When a model claims to be "state-of-the-art" across every benchmark, skepticism is the natural first reaction. But after anthropic released Claude Fable 5 and the author put it through a real coding challenge—analyzing a live project called WeSight and then building a functional Obsidian plugin from scratch—the results were undeniably impressive. However, the experience also laid bare a critical trade-off that developers must consider: unmatched reasoning power comes with a staggering token appetite.
The test began with a simple request: feed Fable 5 the entire WeSight GitHub repository and ask it to diagnose bugs, architectural issues, and UI problems. The model returned a detailed analysis that read like a senior developer reviewing a junior’s pull request—covering everything from functional errors to interaction design gaps. According to Anthropic’s published benchmarks, Fable 5 achieves 57.1% on SWE-bench V2 (compared to GPT-4o’s 31.1% and Gemini 2.5 Pro’s 41.6%), and 75.5% on TAU-bench, surpassing all prior models. The author’s qualitative experience matched this: the analysis was thorough, structured, and actionable. Yet within five hours of fixing a handful of identified bugs, the author had already exhausted one Claude account’s quota. The model’s ability to think deeply is directly proportional to the tokens it consumes—and for simple tasks, that depth becomes an expensive luxury.
Switching to a second account, the author decided to push Fable 5 to its limit: design and develop a full Obsidian plugin that leverages WeSight’s underlying engine. The prompt included the entire codebase and the requirement to integrate multiple CLI tools—Claude Code, Codex, OpenCode—into a single Obsidian interface. Within two hours, Fable 5 produced a working plugin that automatically detects local CLI environments, allows custom third‑party model routing, and lets users control Claude Code directly from Obsidian without needing separate tools like Claudian or cc-switch. It’s one thing for AI to generate boilerplate, but building a multi‑tool orchestrator with zero manual debugging is a different league altogether.
This success highlights a key strength of Fable 5: its ability to maintain coherence across extremely long contexts. The model handled thousands of lines of code, a complex specification, and iterative constraints without losing track. Yet the speed remained modest—around 50–70 tokens per second in the author’s tests, far below the 300–400 tokens/s seen in models like glm-5.1-highspeed. Community feedback on platforms like Hacker News and Reddit echoes this: Fable 5 is ideal for intricate, multi‑step agentic tasks but overkill for quick edits or simple refactors. Choosing the right model for the task is becoming as important as choosing the model itself.
Looking beyond this single test, the implication for developer workflows is significant. The plugin essentially turns Obsidian into a command center for AI‑assisted development—a pattern that may soon become standard as knowledge management and coding tools converge. For instance, a developer could query a local codebase through Obsidian, ask Fable 5 to rewrite a module, and have the result synced automatically. Yet the cost structure still limits such use: at roughly $15 per million input tokens and $75 per million output tokens (Pro tier), a single complex interaction can burn through $10–$20. OpenAI’s GPT‑4o and Google’s Gemini 2.5 Pro offer lower pricing, though with inferior reasoning on hard tasks. The value of Fable 5 lies not in speed or price, but in its ability to grasp intricate logic where other models fall apart.
For developers tempted to use Fable 5 for every coding need, the lesson from this hands‑on test is clear: reserve it for complex problem‑solving—like building a custom plugin from an unfamiliar codebase—and rely on faster, cheaper models for routine short tasks. The author plans to publish the plugin to the Obsidian marketplace soon, and it will likely spark debates about whether AI‑driven IDEs will replace traditional ones. The real shift isn’t that AI can write code—it’s that AI can now understand an entire codebase and build a production‑ready tool on top of it, autonomously. The question is whether your budget allows you to keep asking.