Forward FutureSIGNALSHow we test ↗
ANTHROPIC / MODEL REVIEWFIELD TESTED · NOV 9, 2026

Fable 5.2,
put to work.

What can you actually hand over?
We gave Anthropic’s latest model five demanding assignments, compared it with our current favorites and found out where it earns a place in our daily workflows.

MB
Matthew Berman & the Forward Future teamHands-on tests, with the original recordings
↓
↳

“The best model I have ever used, by far.” Matthew spent more than 100 billion tokens testing it. The builds below are working applications, and some are already part of our production workflow.

THE SHORT ANSWER

Software is solved.
If you have the tokens.

Fable takes massive, complex, underspecified assignments and figures out how to execute them. The apps still have bugs, but knowing how to code is no longer what stops you from building one.

Best use
Ambitious, underspecified builds
Main tradeoff
Cost: up to ~$15,000 per build
Who should care
Developers, game makers, designers
Worth switching?
Yes, for building things
02 / WATCH THE REVIEW

The video, chapter by chapter

Press any timestamp on this page
to play the video from that point.

CHAPTERS01 / 24
MORE FROM THE VIDEO

Demos from the video outside the five field tests, including the one that didn’t work.

Matthew

Bullet vs. water balloon

41M tokens · ~2.5 hrs · $91

The balloon skin and bursting water are ray-traced, and sunlight passing through the rubber tints the shadow purple. It’s one HTML file using WebGL2, with no frameworks or libraries. You can freeze time, orbit the scene and change the bullet caliber, speed and lighting while it runs.

The prompt +
Make a ray tracing simulation of a bullet piercing through a water balloon. Make the balloon pop and the water burst, deforming from its original shape and eventually falling to the ground. Make it look as realistic and physically accurate as possible. Allow me to freeze the animation at any point in time and view the scene from any angle. Add sliders for bullet speed, balloon size, lighting, gravity, and other parameters. Most important, do not use 3JS or any external web libraries. I want to test your physics understanding. So, code this completely from scratch.
Try the simulation ↗
Matthew

A house in Blender

13M tokens · $30

The house has a pool, a terrace, a garage and solar panels, and the final renders show the pool lighting, reflections, glass and shadows. Matthew spotted one mistake: a tree in the middle of the driveway.

The prompt +
Build a 3D render of a beautiful house in Blender. Make it so it feels like a real estate developer's dream.
Matthew

Tokyo in Unreal Engine

6.72B tokens · ~70 hrs · ~$11,000

Every building was made from scratch. The city has traffic, pedestrians, arriving trains, the famous tower and ray-traced reflections on the buildings.

Matthew

A European town

An old-world town with futuristic cars, cows, sheep and a dog wagging its tail. AI drives what the townspeople do and say, and the model also made a video demo of the whole thing.

MatthewThe first bad one

Tesla Roadster event

Recreated from pictures and videos of the real event. It didn’t land.

Alex & Brian

Call of Duty, block style

A blocky multiplayer shooter. Matthew’s verdict: “it just works.”

Alex & Brian

Melee on Android

Decompiled and ported to an Android phone, so Melee is playable anywhere.

Alex

Portal

Portal-inspired games for desktop and mobile. The graphics leave a little to be desired.

Alex & Brian

Starlane

An F-Zero-style racer with strong graphics and AI-generated narration.

Brian

Tidekeeper

The second version of a crab game Brian first made with a previous-generation model. The crab is a lot of fun to control.

Alex & Brian

Dragon Rider

A dragon-flying game with impressive graphics, seen for the first time on camera.

03 / EXPLORE THE TESTS

The five field tests

Each test has the prompts we used,
what got built and what it cost.

FIELD TEST 01 / SOFTWAREEXCELLENT

Can Fable replace software I’m already paying for?

Build a fully functional spreadsheet in the browser with the features people actually use: formulas, multiple sheets, data management and pivot tables. Meanwhile, Alex rebuilt Photoshop and Premiere, customized to the way he already works.

WHAT HAPPENED

Fable produced the familiar ribbon (Home, Insert, Draw, Page Layout, Formulas, Data, Review, View, Automate and Help), a sample sales tracker and a working pivot table. Alex’s Safelight has already made thumbnails for the channel, and his Claude-powered Premiere alternative let him cancel his Adobe subscription.

Excel cloneBy Matthew

Every cell is a real formula. Example sales tracker, extra sheets and a fully functional pivot table. About 50 hours to build.

Try the spreadsheet ↗
Kestrel, a DaVinci Resolve-style editorBy Matthew

Runs locally on a Windows PC: timeline, waveforms, clip splitting, magnetic snapping, keyboard shortcuts, zooms and keyframes.

Build time~50 hrsExcel clone
Tokens~8BCached input + 15M output
Est. API cost$11,700At API pricing
OutcomeExcellentScorecard rating
THE HUMAN READ

I really believe we’re heading toward an era of “software for one.”

You need an application, you describe it and AI builds it around your workflow.
The assignment & prompt +

Matthew’s Excel prompt

Build a full Excel clone in the browser. It should have full feature parity. Read the spec docs to figure out all the features.

Steal this workflow · For personalized software, take Alex’s approach: tell the model how you use the application, which features matter most and which controls and shortcuts you want to preserve.

Caveats & verification +
  • I didn’t verify every feature in Microsoft Excel, but the clone behaved exactly as I expect a spreadsheet to.
  • Recreating the functionality you personally need is very different from replacing every feature, integration and reliability guarantee of a mature professional application.
04 / THE FRICTION REPORT

Where it still
needs you.

Problems we hit while testing, how we
noticed them and what to do about each.

What went wrong

The Gears clone’s reach-over-the-barrier animation looked awkward, and the difficulty crushed Matthew.

How we caught it

Playing the level live on camera.

What to do about it

A polish and tuning pass. Everything else in the level held up.

What went wrong

LEGO Iron Man shipped with confusing controls, audio issues and a few bugs.

How we caught it

Alex playtesting his own build.

What to do about it

Still needs bug fixing before it feels finished.

What went wrong

The Tesla Roadster event recreation simply wasn’t good.

How we caught it

Watching it for the first time.

What to do about it

Recreating a real event from photos and video isn’t a sure thing.

What went wrong

The Blender house put a tree in the middle of the driveway.

How we caught it

Orbiting the finished scene.

What to do about it

Review the layout before the final render.

What went wrong

The AirPods app shows controls the OS may not let it execute.

How we caught it

Reviewing what the demo did and didn’t prove.

What to do about it

Test across repeated connections before calling it complete.

What went wrong

The Gears clone burned 8.5 billion tokens, about $15,000 at API rates, and Ultra Code is extremely expensive to run.

How we caught it

The token and cost readouts after each build.

What to do about it

Pick projects carefully. Matthew doesn’t recommend Ultra Code, and some builds cost under $100.

Across the builds we also saw smaller bugs, performance issues and features that need more testing.

05 / THE PRACTICAL DECISION

Does it earn a place?

Our rating for each kind of work,
with the evidence behind it.

CapabilityOur assessmentThe evidence
Complex software developmentExcellentExcel, DaVinci Resolve and Adobe alternatives with working features, not just convincing interfaces
Game developmentExcellent, a few rough edgesPlayable 3D shooters and multiplayer games with physics, enemy AI and multiple levels
Autonomous executionStandout strengthHandled massive, underspecified assignments with relatively little intervention
Creative productionExcellentProfessional-looking motion graphics, customized editing tools and a complete animated explainer
Physical designPromisingA 3D-printed tower built to real manufacturing constraints and tested under load
Efficiency and costHighly variableAbout $25 for a motion-graphics explainer versus $15,000 for the Gears of War project
Polish and reliabilityNeeds workAwkward animations, occasional bugs, performance issues and features requiring further validation
Strongly recommended

Developers and builders

It takes a short prompt and builds sophisticated applications, games and interactive environments. You’ll still test, debug and validate.

Worth exploring

Knowledge workers and operators

Need a specialized dashboard, a custom integration or a tool that doesn’t exist? You don’t have to wait for a software company.

Not on these tests alone

Writers, researchers and analysts

We only tested building things. Research accuracy, writing quality and analytical reasoning weren’t part of these tests.

Absolutely try it

AI power users

Even if you prefer another model for everyday chat, its ability to execute complicated projects earns it a place in your toolkit.

OUR FAVORITES
Best time-saverMotion graphics

A complete 1:52 explainer in about 45 minutes for an estimated $25.

Best quality improvementGame development

Physics, level complexity and multiplayer were dramatically better.

Best new capabilityPersonalized software

Alex already uses apps he built with Fable in our production workflow.

FABLE 5.2 OR OPUS 5.5?

“Without a doubt, Fable 5.2 is better than Opus 5.5.”

But Opus 5.5 is very efficient, very fast and much less expensive. Some of our projects needed billions of tokens and others cost less than $100, so pick your projects carefully.
06 / TAKE IT FOR A SPIN

Run the tests
yourself

The prompt kit has the exact prompts we used, with the build stats and caveats for each test. The links open the builds we published.

JSON · 5 field tests · 9 prompts · Video timestamps
01

Excel cloneThe spreadsheet from Field Test 01

TRY IT ↗02

Bullet vs. water balloonThe from-scratch WebGL2 simulation

TRY IT ↗03

Software decompiling explainerThe 1:52 motion-graphics video

TRY IT ↗04

Photon image editorLinked from the video description

TRY IT ↗
07 / BEHIND THE REVIEW

How we tested

Our testing approach

We built real projects from short, often underspecified prompts, tried every thinking effort and used Ultra Code mode for the biggest ones. Costs are estimates at API pricing. Each build is a single case study, and we only tested building things, so research and writing quality went untested.

Claims vs. observations

The benchmark scores come from the model providers, as shown in the video, and aren’t from our own tests. Pricing and access are as announced on November 9, 2026.

Extremely autonomous

It made decisions on my behalf, and nine times out of ten, maybe more, they were what I would have done.

Fast to finish

Tasks I’d expect to take 60, 90 or 120 minutes take about 10 minutes at most. It seems to know the straight-line path to done.

Ultra Code finds real issues

Max thinking effort plus Ultra Code across every codebase found bugs, security issues and UI inconsistencies, and finished quickly. It’s also extremely expensive; I don’t recommend it.

Fable or Opus 5.5?

Fable 5.2 is better, without a doubt. Opus 5.5 is still very efficient, very fast and much less expensive.

Benchmarks, pricing & access +
Frontier Code 2.0
  • Fable 5.2Leads

Fable 5.2 dominates every other model on the list, including Opus 5.5.

Frontier SWE
  • Fable 5.269.9

The top score on the list.

Automation Bench
  • Carbon53
  • Fable 5.251

The only benchmark Carbon wins.

Terminal Bench Science
  • Mythos 5.275.3
  • Fable 5.274.6

Mythos 5.2, also released today, barely edges ahead.

OSWorld 2.1
  • Fable 5.289.3
  • Opus 5.587
  • Astra79.5

Measures how well a model controls an operating system.

Humanity’s Last Exam
  • Fable 5.274.2

10 points above Fable 5.1 and just under 10 above Opus 5.5.

PricingInput / 1M tokensOutput / 1M tokensNote
Standard$8$40About 20% cheaper than Fable 5.1
Fast mode$16$802.5× the speed

Access · Anthropic is giving out Fable usage resets this week, depending on whether your plan includes Fable and how long you’ve been subscribed.

Review history +

November 9, 2026: Published the five field tests, the overview video, the scorecard, provider benchmarks and pricing.

October 20, 2026: Created the page with planned tests and no results.

WATCH

The overview video

Every demo from the field tests, with Matthew’s reactions.

Watch on YouTube ↗
READ

The full story

What we learned from putting Fable 5.2 to work.

Read the field tests ↗