Source: raw/I_Tested_Claude_Fable_5.2_vs_Opus_5.5_WILD_RESULTS.md, raw/I_Made_Claude_Fable_5.2_Opus_5.5_Build_the_Same_App_RAW_RESULTS.md, raw/You_re_Using_Fable_5.2_WRONG_It_s_a_Superpower_With_Opus_5.5.md, raw/Fable_5.2_or_Opus_When_to_use_each_model_in_Claude_Code.md (four YouTube videos, all fetched 2026-11-11, launch week of Fable 5.2). Added 2026-11-13: raw/The_NEW_Claude_Fable_5.2_-_No_Hype_Tests_and_7_Great_Use_Cases.md (Jay, https://www.youtube.com/watch?v=QCSavlDtZOE) and raw/Fable_5.2_Is_INCREDIBLE_at_Video_Editing_WILD_RESULTS.md (Brock Mesarich ^[inferred], https://www.youtube.com/watch?v=1fBhLGhQQ2M).
Four videos from Fable 5.2’s launch week put it against Opus 5.5: two same-prompt bake-offs with measured cost and time, one practitioner’s model split across five real workflows, and a short Anthropic explainer on which model to use. Fable won on output quality in most tests, but the cost gap ranged from about 14% more to about 2.9 times as much, depending on the task and the effort setting. All four end in the same place as Anthropic’s own advice: Opus 5.5 for daily work, and Fable for the large, open-ended job, often as the planner or lead with a cheaper model doing the scoped parts.
Key Takeaways
- Fable won output quality in 4 of 6 same-prompt tests across the two bake-offs, with one near-tie (Brock’s website) and one Opus win (video editing, where Fable left a repeated word in the cut).
- The cost gap is task-dependent, not a fixed 2x. ^[ambiguous] Brock’s three tests at
mediumcost Fable only about 14% more in total, and Fable was cheaper on the game. The second tester’s runs atxhighcost Fable about twice as much on a cartoon and about 2.9 times as much on a bug hunt (114). - Speed also varied. Fable finished faster in two of Brock’s three tests. In the second tester’s runs it was 20 minutes slower on one, tied on one, and slower on the bug hunt.
- Bug hunting is where Fable earned its price most clearly. 63 confirmed bugs against Opus’s 54 in one test, with one hallucinated bug; Peter Yang’s Fable audit found 30 issues in an app he had already cleaned up with Opus.
- The common split: Fable plans or leads, a cheaper model executes. Anthropic’s video has Fable lead a framework migration and hand package upgrades to Sonnet subagents; the second tester ends on “Opus as my daily driver with Fable as the planning agent for my bigger build”; Peter Yang builds with Opus and uses Fable to audit and plan big features.
- Effort matters for the comparison. Brock ran both models at the
mediumdefault; the second tester ran atxhigh. Results at one effort level do not transfer to another.
The four sources
| Video | Creator | Format | Effort |
|---|---|---|---|
| ”I Tested Claude Fable 5.2 vs Opus 5.5 (WILD RESULTS)”, https://www.youtube.com/watch?v=AQWgE5LWQ4A | Brock Mesarich (channel confirmed from YouTube metadata on 2026-11-13; the transcript gives only his sponsor code “Brock”; compare his earlier test in [[claude-ai/which-claude-model-for-which-job | Which Claude Model for Which Job]]) | 3 tests, one run each, cost and time per run |
| ”I Made Claude Fable 5.2 & Opus 5.5 Build the Same App (RAW RESULTS)”, https://www.youtube.com/watch?v=PhCpq0Eagak | Not named in the transcript; sponsored by Fish Audio | 3 tests, 4 points each (2 output quality, 1 cheaper, 1 faster) | xhigh stated for test 1 |
| ”You’re Using Fable 5.2 WRONG (It’s a Superpower With Opus 5.5)”, https://www.youtube.com/watch?v=fAkm3YQg73s | Peter Yang (Behind the Craft); sponsored by Anthropic | Five workflows and which model he runs each on | Relays “medium or high" |
| "Fable 5.2 or Opus: When to use each model in Claude Code”, https://www.youtube.com/watch?v=C33hJGt-ySg | Anthropic (the narrator speaks for Anthropic: “the starting point we recommend”; channel not named in the transcript) | Short explainer with one example shop | Opus at medium |
All four: Platform: YouTube. Durations not recorded in the transcripts.
Test 1: Brock, three tasks at default effort
New session per test, identical prompt, no help mid-run, both at medium “because that’s probably what most of us are going to be using.”
| Task | Opus 5.5 | Fable 5.2 | His verdict |
|---|---|---|---|
| Launch site for a made-up e-bike brand (bike drawn in code, comes apart on scroll, configurator, checkout) | $11.66, ~34 min | $13.25, ~26 min | Close. Fable added a range calculator and savings breakdown; “not a huge difference” |
| Browser 3D jungle-temple game, told to play through it and screenshot-check its own work | $22.65, ~56 min | $22.52, ~47 min | Fable clearly: the rope bridge broke as asked, the character moved naturally. Opus’s bridge did not break |
| Edit a raw talking-head short from reference videos and a one-page style brief, no templates or editing software | $5.78, ~13 min | $9.77, ~15 min | Opus: Fable left a repeated “and” in the cut |
- Totals: Opus 45.54, about 14% more for Fable. ^[inferred] (sum of his figures; he says “about 14% more”)
- Why per-task cost differs from per-token price: “the two models don’t use the same amount of work in order to finish the job.”
- His rule: Opus 5.5 for most everyday work (“I really like the way that Opus communicates with me”); Fable for “something big and complex … where I need the result to look as good as possible.”
- The self-check instruction in the game prompt (play it, screenshot, keep improving) is where he thinks “we see the difference between these different models.”
Test 2: three tasks at xhigh, with a cross-vendor judge
The second tester chose xhigh because Anthropic reported Fable’s top agentic-coding score at xhigh, with a dip after max (the launch page’s 49.1% FrontierCode score is at xhigh; see Fable 5.2). A usage tracker logged tokens, time and cost.
| Task | Output points | Time | Cost |
|---|---|---|---|
| 60-90 s 2D cartoon on “the evolution of the Claude mascot”, drawn in code, narrated via the Fish Audio MCP | Fable (better jokes and sound mix; he preferred Opus’s animation style) | Opus (Fable took “an extra 20 minutes”) | Opus (“51% less than Fable”) |
| Plants vs. Zombies clone with an economy and escalating waves | Fable (cleaner art, original plant designs) | Tie | Opus (Fable “$5 more”; Opus used more tokens) |
Bug hunt on his own app, Ace; GPT-6 Astra at xhigh judges each bug, -1 per hallucinated one | Fable: 63 confirmed, 1 rejected, 3 duplicates, score 62. Opus: 54 confirmed, 1 duplicate | Opus faster (Fable ran 58 min) | Opus 329** |
- Final score: Fable 7, Opus 6. Fable won output in all three; Opus won cost in all three and time in two.
- His cost framing: “imagine paying a senior engineer $330 per hour.”
- His conclusion: Opus gives “I’m going to say 80% of the results,” so “I’m using Opus as my daily driver with Fable as the planning agent for my bigger build.”
- Limit: in the PvZ test he says both that Opus used “significantly more tokens” and that it cost “nearly half the amount of tokens in terms of API usage.” The cost ranking is clear; the token claim is not.
Peter Yang: which model for which workflow
Peter Yang says he has spent “over 500 hours” with both models in Claude Code. The video is sponsored by Anthropic. His rule: “Opus is still the best daily driver,” and Fable “is for your hardest and most important work.”
| Workflow | Model | Why |
|---|---|---|
| Chief of staff: a skill that checks email and calendar for the last 7 days and returns a ranked, numbered list; scheduled at 7 a.m. and 4 p.m. | Opus | ”Doesn’t require the most powerful model” |
| Career and life advisor grounded in a personal-plan doc (goals, principles, what gives and takes energy) | Fable | Strategic; “you’ll be surprised at how blunt and truthful it can be” |
| Podcast production as a chain of skills (Riverside MCP for the recording, Suno MCP for music, a sizzle reel in about 20 minutes) | Opus, Fable if the workflow gets “super complicated” | Execution |
Personal mobile app built with /design in Claude Code | Opus to build; Fable “to find bugs and issues with my app and make plans for big features” | Fable found 30 issues, sorted by priority, in an app he had been cleaning up with Opus |
| Caring for a family member through cancer treatment (one-page brief, clinical-trial research before doctor visits) | Fable | ”Nothing matters more” |
- Tips he relays from “Alex from Anthropic”: ask Fable what is wrong with your idea before it builds; start at
mediumorhigh, “no need to use extra high or max”; say up front what needs approval; give it a whole project, not a next step. He adds: “I know most Anthropic employees use Fable on medium effort.” The same four tips are on Prompting Claude Fable 5.2, credited to Alex Albert. - Audit everything on a new model. “Whenever I get access to a new and better model, I always like to run something like this on all my code bases”: “please find bugs and issues with this project so we can fix and do potential UI cleanup improvements.”
- Skill audit across sessions, run on Fable. He invokes
/claude-api prompt-auditand asks Claude to mine past sessions for repeated feedback; in his example it read 90 days and 149 sessions and proposed exact copy edits to his skills, which he accepts or rejects line by line. Reason: improving a skill from one conversation “tends to overfit.” (He calls it “a secret Claude prompt”; theprompt-auditargument was not checked against Anthropic docs.) - Plan facts he cites: on Max, Fable is included “for up to half of your weekly limit”; on the API it costs twice as much per token. Both match Anthropic’s pages.
Anthropic’s explainer: four models, one shop
Anthropic’s video uses an online shop to place each model:
- Opus 5.5 at
medium: add monthly subscriptions (touches data model, API and UI). “Most of the time, the answer is Opus 5.5 on medium effort.” If it falls short, raise effort first: “a smaller step than changing the model, and it usually keeps the prompt cache.” - Fable 5.2: migrate the whole shop to a new architecture, “a large, loosely defined problem … with dozens of milestones and decisions that nobody has written down yet.” Give it the goal and the non-negotiables (“Checkout can’t go down and orders can’t be lost”); “it makes the plan, does the work, and tests as it goes.”
- Sonnet: a bug report with a clear reproduction, where “I can say what done looks like in one sentence.” Faster than Opus “at half the price per token.”
- Haiku: sorting each incoming customer message through the Claude API, “high-volume, cost-sensitive work like summarization, classification, and routing.”
- Split one job across models: for a framework upgrade across every package, “I’ve asked Fable to lead the migration and hand those parts to Sonnet” as subagents, since “upgrading each package and fixing its tests is a well-scoped task.”
2026-11-13: Jay’s “no hype” tests across seven use cases
Source: raw/The_NEW_Claude_Fable_5.2_-_No_Hype_Tests_and_7_Great_Use_Cases.md · Creator: Jay E (channel “Jay E | RoboNuggets”, from YouTube metadata on 2026-11-13; says he leads the AI business of “one of the largest AI communities globally”) · URL: https://www.youtube.com/watch?v=QCSavlDtZOE · Platform: YouTube · Duration: not recorded. Fresh Claude Code instance, same prompt and effort level for both models; effort level not stated.
| Use case | His verdict | Cost and time (where given) |
|---|---|---|
| 3D room animation coded in Blender, four improvement loops | Fable “more cinematic” (depth of field, lighting) | Fable 24, 2 h 20 min |
| Explorable 3D Sydney Opera House | Fable more accurate: correct two levels, plaza, Royal Botanic Gardens; Opus put trees where there are none | Fable “significantly much more expensive” (no figure) |
| Human-eye anatomy explainer | About equal; Fable added veins and eye sockets, “executes each task more completely” | Not given |
| Motion graphics (no design system, a collage style, and his own design system) | “Pretty much like for like” | Opus cheaper |
| Landing pages from a rubric and palette | ”No clear winner”; with his design system Fable “cost it twice as much” | Fable about 2x |
| Marketing copy on an infographic page | Neither passes; both “smell like AI” (eyebrow headings, extra lines such as “the whole secret is this loop”) | — |
- First-pass output from a single prompt was well short of the demos he had seen. What closed the gap was a loop: show the model its current output and ask it to improve the visual look, repeated (four versions here). He points to Matt Shumer’s gauntlet loop for the extreme version; see Gauntlet Loop.
- Cost figures he relays from Artificial Analysis: Fable 5.2 about 2.6x Opus per task and Intelligence Index 65 vs 59. Both match the AA figures already relayed on Fable 5.2 (2.6x at
max); AA was still not checked directly. - Usage limits: he warns that heavy Fable runs will hit the 5-hour usage limit quickly.
- His rule: Opus 5.5 is “clearly enough” for most people; Fable for tasks complex enough to use it fully, such as AAA-quality 3D or accurate real-world scenes. He would “rather hire someone who’s really good … just using Opus versus a complete beginner who just uses Fable.”
- For motion graphics and landing pages, skills and design systems matter more than the model. This matches Claire Vo: Fable follows a style guide well.
- Build your own benchmark. His starting prompt idea: ask Claude which skill or workflow was the most token-intensive over the past week, then run that same task on each new model and compare quality and cost on one page. His own is a 10-second animation run across Opus, Sonnet and OpenAI models.
- Demos he cites (not checked): a Blender room by “Riley” with his own Blender skills; a game built as a Claude artifact and a 3D personal homepage by Anthropic’s Felix Riesberg (transcribed “Felix Wiesberg”); a code-only LEGO Lord of the Rings animation.
2026-11-13: Brock moves his own video editing to Fable
In a follow-up video (raw/Fable_5.2_Is_INCREDIBLE_at_Video_Editing_WILD_RESULTS.md; full system in Brock Mesarich’s video-editing system) Brock says he has switched his production editing from Opus 5.5 to Fable 5.2. His reason: Opus left chopped words and bad takes in cuts. He puts Fable at “about 14% more than Opus 5.5” on “similar editing skills”.
- This reverses his bake-off result above, where Opus won the editing test because Fable left a repeated “and”. ^[ambiguous] The bake-off was one run; the switch is based on weeks of use he does not quantify.
- His 14% matches his three-task bake-off total; the editing task alone cost Fable about 69% more (5.78).
Where the sources disagree
| Question | Brock (medium) | Second tester (xhigh) | Earlier tests in the wiki |
|---|---|---|---|
| Fable cost vs Opus, per task | +14% total; cheaper on the game | About 2x (cartoon); 2.9x (bug hunt) | Ben AI, both at high: about 70% more over three tasks ([[claude-ai/claude-fable-5-2-mythos-5-2 |
| Fable speed vs Opus | Faster on 2 of 3 | Slower or tied | Ben AI: slower on both timed tasks |
| Jay (2026-11-13), 3D loop and landing pages | 24 (about 3.3x) and about 2x; slower on the 3D task (4 h 30 vs 2 h 20) | ||
| Anthropic’s figure | FrontierCode 2.0: Fable high matched Opus max “for roughly half the cost per task” |
- These are single runs on different tasks at different effort levels, so the spread is not a contradiction of Anthropic’s benchmark claim, which compares Fable
highwith Opusmax. - What all four agree on: start on Opus 5.5; move to Fable when the job is large, open-ended or high-stakes.
Try It
- Run your own three-task bake-off at the same effort on both models, fresh session each, one prompt, no help. Record cost and time from
/usage. Include one task Opus already does well (Brock’s editing test) and one hard, open-ended one. - Use Fable as the planner on your next big change. Give it the goal and two or three non-negotiables, have it write the plan and milestones, then hand scoped steps to Opus or Sonnet subagents (set
model:in the subagent definition). See Orchestrator + Cheap-Worker Routing. - Run one Fable bug audit on a codebase you know well, and check each finding before fixing. The second tester used a different vendor’s model as judge to catch hallucinated bugs.
- Tell the model to check its own output (play the game, screenshot, iterate). Brock’s game prompt did this, and it is the test where Fable pulled furthest ahead.
- Mine your past sessions for repeated feedback before editing a skill, rather than tuning it from one conversation.
- Build a personal benchmark from your heaviest skill (Jay): ask Claude which skill used the most tokens last week, then rerun that exact task on each new model and record quality and cost side by side.
- Loop visual output before judging a model. Give the same “improve the visual look” prompt three or four times; Jay’s first passes were far below what the loops produced.
Open Questions
- Second tester’s identity and efforts for tests 2 and 3. The transcript gives neither;
xhighis stated only for the cartoon. /claude-api prompt-audit. Peter Yang’s command was not checked against Claude Code documentation.- Durations of the four videos were not captured.
- Jay’s effort level is not stated; his Sydney Opera House and eye-anatomy runs give no cost figures.
- Felix Riesberg’s game artifact and homepage, and Riley’s Blender room, were not checked.
Related
- Claude Fable 5.2 + Mythos 5.2 — pricing, benchmarks and Anthropic’s routing rule.
- Claude Opus 5.5 — the daily driver all four sources keep.
- Which Claude Model for Which Job — routing across Haiku, Sonnet, Opus and Fable.
- Prompting Claude Fable 5.2 — effort settings and Alex Albert’s tips.
- Which Effort Level to Use — why
mediumandxhighresults differ. - Orchestrator + Cheap-Worker Routing — the same split across ecosystems.
- Picking the Right Model — running your own comparison.
- Brock Mesarich’s video-editing system — the production setup he moved to Fable.