Not established from the available sources.
ANTHROPIC · MODEL RELEASE TRACKER
Claude Fable 5
Claude Fable 5 is an Anthropic model/agent variant that was tested for retaining knowledge across sessions by writing and reading note files. Anthropic reports that Fable 5 uses selective memory strategies and a post-session 'Dreaming' process to correct and structure its memories, and it performed better than an older Sonnet 3.7 agent in a Pokémon Red benchmark.CURRENT SNAPSHOT2/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Not established from the available sources.
CursorBench 3.2 performance comparisonAnthropic reports that on CursorBench 3.2 at max effort, Opus 5 performs within 0.5% of Claude Fable 5's peak score, indicating Fable 5's peak is marginally higher on that benchmark.
Benchmark evaluation (InnovationEval)Evaluated by Epoch AI on the 'InnovationEval' benchmark, which tests whether AI agents can independently invent, implement, test, and refine research methods.
Evaluation tasksTested on short-answer tasks (e.g., science questions) and coding tasks as part of the InnovationEval evaluation.
Compute and connectivity during evaluationDuring the evaluation, each agent had access to up to 3,000 hours of compute on high-end chips and no internet access.
Prior knowledge of SDPOEpoch reported that Claude Fable 5 had no prior knowledge of the SDPO method for the experiment.
MathVis benchmark pass@1Achieved a pass@1 score of 26.56% on MathVis, a 512-task benchmark for mathematical figure generation.
Included in the "Into the Unknown" benchmarkNamed as one of nine AI agents evaluated in the Carnegie Mellon "Into the Unknown" benchmark (agents tested included Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra, Deepseek V4 Pro, Kimi K3, and GLM-5.3).
Benchmark outcome and evaluation conditions (as reported)The benchmark report states that none of the 135 submissions from the nine agents met all requirements; during the evaluation agents (including Claude Fable 5) had internet connectivity, access to powerful machines, no time limit, and up to five restarts, with the runs totaling nearly 2,000 hours.
Pokémon Red memory-agent benchmark performanceIn Anthropic's Pokémon Red test, Claude Fable 5 earned four badges by using limited, selective memory, while the older Claude Sonnet 3.7 created many files and finished with one badge.
Cited as running in Replit Agent productionA Replit blog post in September cited Claude Fable 5 (and Claude Fable 5.1) as models running in Replit Agent’s production environment.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
CAPABILITIES · Described as 'frontier' level capabilityDEVELOPER CLAIM
Anthropic’s Opus 5 announcement characterizes Opus 5 as coming close to the "frontier intelligence" of Claude Fable 5, implying Fable 5 represents a frontier-level capability.
BENCHMARKS · CursorBench 3.2 performance comparisonDEVELOPER CLAIM
Anthropic reports that on CursorBench 3.2 at max effort, Opus 5 performs within 0.5% of Claude Fable 5's peak score, indicating Fable 5's peak is marginally higher on that benchmark.
SAFETY · Extra safeguards motivated by predecessor's risky cyber capabilitiesDEVELOPER CLAIM
Anthropic cited Claude Mythos Preview's advanced cyber capabilities posing high risks as the reason for adding safeguards that delayed Claude Fable 5.
AVAILABILITY · Cited as running in Replit Agent productionDEVELOPER CLAIM
A Replit blog post in September cited Claude Fable 5 (and Claude Fable 5.1) as models running in Replit Agent’s production environment.
BENCHMARKS · Benchmark evaluation (InnovationEval)DEVELOPER CLAIM
Evaluated by Epoch AI on the 'InnovationEval' benchmark, which tests whether AI agents can independently invent, implement, test, and refine research methods.
BENCHMARKS · Evaluation tasksDEVELOPER CLAIM
Tested on short-answer tasks (e.g., science questions) and coding tasks as part of the InnovationEval evaluation.
BENCHMARKS · Compute and connectivity during evaluationDEVELOPER CLAIM
During the evaluation, each agent had access to up to 3,000 hours of compute on high-end chips and no internet access.
BENCHMARKS · Prior knowledge of SDPODEVELOPER CLAIM
Epoch reported that Claude Fable 5 had no prior knowledge of the SDPO method for the experiment.
CAPABILITIES · Retries failed tasks with previous attemptsDEVELOPER CLAIM
Claude Fable 5 used a technique of retrying failed tasks while feeding it the previous failed attempts.
LIMITATIONS · No measurable improvement from retry strategyDEVELOPER CLAIM
The retry-with-previous-attempts technique produced no measurable improvement in the experiment.
LIMITATIONS · Performance vs human referenceDEVELOPER CLAIM
Epoch reported that Claude Fable 5 did not come close to the human reference on the InnovationEval tasks.
LIMITATIONS · Scientific judgment and creativityDEVELOPER CLAIM
Epoch (and Anthropic) found that models including Claude Fable 5 lack scientific self-criticism and genuine creative thinking when running experiments.
BENCHMARKS · MathVis benchmark pass@1INDEPENDENTLY SUPPORTED
Achieved a pass@1 score of 26.56% on MathVis, a 512-task benchmark for mathematical figure generation.
RELEASE · Release delayed for additional safeguardsDEVELOPER CLAIM
Anthropic delayed the release of Claude Fable 5 in order to add extra safeguards.
LIMITATIONS · Delay reduced revenue and market shareDEVELOPER CLAIM
Anthropic said the extra safeguards delaying Claude Fable 5 cost the company revenue and market share.
BENCHMARKS · Included in the "Into the Unknown" benchmarkDEVELOPER CLAIM
Named as one of nine AI agents evaluated in the Carnegie Mellon "Into the Unknown" benchmark (agents tested included Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra, Deepseek V4 Pro, Kimi K3, and GLM-5.3).
BENCHMARKS · Benchmark outcome and evaluation conditions (as reported)DEVELOPER CLAIM
The benchmark report states that none of the 135 submissions from the nine agents met all requirements; during the evaluation agents (including Claude Fable 5) had internet connectivity, access to powerful machines, no time limit, and up to five restarts, with the runs totaling nearly 2,000 hours.
CAPABILITIES · Session memory storage via note files and selective recordingDEVELOPER CLAIM
Fable 5 stores information across sessions by writing notes to files that it can read later. It limited its memory to a few files and avoided recording information already visible on screen, instead writing down plans, strategies, and rules it had tested.
CAPABILITIES · Post-session 'Dreaming' review processDEVELOPER CLAIM
Anthropic describes a post-session process called 'Dreaming' in which the system analyzes earlier session transcripts during idle periods to independently correct faulty notes and build structured knowledge collections.
BENCHMARKS · Pokémon Red memory-agent benchmark performanceDEVELOPER CLAIM
In Anthropic's Pokémon Red test, Claude Fable 5 earned four badges by using limited, selective memory, while the older Claude Sonnet 3.7 created many files and finished with one badge.
LIMITATIONS · Risk that faulty or outdated notes may be trustedDEVELOPER CLAIM
Anthropic notes a core problem is selection: faulty or outdated notes can be blindly trusted later by the agent, and rigid instructions about what to record can eliminate advantages that stronger models otherwise provide.
WHAT CHANGED
Stored passport versions, without reconstructed history.
CAPABILITIES · Post-session 'Dreaming' review process+ Anthropic describes a post-session process called 'Dreaming' in which the system analyzes earlier session transcripts during idle periods to independently correct faulty notes and build structured knowledge collections.
BENCHMARKS · Pokémon Red memory-agent benchmark performance+ In Anthropic's Pokémon Red test, Claude Fable 5 earned four badges by using limited, selective memory, while the older Claude Sonnet 3.7 created many files and finished with one badge.
LIMITATIONS · Risk that faulty or outdated notes may be trusted+ Anthropic notes a core problem is selection: faulty or outdated notes can be blindly trusted later by the agent, and rigid instructions about what to record can eliminate advantages that stronger models otherwise provide.
CAPABILITIES · Session memory storage via note files and selective recording+ Fable 5 stores information across sessions by writing notes to files that it can read later. It limited its memory to a few files and avoided recording information already visible on screen, instead writing down plans, strategies, and rules it had tested.
BENCHMARKS · Benchmark outcome and evaluation conditions (as reported)+ The benchmark report states that none of the 135 submissions from the nine agents met all requirements; during the evaluation agents (including Claude Fable 5) had internet connectivity, access to powerful machines, no time limit, and up to five restarts, with the runs totaling nearly 2,000 hours.
BENCHMARKS · Included in the "Into the Unknown" benchmark+ Named as one of nine AI agents evaluated in the Carnegie Mellon "Into the Unknown" benchmark (agents tested included Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra, Deepseek V4 Pro, Kimi K3, and GLM-5.3).
SAFETY · Biology safeguards updated to reduce false positives and fallbacks− Anthropic updated Claude Fable 5’s biology safeguards to substantially reduce false positives; Fable 5 users will experience many fewer “fallbacks” where the system switches to a less capable model after biology-related queries.
LIMITATIONS · Delay reduced revenue and market share+ Anthropic said the extra safeguards delaying Claude Fable 5 cost the company revenue and market share.
SAFETY · Extra safeguards motivated by predecessor's risky cyber capabilities+ Anthropic cited Claude Mythos Preview's advanced cyber capabilities posing high risks as the reason for adding safeguards that delayed Claude Fable 5.
RELEASE · Release delayed for additional safeguards+ Anthropic delayed the release of Claude Fable 5 in order to add extra safeguards.
SAFETY · Biology safeguards updated to reduce false positives and fallbacks+ Anthropic updated Claude Fable 5’s biology safeguards to substantially reduce false positives; Fable 5 users will experience many fewer “fallbacks” where the system switches to a less capable model after biology-related queries.
SAFETY · Subject to stricter safeguards that block some life‑science tasks− Anthropic describes its "generally available Fable models" as blocking a range of life-sciences tasks (examples given include drug discovery, research biology, clinical development, and manufacturing), and contrasts these safeguards with the Life Sciences Verification Program grants that allow broader access to other models.
AVAILABILITY · Cited as running in Replit Agent production+ A Replit blog post in September cited Claude Fable 5 (and Claude Fable 5.1) as models running in Replit Agent’s production environment.
AVAILABILITY · Reported production use in Replit Agent− A September blog post was cited as reporting that Claude Fable 5 (and Claude Fable 5.1) were running in Replit Agent’s production environment.
AVAILABILITY · Reported production use in Replit Agent+ A September blog post was cited as reporting that Claude Fable 5 (and Claude Fable 5.1) were running in Replit Agent’s production environment.
AVAILABILITY · Used in coding‑agent tests− An AWS blog describing agent-based SageMaker deployments cites tests that used "Auto or Claude Fable 5" with a coding agent, indicating Fable 5 was used in coding-agent deployment experiments.
BENCHMARKS · MathVis benchmark pass@1+ Achieved a pass@1 score of 26.56% on MathVis, a 512-task benchmark for mathematical figure generation.
BENCHMARKS · Compute and connectivity during evaluation+ During the evaluation, each agent had access to up to 3,000 hours of compute on high-end chips and no internet access.
BENCHMARKS · Evaluation tasks+ Tested on short-answer tasks (e.g., science questions) and coding tasks as part of the InnovationEval evaluation.
BENCHMARKS · Benchmark evaluation (InnovationEval)+ Evaluated by Epoch AI on the 'InnovationEval' benchmark, which tests whether AI agents can independently invent, implement, test, and refine research methods.
LIMITATIONS · No measurable improvement from retry strategy+ The retry-with-previous-attempts technique produced no measurable improvement in the experiment.
LIMITATIONS · Performance vs human reference+ Epoch reported that Claude Fable 5 did not come close to the human reference on the InnovationEval tasks.
CAPABILITIES · Retries failed tasks with previous attempts+ Claude Fable 5 used a technique of retrying failed tasks while feeding it the previous failed attempts.
LIMITATIONS · Scientific judgment and creativity+ Epoch (and Anthropic) found that models including Claude Fable 5 lack scientific self-criticism and genuine creative thinking when running experiments.
BENCHMARKS · Prior knowledge of SDPO+ Epoch reported that Claude Fable 5 had no prior knowledge of the SDPO method for the experiment.