Comparisons

GPT-6 Astra vs Fable 5.1: Four Builds, One Cost Surprise

Direct Answer

GPT-6 Astra did not sweep Claude Fable 5.1 in Pat Simmons's four-build comparison. The test ended in a 2-2 tie, but Astra's reported completion cost made the result more interesting than the score. Fable produced the preferred kart racer and 3D dive-watch page. Astra produced the preferred physics simulator and website recreation. GPT-5.6 Sol generally trailed both on visual quality.

The surprise was total task economics. Pat says Astra and Fable carried the same listed token price in his comparison, yet Astra's recorded runs were often much cheaper because it completed tasks with fewer or shorter agent turns. On the Theo Jansen simulator, he reported about $6.37 for Astra versus $84 for Fable. On the site-recreation task, the figures were $71 and $223.

The practical conclusion: compare the cost of a result your team would actually accept, including retries, elapsed time, repair work, and verification.

Watch the Four-Build Comparison

Credits and disclosure: the builds, rankings, elapsed times, and task-cost figures come from Pat Simmons's AI for Mortals video, published on 4 September 2026. The hidden model columns were revealed after ranking. This is a creator-run qualitative comparison, not a repeated independent benchmark.

How the Test Worked

Pat sent four prompts to GPT-6 Astra, Claude Fable 5.1, and GPT-5.6 Sol. Six terminal windows ran the builds in parallel, and he reviewed the applications without initially knowing which model created each column. The tasks were visual and interactive: a kart racer, a mathematically constrained walking machine, a scroll-driven product page, and a recreation of a site selected from Awwwards.

That blind reveal reduces brand bias, but it does not remove every variable. Each task appears to have one output per model, the reviewer knew the three candidates, and the criteria were not converted into a scored rubric before inspection. Agent duration, tool use, and generated volume also differed.

What the test showsWhat it does not proveWhy it matters
First-pass visual and interactive qualityRepeatability across multiple runsUseful for discovering tendencies
Creator-observed task totalsA universal cost multiplierHighlights execution efficiency
Relative performance on four promptsOverall model intelligenceKeeps the verdict task-specific
Hidden identities during rankingA fully blinded independent studyReduces reviewer bias

The Four-Build Scorecard

BuildPat's winnerWhy it wonCaveat
3D kart racerFable 5.1Better polish, environment, and play feelOne game per model
Theo Jansen walkerGPT-6 AstraCleaner controls and stronger 3D presentationPhysics not independently checked
3D dive-watch pageFable 5.1Smoother choreography and richer detailVisual judgment
Site recreationGPT-6 AstraClosest recreation and strongest workflow resultModels chose different targets

Build 1: Fable Won the Kart Racer

The first prompt asked for a polished 3D kart-racing game with responsive controls, items, opponents, barriers, and a credible track. Pat preferred Fable's result. Astra's build was playable but less polished, while Sol's version suffered from confusing controls, weak graphics, and missing environmental boundaries.

Coding strength does not automatically create art direction or game feel. A production review should test collision behavior, frame rate, input latency, opponent logic, restart states, accessibility, and whether a new player understands the controls.

Build 2: Astra Won the Walking Machine

The second build modeled a Theo Jansen-style strandbeest, a linkage system whose proportions determine whether its legs can walk. Astra earned Pat's top rank for its 3D treatment and cleaner controls. Fable offered more controls, but the interface felt busier.

Pat openly says he cannot certify the mechanism's mathematics. A convincing animation can still be mechanically wrong.

For technical simulations: separate visual quality from domain correctness. Validate equations, constraints, units, and edge cases before using the result for teaching or engineering decisions.

Build 3: Fable Won the Dive-Watch Page

The third prompt requested a complete scrolling product page centered on a real-time 3D dive watch. The watch needed to move from wireframe to solid, explode into components, display annotations, and reassemble as the visitor scrolled.

Fable's page was the clearest win. It combined smooth transitions, convincing lighting, detailed components, readable annotations, and a longer product narrative. Astra had attractive moments, but its watch and transitions did not reach the same level. Sol's output was less realistic and less functional.

High-fidelity product storytelling combines geometry, lighting, motion, typography, and pacing. A technically complete page can still lose when those systems do not feel like one art-directed experience.

Build 4: Astra Won the Site Recreation

For the final task, each model selected a site from Awwwards and attempted to reproduce it with Pat's public site-recreation skill. The workflow asks agents to inspect views, gather screenshots and styles, build the interface, and check their work.

Astra produced the preferred result. Pat found its target selection distinctive and its recreation impressively close in layout, interactions, transitions, and overall character. Fable also delivered a strong page. Sol failed to reproduce its selected target convincingly.

This tests more than visual generation. The agent must browse, extract evidence, follow a staged method, maintain context, write code, and compare its output with a reference. Astra's win supports a narrow but useful claim: it handled this long, tool-using workflow especially well.

Website recreation also has a rights boundary. Use it for authorized redesigns, migrations, internal prototypes, and implementation study. Replace third-party trademarks, photography, copy, and proprietary assets before publishing unless you have permission.

The Cost Surprise: Sticker Price Is Not Task Price

Matching per-token prices did not produce matching bills in Pat's runs:

TaskFable 5.1GPT-6 AstraInterpretation
Kart racerAbout $55About $10Fable won quality; Astra spent less
Walking machineAbout $84About $6.37Astra won quality and cost
Site recreationAbout $223About $71Astra won with a lower total

These figures belong to Pat's setup. Total cost changes with the harness, context caching, tool calls, retry policy, reasoning setting, provider billing, and allowed run time. The video does not establish a permanent 5x to 15x advantage across every workload.

It does establish the right unit of analysis. A cheap model can become expensive through loops, retries, and repair. A higher-priced model can be economical if it reaches acceptance quickly. Track cost per accepted result, not price per million tokens alone.

A Better Model Test for Your Team

  1. Choose real tasks. Select completed jobs that represent your actual coding, design, research, and tool use.
  2. Freeze the environment. Match the prompt, files, tools, timeout, reasoning level, and stopping rule.
  3. Run each condition three times. One generation can be unusually good or bad.
  4. Review without model names. Score functionality before visual preference.
  5. Use a written rubric. Include correctness, visual quality, accessibility, security, maintainability, and instruction following.
  6. Verify the artifact. Run tests, inspect the console, exercise every control, and compare against the brief.
  7. Calculate accepted-result cost. Add model spend, failed runs, wall time, and human repair minutes.

The output should be a routing policy, not a single winner. One model may be your default for interactive design, another for fast implementation, and a cheaper model for drafts. Re-run the test when versions, prices, or your workflow change.

Video Chapters

TimeTopicTimeTopic
00:00Introduction04:36Build 1: 3D kart racer
00:43Kicking off the builds09:37Build 2: Theo Jansen walking machine
00:55The benchmarks15:11Build 3: 3D dive-watch page
02:36Coding benchmarks19:59Build 4: recreating an Awwwards winner
02:56Knowledge-work benchmarks26:07Final reveal and cost
26:47The verdict27:22Has OpenAI caught Anthropic?
28:13What comes next

Verdict

Pat's test supports a measured update: GPT-6 Astra is competitive with Claude Fable 5.1 on ambitious visual coding work. It won two of four builds, including the workflow that required the broadest combination of browsing, evidence extraction, code generation, and visual comparison. Fable remained stronger on the kart racer and polished product page.

The larger shift is economic. In these runs, Astra often reached a competitive result with much less recorded spend. That makes agent efficiency a first-class model characteristic, alongside intelligence, speed, and style.

OpenAI may have closed the visible capability gap in this test, but a 2-2 creator comparison is not a permanent crown. Reopen your own evaluation, measure accepted-result cost, and let repeatable evidence decide which model gets each job.

Sources and Links

This article uses the primary video's official YouTube publication date of 4 September 2026 and was researched and published on 6 September 2026. Rankings and cost totals are attributed to the creator's recorded test. Model behavior, access, and pricing can change.

Common questions

Did GPT-6 Astra beat Claude Fable 5.1 in Pat Simmons's test?
The creator-run comparison ended in a 2-2 tie. Fable won the kart racer and dive-watch page. Astra won the walking-machine simulator and site recreation.
Why did Astra cost less if its token prices matched Fable 5.1?
Pat attributes the difference to task efficiency. In his runs, Astra used fewer agent turns or finished sooner. This is a result from his harness, not a universal cost ratio.
Was GPT-5.6 Sol competitive?
Sol was usually the weakest visual result in this test, although it was inexpensive. That may still make it useful for drafts or narrow edits.
Can four prompts prove which model is best?
No. Model outputs vary and the scoring is subjective. A production decision needs repeated runs, hidden model names, fixed limits, and a rubric based on real work.
What is the best way to compare coding agents?
Measure accepted-result cost: model spend, elapsed time, retries, human repair time, and whether the result passes functional, visual, accessibility, and security checks.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call