I ran a detailed comparison between Fable 5.1, Astra, and SWE-2 in Devin to map out their strengths and weaknesses: SWE-2 – Strengths: It explores territory other models never touch, looking at problems with a much wider lens. It picks up on subtle cues with very little context. You can trust it with minimal prompting, like telling it to delete something without worrying it might wipe out unrelated code. It has great discipline. When evaluating a problem, it targets the main pain point right away instead of wasting time on trivia. Its "what not to do" advice is genuinely useful and keeps you from spinning your wheels in solved areas. Unlike Astra, it avoids arbitrary, nonsensical refusals. It feels much more honest about what it can and cannot do. – Weaknesses: It lags far behind its base model, Kimi K3, in UI sense, aesthetics, and design taste. It also tends to stumble on a great idea and then hesitate to actually ship it. Astra – Strengths: Pure tenacity once it identifies a bug. It simply doesn't stop until it solves the problem. In one test, Opus 5 gave up and told me to file a support ticket, while Astra kept digging until it fixed the issue completely. It shows real engineering creativity, picking tight research paths with concise, structurally sound solutions and solid recommendations. It also has a great sense of success probability and offers precise statistical qualification for tasks. – Weaknesses: It tends to plan around small sample sizes and thin evidence. It gets bogged down in minor details instead of looking at the big picture and overall impact. It also struggles to synthesize ideas across different domains, overspecializing until the scope becomes too narrow. Fable 5.1 – Strengths: The highest evidence accuracy by far. It reliably delivers on what it promises, unlike models that promise everything and underdeliver. It has strong design taste, clear recommendations, and a holistic grasp of the end goal. It weighs priorities accurately. It can also pivot your entire strategy convincingly without drifting off into irrelevant territory. – Weaknesses: Its answers are dense and exhausting to read. Once it settles on an idea, it's hard to steer it away. Its research skims the surface rather than digging up rare findings. It can suffer from tunnel vision, sticking strictly to technical aspects while ignoring costs or viability metrics unless prompted. Patterns and Behaviors – Negation: Fable and SWE-2 set broad boundaries. Astra uses rigid negative rules that abruptly flip decisions when triggered. – Stability: Fable and Astra are steady. SWE-2 can be inconsistent, sometimes generating an idea only to backtrack on it. – Length: Astra is the most concise. Fable is the longest. SWE-2 sits in the middle, though it has repetitive filler that needs trimming. – Main Risks: SWE-2 over-explores possibilities at the expense of verifying existing code. Astra loses the forest for the trees. Fable turns practical technical decisions into academic literature reviews. – Standout Traits: SWE-2 wins on thoroughness, Astra wins on grit, and Fable wins on raw reasoning and persuasion. Scores and Recommendations – Scores: Fable 5.1: 8.5/10. Astra: 7.5/10. SWE-2: 7/10. – Pick Fable or SWE-2 if you want strict prompt adherence, clean execution, or deep mechanical logic. – Pick Astra if you need precise verification evidence and relentless debugging. – Scope and context: SWE-2 and Fable are tied, while Astra falls slightly behind. This is an initial assessment based on tests in Devin, and I'll need more runs to map out the nuances completely.