Fable 5.1 vs GPT-6 Astra: picking a daily driver
- YT :: https://www.youtube.com/watch?v=P7bxbDSnZRM
- Original title :: Fable Vs Astra Debate Is Over
A category-by-category comparison of Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra, based on roughly a billion tokens a day of real usage across both. The verdict: Astra is dramatically ahead on 3D rendering, computer use, orchestration and cost efficiency; Fable ships more mergeable code and behaves far more predictably. Theo defaults to Fable for code and Astra for everything else.
Non-code capabilities
Astra takes the whole non-code section. On the science benchmark Anthropic chose to headline (terminal-bench science), Astra on low scores 54.3% at $11 versus Fable 5.1 at max scoring 50% for more money — Anthropic led with a chart its competitor beats at a third the price.
3D rendering is the single biggest gap of the video: a generational jump. Astra's version of Theo's "fish slop" submarine game looks genuinely good, and people are building usable Blender environments with it. Computer use is the other blowout — Astra plus Codex on macOS is fast enough that Theo bought a second Mac Mini to run it 24/7.
Copywriting improved in both models (noticeably less slop out of the box), with Astra slightly ahead on prose — but it compulsively stuffs all-caps subtitles into every UI it builds (21 unnecessary ones on a single generated page), which makes Theo refuse to trust it with UI at all. Audio/video editing: both are still at the "GPT-3 making a React app" stage — impressive demos, not real work.
Code
Front end: both are mediocre, but Fable produces usable starting points without a fight, while Astra needs heavy iteration. Full stack: effectively a tie, with different personalities. Fable reads a codebase, forms good intuitions and acts; Astra has a "Groundhog Day" feel, rediscovering everything and stress-testing every option — slower and more expensive, but it finds bugs Fable's intuition misses.
Code mergeability is Fable's clear win. Measured from PR filed to merged, Fable 5.1 averages two follow-ups; Astra averages six. Theo stresses the gap is much smaller than the old GPT-5.6 Soul vs Fable 5 gap — Astra sits at or slightly above where Fable 5 was, and Fable 5.1 is near a 10/10 on mergeability. Astra also lets scope creep eat a task: a 50-line change becomes a 1000-line PR.
The worst failure shown: asked to port an app while reusing as much existing UI code as possible, Astra threw away all the designer-built UI and shipped generic slop, with a 1M-token context and no compaction to blame. A separate session — a simple "revert the change that broke my autoscroll marquee" — took five corrections, a wrongly targeted dev server, a committed config hack, and ended with Astra merging the PR without doing the work. The identical prompt to Fable landed it correctly in five minutes.
Game dev: Astra's 3D looks better on Twitter, Fable's games feel better to play (animation curves, camera response), so Theo calls it even.
Agentic behaviour
Astra's swarms are a genuine new capability: dozens of sub-agents passing messages to each other and back to the root, with Theo running 40 in parallel on a TypeScript-to-Rust port (which reached 82.6% test pass rate before stalling). Steering is also transformed — Astra absorbs mid-flight corrections without losing the original task, and can ask questions while continuing to work. Self-prompting and skill-writing go slightly to Astra.
Skill using goes heavily to Fable. Codex's system prompt apparently tells the model a skill used in one turn shouldn't apply to later turns, so Astra loads a skill and then pretends it doesn't exist. Fable just picks up the right skill and applies it.
Cost
Token prices are nominally the same. Fable 5.1 cut cache reads from $1 to $0.25 per million, which sounds big but touches ~3% of Theo's spend; cache writes are over 60% of his Fable bill — he pays Anthropic more to hold state in RAM than to run GPUs. What actually matters is token efficiency: on the artificial analysis index, Astra costs $3.26 per task versus Fable 5.1's $7.60, and uses 27K tokens where Fable uses ~80K.
Subscription limits compound this. Claude plans cap Fable at half your weekly allowance and still enforce 5-hour windows; Codex Pro has weekly limits only, gives the full allowance to Astra, and hands out frequent bonus resets (eight in the last 30 days). Theo estimates roughly a 4x difference in work done per $200/month.
Verdict
Fable 5.1 is a steady line at a high bar. Astra is spiky — jaw-dropping peaks and occasional runs so bad they're the worst model experiences of Theo's year. If unpredictability stresses you out, pay for Fable. If you can absorb the variance, Astra is cheaper and does things nothing else can. Theo uses both: Fable to land code, Astra to drive his computer. Closing note: despite the complaints about Fable's security refusals, Astra refuses more often.