Published: 2026-09-05
A fully local agent with tools, in about 40 lines: Ollama plus Pydantic AI
Running Nex-N2.5 Mini on two H100s: the SGLang container setup, and an honest benchmark read
Semantic grep cut agent tool calls 58% and input tokens 47% in the project's own benchmarks
The instruction ceiling moved 10x in a year: 200 rules became 2,000
How LinkedIn made coding agents work on 1,000+ internal repos without fine-tuning
Two context approaches that look right and stall: the curated-context trap and the MCP plateau
GPT-6 Astra vs Fable 5.1: five identical builds, and a looks-versus-works split
Chapters / key moments (click to jump — plays here on the page)
The finding that generalises past this test: Fable 5.1 optimised for looking like the target, and Astra optimised for working like it. On the Notion clone, Fable produced the more convincing replica — better drop-downs, richer cover options, closer to the real workspace — but several of its buttons did nothing. Astra's looked plainer and every button it drew was wired up. Which one is better depends entirely on whether you are building a demo or a tool. Worth noting on speed: Astra finished all five builds in about 30–40 minutes, and Fable took roughly another 30–40 minutes to finish the rest after its first.
Source video
"I Made GPT-6 Astra and Fable 5.1 Build the Same App (RAW RESULTS)" by Bart Slodyczka — Watch on YouTube →
Gotchas & Caveats
- Both models were run at
higheffort through their desktop apps, five chats each in parallel. Results at other effort levels would differ — and on Fable 5.1 specifically, effort level also changes tool-calling behaviour. - The prompts were deliberately specific, which the reviewer notes is why the two CRMs came out structurally similar — a looser prompt would widen the gap in either direction.
- Judgements about which UI is better are the reviewer's, though the functional findings (dead buttons, non-scrolling panel, missing entity) are demonstrated on screen rather than asserted.
- This is a build-quality comparison, not a cost comparison — for recorded cost and time per task see our fifteen-use-case head-to-head.
Key Takeaways
- Every one of the ten sessions produced a working app with no steering. The only intervention across the whole test was hitting a session limit and upgrading the plan. That is the actual headline — five non-trivial apps each, unattended.
- Looks versus works is the consistent split. On the Notion clone Fable was visually closer to the real product but shipped dead buttons; Astra was plainer and functionally complete. The reviewer's own conclusion: a button that does nothing is worse than no button, because it is how a user learns the app is fake.
- On the CRM, Fable won on workflow design. It let you create a new customer from inside the new-job dialog; Astra forced you out to the customers tab and back. It also had a dedicated jobs board Astra lacked. Both are the sort of thing that only shows up when you actually try to complete a task in the app.
- Astra's CRM had two concrete gaps: a settings panel that did not scroll (so content below the fold was unreachable), and no way to add vans or employees — which for a scheduling-based business is a missing entity, not a missing field.
- Speed favoured Astra substantially in this run: all five finished in 30–40 minutes against roughly double that for Fable.
- The clothing-brand build was expected to favour Astra because of built-in image generation, and the storefront it produced included animated brand-appropriate detail.
More OpenClaw & Claude Code news
A fully local agent with tools, in about 40 lines: Ollama plus Pydantic AI
Running Nex-N2.5 Mini on two H100s: the SGLang container setup, and an honest benchmark read
Semantic grep cut agent tool calls 58% and input tokens 47% in the project's own benchmarks
The instruction ceiling moved 10x in a year: 200 rules became 2,000
How LinkedIn made coding agents work on 1,000+ internal repos without fine-tuning
Two context approaches that look right and stall: the curated-context trap and the MCP plateau