The Claude Fable family
This family has carried two active versions since 1 September 2026, Fable 5 and Fable 5.1, both priced at $10 in and $50 out, exactly double Opus 5, while both carry a one-million-token context window and a 128K-token output ceiling. The cost delta is not traceable to any row in the spec table.
- double the Opus 5 rate
- adaptive thinking permanently on
- the most expensive tier in the Claude architecture
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
Where the spec delta actually sits
No row in the spec table accounts for the price gap, and that is the load-bearing fact of this page. Line up Fable 5 against Opus 5 on context window, output ceiling, deployment routes, and even knowledge cutoff: both carry a one-million-token window, both cap output at 128K tokens, and both ship on the same five deployment routes, Anthropic direct API, Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Fable 5 knowledge cutoff is January 2026 against May 2026 for Opus 5, so on that one parameter the costlier model is the older build.
The real difference is logged on two other parameters. Anthropic describes Fable 5 as the most capable model it has widely released, positioned for next-generation intelligence for long-running agents. Its adaptive-thinking field reads yes, always on, where Opus 5 and Sonnet 5 read simply yes, meaning the parameter cannot be disabled on Fable.
Two benchmark data points pointing in opposite directions
The operational recommendation on this page comes straight from this data. On the Arena WebDev leaderboard, Fable 5 scores 1628, below Opus 5 1692. On the general-text leaderboard the direction reverses: Fable 5 sits at 1506, above Opus 5 1494.
The gap runs 64 points on the coding workload and 12 points on the text workload. Most capable does not translate to highest quality across every workload. If your workload is coding, the model charging double sits lower on that exact benchmark. Our coding-category table shows the same pattern: Fable 5 ranks second, Opus 5 ranks first, a 15.4-point gap.
Two operational parameters you lose at the higher rate
First, response latency. In Anthropic table, the comparative-latency field reads slower for Fable 5, moderate for Opus 5, and fast for Sonnet 5. Before Fable 5.1 shipped, Fable 5 was the only model in that table carrying the slower tag; now both members of this family carry it. In architectures where an agent runs unattended overnight, this parameter is irrelevant. On any path where a user waits on the response, it is a direct cost.
Second, long-batch output capacity. Anthropic runs a beta on the Message Batches API that lifts the output ceiling to 300K tokens, and the list of supported models is explicit: Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 5, and Sonnet 4.6. Fable 5 is absent from that list. On the workload of producing very long output in batches, the cheaper model has capacity the costlier one lacks.
A thinking parameter this architecture will not let you disable
Adaptive thinking running permanently on carries a direct financial consequence invisible in the unit rate: thinking tokens bill as output tokens, and output here costs $50. On Opus 5 the effort level can be lowered to control cost; this family does not expose that lever.
One more parameter moves the invoice, and it is not specific to Fable but belongs to the whole tokenizer generation: the tokenizer introduced with Opus 4.7 turns the same input text into roughly thirty percent more tokens. Fable 5 runs the same tokenizer. So comparing the nominal rate of two models from different generations understates the actual bill.
The second version, and two that sit outside the public deployment tier
Claude Fable 5.1 shipped on 1 September 2026, bringing this family to two active versions. Its base rate matches Fable 5, and the context window and output ceiling are unchanged. Three parameters moved: cache-read cost dropped from $1 to $0.25, the knowledge cutoff moved from January to June 2026, and benchmark scores rose on long agentic workloads. Full arithmetic and its boundaries are documented on the dedicated Fable 5.1 page.
Two further versions exist in this family architecture but sit outside the public deployment tier. Claude Mythos 5 carries the same spec sheet and rate as Fable 5 and shipped on the same date, 9 June 2026; Mythos 5.1 repeated that pattern on Fable 5.1 release day. Neither is generally available: deployment is invitation-only, inside Project Glasswing, with no self-serve enrollment path. We do not count them as procurable versions of this family, because a build outside the public deployment tier has no place on a procurement reference page.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
Release timeline
Each version with its own release date, and what changed against the one before it.
- Claude Fable 5.1 current Base rate unchanged, cache rate at a quarter, and the knowledge cutoff now sits ahead of Opus 5.
- Claude Fable 5 legacy The highest-cost model in general circulation on Anthropic's production line: rank one in general conversation, rank two on coding workloads.
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- blocked
- Payment
- no working route
- Free tier
- no
- Measured on
How we checked: Iran does not appear on Anthropic's supported-countries list. This status is sourced from Anthropic's own documentation, not a direct network measurement.
Documented strengths
- The highest Claude benchmark score on the Arena general-text leaderboard, 1506 against 1494 for Opus 5
- A one-million-token context window and a 128K-token output ceiling, matching the top tier of the Claude architecture
- Full coverage across all five deployment routes, from Anthropic direct API to Microsoft Foundry
- Adaptive thinking runs permanently on, so behaviour is more uniform across consecutive requests
Known weaknesses and limits
- Its rate is double Opus 5 while the context window, output ceiling, and deployment routes are fully identical.
- It sits below Opus 5 on the WebDev leaderboard, 1628 against 1692, meaning the costlier model performs worse on the coding workload.
- It is the only model whose comparative-latency field reads slower in Anthropic table, true for this version and for the newer one alike.
- The 300K-token output beta on Message Batches does not cover Fable 5, while it covers Opus 5 and four other models.
- Adaptive thinking is always on, so the effort-tuning lever for reducing cost is not available.
- Its knowledge cutoff is January 2026, four months behind the May 2026 cutoff of Opus 5.
- Iran is not on the Anthropic supported-countries list, so no official deployment route exists.
Technical verdict
If your workload is coding, Opus 5 is the correct choice: higher benchmark score, half the rate. Justifying Fable 5 only holds when the workload matches exactly what it was built for, long-running agents and general text, and when neither response latency nor long-batch output capacity is an operational constraint.
Frequently raised questions, with documented answers
Does Fable 5 outperform Opus 5
It depends on the workload, and the two Arena data points point in opposite directions. Opus 5 leads on WebDev at 1692, while Fable 5 leads on general text at 1506. Given Fable rate is double, the clear technical choice for a coding workload is Opus 5.
How many active versions does the Fable family have
Two: Fable 5, generally available since 9 June 2026, and Fable 5.1, released 1 September 2026 with a lower cache cost and a more recent knowledge cutoff. Claude Mythos 5 and Mythos 5.1 carry the same specs and rate but ship invitation-only inside Project Glasswing with no self-serve enrollment, so they are not procurable versions of this family.