Fable 5 is generally available for GitHub Copilot.” June 9, 2026. https://github.blog/changelog/2026-06-09-claude-fable-5-is-generally-available-for-github-copilot/
3. Anthropic. “Pricing.” Claude Platform docs. https://platform.claude.com/docs/en/about-claude/pricing
4. Anthropic. “Migration guide.” Claude Platform docs. https://platform.claude.com/docs/en/about-claude/models/migration-guide
5. Developers Digest. “How Claude’s Usage Limits Actually Work With Fable 5.” June 2026. https://www.developersdigest.tech/blog/claude-usage-limits-fable-5-explained (secondary source for the Claude Code UI string; corroborated by HN reports)
6. Anthropic Help Center. “Why Claude switched models in your conversation with Fable 5.” https://support.claude.com/en/articles/15363606
7. Hacker News thread on Fable 5 subscription plans, June 2026. https://news.ycombinator.com/item?id=48463982
8. Hacker News, “Tell HN: Fable is too expensive.” June 2026. https://news.ycombinator.com/item?id=48485950
9. Willison, Simon. “Initial impressions of Claude Fable 5.” June 9, 2026. https://simonwillison.net/2026/Jun/9/claude-fable-5/
10. Anthropic Help Center. “Manage usage credits for paid Claude plans.” https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plans
11. claudefa.st. “Fable 5 Usage Credits Explained.” June 2026. https://claudefa.st/blog/guide/development/fable-5-usage-credits (third-party; post-cliff flow details not officially documented)
12. Latent.Space AINews. “Mythos but Safe, with Controversial Terms.” June 2026. https://www.latent.space/p/ainews-anthropic-claude-fable-5-mythos
13. Artificial Analysis. “Claude Fable 5 launches at #1 on the Intelligence Index.” June 2026. https://artificialanalysis.ai/articles/claude-fable-5-mythos-intelligence-index
14. Anthropic. “Models overview” (refusal stop_reason and fallbacks beta). Claude Platform docs. https://platform.claude.com/docs/en/about-claude/models/overview
15. Singh, S., et al. “The Leaderboard Illusion.” Preprint, not peer-reviewed; statistics disputed by LMArena. https://arxiv.org/abs/2504.20879
16. Lambrecht, A., and Skiera, B. “Paying Too Much and Being Happy About It: Existence, Causes, and Consequences of Tariff-Choice Biases.” Journal of Marketing Research, 2006. https://www.marketing.uni-frankfurt.de/fileadmin/Publikationen/Lambrecht_Skiera_Tariff-Choice-Biases-JMR.pdf
17. ccusage releases (no claude-fable-5 pricing entry as of v20.0.11, June 11, 2026). https://github.com/ryoppippi/ccusage/releases ; cost-tracker fallthrough example: https://github.com/danielmiessler/Personal_AI_Infrastructure/issues/1341

Footnotes

1. Anthropic launch post, June 9, 2026. All included-window quotes verbatim. anthropic.com ↩ ↩² ↩³ ↩⁴

2. GitHub Changelog, June 9, 2026: “This model is billed at provider list pricing under Usage Based Billing.” github.blog ↩

3. Claude Platform pricing docs: full rate table; tokenizer note (“may use up to 35% more tokens for the same fixed text,” applying to Opus 4.7 and later); fast-mode availability. platform.claude.com ↩ ↩² ↩³

4. Claude Platform migration guide: “Token counts are roughly unchanged because both models use the same tokenizer.” platform.claude.com ↩

5. The “~2x faster than Opus” weighting appears in the Claude Code interface and in secondary writeups; no official Anthropic web page documents it. Secondary: developersdigest.tech. Corroborating community report: HN user cortesoft, thread 48463982. ↩

6. Anthropic Help Center article 15363606: fallback responses “count toward your usage limit or consumption”; “fewer than 5% of sessions on average.” support.claude.com ↩ ↩²

7. HN user joshstrange (Max $100 tier): five-hour window consumed in under eight minutes of max-effort agentic use. Anecdote, not a controlled measurement. news.ycombinator.com ↩

8. HN thread 48485950: thread OP’s 45%-of-plan-in-one-day report; simonw’s “underpriced right now” counterpoint; practitioner reports of incremental gains over Opus 4.8. news.ycombinator.com ↩ ↩² ↩³

9. Willison, June 9, 2026: $110.42 in one day, $99.26 from a single 78.2M-token session. Primary source. simonwillison.net ↩

10. Anthropic Help Center article 12429409: general usage-credits mechanics, $2,000/day redemption cap, standard API rates. No Fable-specific section. support.claude.com ↩

11. claudefa.st credits guide: post-cliff activation flow details. Third-party; not verified against official docs. claudefa.st ↩

12. Latent.Space AINews roundup: RSI-suppression language walkback. latent.space ↩

13. Artificial Analysis: Index 64.9; eval entity “Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)”; ~8% Index / 9% HLE / 9% AA-Omniscience fallback rates; ~$2.2k HLE run. artificialanalysis.ai ↩

14. Claude Platform models overview: HTTP 200 refusals with stop_reason: "refusal" and stop_details.category; beta fallbacks parameter. platform.claude.com ↩

15. Singh et al., “The Leaderboard Illusion,” arXiv:2504.20879. Preprint; quantitative claims disputed by LMArena. Used here only for the qualitative point that leaderboard methodology hides in headline numbers. arxiv.org ↩

16. Lambrecht and Skiera, JMR 2006: flat-rate bias and the taxi-meter effect. Counterpoint: Krämer and Wiewiorra (Telecommunications Policy, 2012) document a flexibility effect; segment sizes are disputed. PDF ↩

17. ccusage v20.0.2 through v20.0.11 ship no claude-fable-5 pricing entry; danielmiessler/Personal_AI_Infrastructure issue #1341 documents a ~3.3x under-costing fallthrough to Sonnet defaults. github.com ↩