Model
Model explorer

Claude Fable 5

CLOSED
Anthropic · Claude 5 family · released Jun 9, 2026

Model id claude-fable-5. This is the correct current name of the family the task flagged as 'a newer Claude model family' — it is Anthropic's first widely-released 'Mythos-class' model (a tier above Opus), positioned above Opus 4.8. It shares identical underlying weights with the invitation-only Claude Mythos 5 (claude-mythos-5, Project Glasswing only) but adds safety classifiers that can refuse requests; Mythos 5 does not have those classifiers. Notably, access to Fable 5 was briefly restricted/pulled and then restored in July 2026 following a jailbreak-linked export-control issue (see anthropic.com/news/redeploying-fable-5) — this does not change its specs. NO SEPARATE '(High)' PRODUCT: like Sonnet 5, 'high' is the documented default effort level and is not a distinct marketed SKU; not added as a sibling entry. Claude Mythos 5 itself was NOT added as a separate corpus entry given its restricted, non-public availability and the difficulty independently sourcing benchmark numbers specific to it (as opposed to Fable 5).

ReasoningCodingVisionFunction callingTool useAgentic
2928.6
Elo · rank #11
Parameters
Undisclosed
Active params
Context
1M tokens
Architecture
Dense transformer, architecture undisclosed; always-on adaptive thinking, effort-parameter controlled
License
Proprietary
Languages
API price (in/out)
$10 / $50
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
AA-BriefcaseAgents1574#4
best: Claude Opus 5 · 1720
Agents' Last ExamAgents48.5%#6
best: GPT-6 Astra · 59.3%
Arena EloHuman preference1505#1
best: this model · 1505
AutomationBenchAgents17.4%#10
best: Qwen3.8-Max-0902 · 50.8%
CursorBenchCoding70.5%#2
best: Claude Fable 5.1 · 73.4%
CyberGymCoding0.0%#15
best: GLM-5.3 · 84.5%
DeepSWECoding69.7%#4
best: Muse Spark 1.3 · 75.4%
FrontierBenchAgents33.8%#3
best: Claude Opus 5 · 43.3%
best: this model · 64.9%
FrontierCodeCoding53.5%#1
best: this model · 53.5%
GDPval-AAAgents1932#1
best: this model · 1932
GPQA DiamondReasoning92.6%#16
best: GPT-6 Astra · 96.0%
best: Grok 4.6 · 15.8%
Humanity's Last ExamReasoning53.3%#7
best: Claude Opus 5 · 64.7%
LiveCodeBenchCoding89.8%#7
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MMLU-ProKnowledge91.5%#1
best: this model · 91.5%
MMMUVision89.3%#1
best: this model · 89.3%
OSWorld 2.0Agents66.1%#4
best: Claude Fable 5.1 · 77.9%
OSWorld-VerifiedAgents85.0%#1
best: this model · 85.0%
ScreenSpot-ProVision87.3%#2
best: GPT-6 Astra · 92.7%
best: Claude Opus 5 · 89.5%
best: Claude Opus 5 · 59.4%
SWE-bench ProCoding80.0%#2
best: Claude Fable 5.1 · 81.2%
SWE-bench VerifiedCoding95.0%#2
best: Claude Opus 5 · 96.0%
Terminal-Bench 2.0Coding84.3%#10
best: Gemini 3.8 Flash · 89.4%
Terminal-Bench 3.0Agents34.1%#2
best: GPT-5.6 Sol · 34.6%
Terminal-Bench 4.0Agents42.0%#5
best: Claude Mythos 5.1 · 60.9%
best: GPT-6 Astra · 64.6%
best: GLM-5.3-Flash · 78.4%
API price $10/$50 · each benchmark row carries its own source badge (see methodology)