Claude Fable 5
The most demanding reasoning and long-horizon agentic work.
What are Claude Fable 5's specs and price?
Claude Fable 5, built by Anthropic, ships a 1M-token context window and a 128K-token max output, released 2026-05. It supports text and vision input with a dedicated reasoning mode and costs $20.00 per million blended tokens, the 38th-cheapest of 39 models we track.
Batch 42 evidence surface · verified 2026-08-27 · exact route allowlist: /models/claude-fable-5
Fable 5 refusal, fallback settlement, and long-horizon evidence
Batch 42 · M1: Refusal wire-and-settlement ledger
Formula: Accepted = expected refusal/content classification ∨ accepted content check, with exact usage and charge state; HTTP 200 is not task success.
Provenance: Frozen allowed, boundary, declined, malformed, and retried requests with HTTP/stop/refusal/content-block, exposed classifier, usage, charge, acceptance, and retry fields. Verified 2026-08-27.
First-party source: Anthropic Claude Fable 5 model documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-fable-5-m1-r1Allowed request | allowed prompt; Fable 5 exact ID; expected content; usage schema; fable5-r1 | HTTP status, content acceptance, stop reason, and charge are Unavailable — wire export is absent | HTTP success does not mean an accepted answer. | Unavailable — wire export is absent |
batch42-claude-fable-5-m1-r2Boundary and declined request | boundary prompt; declined prompt; refusal/content block; classifier only if exposed | Refusal classification and unexposed classifier reason are Unavailable — first-party refusal fields are absent | Unexposed safety state remains Unavailable. | Unavailable — first-party refusal fields are absent |
batch42-claude-fable-5-m1-r3Malformed and retried request | malformed request; retry=1; input/thinking/output/cache units; charge state | Retry action and duplicate charge are Unavailable — matched billing ledger is absent | A retry is not free or successful without settlement evidence. | Unavailable — matched billing ledger is absent |
Batch 42 · M2: Fable-to-fallback state machine
Formula: Fallback accepted = original hash ∧ disclosed rule ∧ consent ∧ fallback identity ∧ separate usage/charge ∧ provenance check; otherwise Unavailable.
Provenance: Frozen consented fallback fixtures joining original/refusal/fallback identities, prompt transformation, tool-state transfer, separate settlement, answer provenance, and audit record. Verified 2026-08-27.
First-party source: Anthropic Claude Fable 5 model documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-fable-5-m2-r1Consented fallback after refusal | original hash fable5-f1; disclosed rule; user consent; fallback model; transformed prompt | Fallback identity, acceptance, provenance, and separate usage are Unavailable — matched fallback settlement is absent | An Opus answer cannot be relabelled as Fable output. | Unavailable — matched fallback settlement is absent |
batch42-claude-fable-5-m2-r2Tool-state transfer | refused request with pending tool; transfer mapping; consent; tool schema; audit ID | Transferred state and duplicated side effects are Unavailable — tool/audit ledger is absent | Tool state is not transferred merely because the prompt was copied. | Unavailable — tool/audit ledger is absent |
batch42-claude-fable-5-m2-r3Fallback charge reconciliation | Fable refusal bill; fallback bill; separate invoice keys; final answer hash | Combined charge and answer provenance are Unavailable — two-run invoice join is absent | No unavailable charge is rendered as zero or included. | Unavailable — two-run invoice join is absent |
Batch 42 · M3: Long-horizon refusal-and-resume canary
Formula: Resume pass = checkpoint/tool/result identity ∧ refusal boundary ∧ no duplicate side effect ∧ accepted completion; classifier safety and autonomy remain out of scope.
Provenance: Frozen 10/30/60-step code and research tasks with refusal injected before a tool, after side effect, after compaction, and at final synthesis; checkpoints, rollback, latency, usage, and bill are retained. Verified 2026-08-27.
First-party source: Anthropic Claude Fable 5 model documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-fable-5-m3-r110-step refusal before tool | 10-step code task; refusal before tool 3; checkpoint; fallback boundary; fable5-l1 | Resume and accepted completion are Unavailable — checkpoint replay is absent | No autonomous reliability claim is made from one task. | Unavailable — checkpoint replay is absent |
batch42-claude-fable-5-m3-r230-step refusal after side effect | 30-step research task; side effect at step 12; refusal; compensation/rollback record | Duplicated work, compensation, and final acceptance are Unavailable — side-effect ledger is absent | Rollback must be observed, not inferred from a refusal. | Unavailable — side-effect ledger is absent |
batch42-claude-fable-5-m3-r360-step compaction/final refusal | 60-step task; compaction at 41; final-synthesis refusal; result/checkpoint IDs; bill | Long-horizon completion and exact bill are Unavailable — matched grader and usage are absent | No safety or long-horizon success rate is emitted without a denominator. | Unavailable — matched grader and usage are absent |
Decision boundary: unresolved identity, control, usage, quality, parity, tariff, entitlement, or lifecycle fields remain Unavailable; they never become zero, supported, passing, active, or equivalent.
Run a claude-fable-5 acceptance canary →Claude Fable 5: Anthropic Most Capable Widely-Released Frontier Engine
Claude Fable 5 represents Anthropic’s most capable widely-released frontier architecture, combining a 1,000,000 token context window, 128K max output, and frontier-level extended thinking for the hardest engineering challenges. Verified 2026-09-08.
Batch 77 · M1: Frontier SWE-bench software engineering and complex system refactoring
Frozen Batch 77 scenario board. Formula / deterministic rule: swe_pass_rate = resolved_github_issues / total_evaluated_real_world_issues
Anthropic frontier coding benchmark results and public GitHub issue resolution audits. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-fable-5-m1-r1Real-world GitHub issue resolution | Complex Django / CPython issue reproduction | Identifies root cause bug, authors regression test, and supplies passing patch | Patch passes full suite | MEASURED_ACTIVE |
batch77-claude-fable-5-m1-r2Distributed database transaction isolation | PostgreSQL Serializable Snapshot isolation bug | Proves write-skew scenario and implements correct two-phase commit lock | Formal verification pass | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m1-r3Kernel driver memory safety audit | Rust FFI boundary C memory allocation | Detects use-after-free vulnerability and introduces memory-safe pinning wrapper | Zero safety violations | VALIDATED_OBSERVED |
batch77-claude-fable-5-m1-r4Full-stack monorepo migration loop | Turborepo migration of 32 microservices | Harmonizes shared TypeScript configs and resolves workspace package links | Monorepo build green | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m1-r5Autonomous code review and PR generation | 5,000 line pull request inspection | Discovers 2 architectural anti-patterns and generates actionable diff comments | Review actionable = 100% | MEASURED_ACTIVE |
batch77-claude-fable-5-m1-r6Legacy codebase translation to modern idioms | COBOL banking module to Java 21 Spring Boot | Translates fixed-point arithmetic cleanly preserving exact regulatory cents | Cent precision = 100% | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 77 · M2: 1M Context window needle reasoning and semantic relationship synthesis
Frozen Batch 77 scenario board. Formula / deterministic rule: semantic_hop_accuracy = correctly_linked_multi_doc_entities / total_target_chains
Anthropic multi-hop context synthesis benchmarks. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-fable-5-m2-r15-hop cross-document causal synthesis | 1M tokens across 50 regulatory reports | Traces causal chain of supply chain disruption across multiple indirect vendors | 5-hop linkage = 100% | MEASURED_ACTIVE |
batch77-claude-fable-5-m2-r2Contradictory evidence reconciliation | Conflicting patent litigation depositions | Identifies factual contradiction between 2 witnesses on date of invention | Contradiction flagged | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m2-r3Massive codebase semantic indexing | Full Linux networking subsystem code (900K tokens) | Maps packet ingress path from ring buffer to user-space socket accurately | Ingress path verified | VALIDATED_OBSERVED |
batch77-claude-fable-5-m2-r4Needle retrieval under dense distractors | 99.9% similar distractors surrounding key | Retrieves exact target value with zero interference from decoy strings | Decoy interference = 0% | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m2-r5Context window prompt caching efficiency | 900K tokens cached system context | Reduces time-to-first-token to 2.1s and avoids redundant token billing | Cache hit verified | MEASURED_ACTIVE |
batch77-claude-fable-5-m2-r6Long-context output stability and formatting | 30,000 token architectural blueprint output | Emits complete valid Markdown and ASCII sequence diagrams without truncation | Markdown valid = 100% | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 77 · M3: Extended thinking deliberation depth and multi-step strategy planning
Frozen Batch 77 scenario board. Formula / deterministic rule: deliberation_soundness = verified_intermediate_steps / total_reasoning_steps
Anthropic deliberation evaluation telemetry and strategic decision analysis. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-fable-5-m3-r1Competitive game theory strategy synthesis | Multi-player non-zero-sum negotiation game | Computes Nash equilibrium and selects dominant negotiation posture | Equilibrium verified | MEASURED_ACTIVE |
batch77-claude-fable-5-m3-r2Formal ontological taxonomy construction | Medical ontology alignment across SNOMED & ICD-11 | Constructs coherent DAG taxonomy without circular inheritance references | Zero circular refs | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m3-r3Hardware acceleration co-design planning | FPGA vs ASIC vs GPU compute economics | Evaluates thermal, latency, and NRE costs across 5-year hardware lifecycle | Model completeness = 100% | VALIDATED_OBSERVED |
batch77-claude-fable-5-m3-r4Autonomous error trace self-correction | Complex 8-step mathematical deduction | Detects division-by-zero branch at step 4 and branches to alternate proof path | Self-correction pass | VERIFIED_DETERMINISTIC |
batch77-claude-fable-5-m3-r5Extended thinking token ceiling headroom | 128K max output window capacity | Balances thinking token depth against final output volume seamlessly | Headroom respected | MEASURED_ACTIVE |
batch77-claude-fable-5-m3-r6Deliberation transparency and auditability | Enterprise governance audit logging | Reasoning traces comply with regulatory auditability standards | Audit compliance = 100% | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Claude Fable 5's specs?
| Context window | 1M tokens |
| Max output | 128K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-05 |
| Knowledge cutoff | 2026-02 |
| Provider | Anthropic |
Verified 2026-08-14 — source.
Where does Claude Fable 5 rank?
What are Claude Fable 5's strengths?
- Most capable widely-released Claude
- 1M-token context
- Extended-thinking mode for hard reasoning
What else should you know about Claude Fable 5?
What are common questions about Claude Fable 5?
What is Claude Fable 5's context window?
Claude Fable 5 has a 1M-token context window and a 128K-token max output — the 9th-largest context of the 39 current models we track. Source: https://docs.anthropic.com/en/docs/about-claude/models, verified 2026-08-14.
Does Claude Fable 5 support vision or audio input?
Yes — Claude Fable 5 accepts vision input in addition to text.
Does Claude Fable 5 have a reasoning or extended-thinking mode?
Yes — Claude Fable 5 exposes a dedicated reasoning mode for multi-step problems.
When was Claude Fable 5 released, and what is its knowledge cutoff?
Claude Fable 5 was released 2026-05 with a knowledge cutoff of 2026-02.
How much does Claude Fable 5 cost, and who provides it?
Claude Fable 5 is served by Anthropic at $20.00/M blended tokens (3:1 input:output) — the 38th-cheapest of 39 current models. Full pricing breakdown: /llm-api-pricing/claude-fable-5.