AI Search's Fable 5.2 review is at its best when the model has to finish a whole artifact, not just answer a prompt. Claude directs a coded history film, plays a virtual piano, produces a song through audio software, reconstructs a property in Blender and builds a playable Unreal scene. Then two image tests puncture the easy “best at everything” story. Here is what those demonstrations actually establish, checked against Anthropic's November 9 launch.
Fable 5.2 looks compelling for long, tool-heavy creative builds in this review. The outputs took roughly 35 minutes to seven hours, sometimes with manual continuation after a usage limit. The creator's comparison is subjective, not a controlled benchmark. For ordinary work, Anthropic still recommends Opus 5.5 at medium effort first; test Fable when the problem is unusually hard and the result is worth its higher cost.
Watch the Full Review
Creator: AI Search. The Higgsfield segment is sponsored. Higgsfield is a separate image and video generation service; do not attribute its outputs to Fable's native generation.
The Creative Build Scorecard
| Test | What the video shows | What it does not prove |
|---|---|---|
| History animation | Claude codes and revises a roughly 90-second film, inspects frames and uses Gemini TTS for narration. The creator reports about 90 minutes and one visual revision. | The film is not made by a native video generator alone. Its history and predictions still need editorial fact-checking. |
| Virtual piano | It composes and plays through an on-screen keyboard. The creator says Opus 5.5 already handles this test well. | One performance does not establish a meaningful Fable lead. |
| Full song | It assembles vocals, instrument parts and effects using a voice tool, digital audio workstation, plug-ins and samples. The creator praises the instrumental but dislikes the synthetic vocals. | The model did not create a polished record unaided; the run crossed a five-hour usage boundary. |
| Property tour | From listing photos, it builds a detailed Blender scene and camera path in about 90 minutes, by the creator's account. | Photos cannot verify unseen rooms or a complete floor plan. Visual similarity is not measured spatial accuracy. |
| Unreal game | It works across a character asset, animation source, Blender and Unreal, with a screenshot-based critic loop. The creator reports about seven hours and two limit interruptions. | “AAA quality” is a prompt's scoring label, not an independent production-grade review. Asset licenses and performance still need checking. |
| Physics scene | A configurable, coded water-balloon impact looks convincing on screen after about 35 minutes. | Appearance alone cannot validate the underlying physics. The controls and output need numerical checks against a known simulation. |
The common thread is orchestration: Claude plans, writes code, operates tools, inspects an output and revises it. That matters more than the “one prompt” framing. The song and game also show the cost of that autonomy: long sessions can hit limits, require a human to resume and still need artistic review.
Where the Demos Did Not Hold Up
At 21:05, Fable cannot confidently find a hidden frog in a crowded image and marks other objects instead. That is a useful counterexample to the stronger creative demos, although a single image is not a vision benchmark. The creator's informal comparison with another model has no published matched protocol here.
At 22:18, the creator asks the model to classify six brain-scan images and scores it at two correct. This is not evidence of diagnostic accuracy: the image set, reference labels, clinical context and scoring were not independently verified. The subsequent leukemia research answer at 23:17 is assessed mainly for perceived completeness. Neither demonstration should guide patient care, and neither supports a claim of medical superiority.
The animal-welfare ideation section is similarly an idea-generation exercise. Evaluate the proposed interventions with domain experts and actual welfare outcomes before calling them innovations.
What Anthropic's Evidence Adds
Anthropic reports gains in coding, reasoning and vision, plus three working electronic prototypes, robotics progress and the formalization of Fermat's Last Theorem by a 1,000-agent team in 30 hours. Formalizing an existing theorem means making a proof machine-checkable, not discovering a new theorem. Anthropic also describes two black-hole theory results; those are research claims supported by linked preprints, not results reproduced by this video.
| Anthropic evaluation | Fable 5.2 | Comparison |
|---|---|---|
| FrontierCode 2.0, agentic coding | 49.1% at xhigh | Opus 5.5: 45.5% |
| FrontierSWE v2, agentic coding | 69.9% | Opus 5.5: 62.3% |
| Remote Labor Index, knowledge work | 41.3% at high | Opus 5.5: 21.3% |
| AutomationBench, business workflows | 51.4% | Gemini 4 Carbon: 53.9% |
These are vendor-published results with different effort settings and test harnesses; the last row also shows that Fable does not lead every reported category. They are more structured than the video demos, but they are not a substitute for testing your own workflow. Anthropic's model-selection guide recommends beginning most coding and knowledge work with Opus 5.5 at medium effort, then moving to Fable for the hardest or longest tasks.
On safeguards, the video uses “nerfed” as a broad description. Anthropic's narrower account says it limited training gains in specific AI-acceleration research, including accelerator-kernel optimization and large-model architecture work, while leaving related agentic coding and traditional ML research unaffected. It says those training changes are not applied as an inference-time downgrade, though some requests may be blocked or transparently routed elsewhere. Anthropic reports fewer unnecessary cyber and biology blocks than Fable 5.1. That is different from saying the model generally gives weaker answers on all biology, cybersecurity or AI research.
Price, Speed and Plan Limits
As of 11 November 2026, Anthropic lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens on the Claude API, versus Opus 5.5 at $4 and $20 in its model guide. Optional fast mode doubles Fable's token price. Anthropic says Fable can generate output up to 30% faster than Fable 5.1, but that does not make every end-to-end agent run quick. The creator's 35-minute to seven-hour jobs include tool use, iteration and interruptions.
Claude plans have a five-hour session limit where applicable and weekly limits; usage depends on model, effort and task. The launch offer gives eligible Pro and standard Team seats included Fable usage up to 50% of weekly limits through 16 November 2026, 11:59 pm PT, then reverts to usage credits. Max, Ultra and eligible premium seats have separate Fable limits and temporary resets. A reset restores usage; it does not raise the cap. Check Settings > Usage before committing to a long build. The video's “twice as fast through quota” phrasing is a useful warning, not a universal conversion rate.
A Better Test for Your Own Work
- Choose one job whose finished output you can inspect: a small animation, bounded 3D scene or game prototype. Specify the tools, acceptance checks and a time or token budget.
- Run the same brief with Opus 5.5 at medium effort and Fable 5.2 at medium or high. Keep the same assets, environment and human interventions.
- Record elapsed time, usage, number of interventions, quality defects and cost per accepted result. Review sources, licenses and physical or medical claims separately.
- Use Fable only where the quality lift or reduced rework pays for the extra spend. Save the prompt, files and review notes so the comparison can be repeated.
My read: Fable's strongest showing here is keeping a long, messy creative build moving across tools. Its weakest evidence is the leap from a visually persuasive demo to factual, physical or clinical correctness. Treat the former as a reason to prototype and the latter as a reason to verify.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Fable 5.2 introduction |
| 00:23 | Claude Code setup |
| 00:54 | Coded animation and narration |
| 04:41 | Virtual piano performance |
| 06:46 | Full song production |
| 13:07 | Blender real-estate tour |
| 15:49 | Higgsfield sponsor segment |
| 17:04 | Unreal game build |
| 19:52 | Water-balloon physics demo |
| 21:05 | Hidden-frog image test |
| 22:18 | Brain-scan classification test |
| 23:17 | Medical research answer |
| 24:17 | Animal-welfare ideation |
| 25:15 | Specs and safeguards |
| 26:18 | Anthropic research examples |
| 28:00 | Benchmarks and cost |
| 31:00 | Usage and verdict |
Sources and Method
Checked 11 November 2026. The AI Search video and supplied transcript establish what the creator attempted, displayed and reported. We did not receive the project files, raw timing logs, medical reference labels or a controlled model-comparison dataset. Anthropic's launch article, model guide, launch-offer terms and usage-limit guide support the official specifications, benchmark and access details. Creator impressions, vendor evaluations and this article's recommendations are intentionally labeled as different kinds of evidence.