Stop Picking The Wrong Model

Summary

The video compares Claude Opus 5.5 and Fable 5.2 for coding work, arguing that they are the two models most worth considering for the creator’s day-to-day use. The central recommendation is not simply “use the smartest model,” but to choose based on whether a task benefits more from careful verification or creative judgment and willingness to ship.

Model-selection guide

• Opus 5.5: The default for most engineering work. The creator describes it as meticulous, persistent, and unlikely to overlook details. It is especially useful when requirements are clear, correctness can be tested, or missing an edge case would be costly.
• Fable 5.2: Try it when Opus is stuck, when you want more varied or unconventional approaches, or when a task calls for exploration and subjective judgment. The creator also finds it useful for reviewing work that Opus has repeatedly second-guessed: Fable may conclude that the work is already good enough to merge.
• Sonnet 5.5: Anthropic’s guidance, as discussed in the video, positions Sonnet for well-scoped fixes and features. The creator is less enthusiastic, arguing that it is not particularly good value in their experience, and suggests using it selectively rather than as the default in a model picker.
• Haiku 5.5: Best suited to fast, high-volume tasks such as classification, routing, extraction, and scanning many files for a recognizable pattern. The creator advises against using it to write code.
• Effort levels: Start Opus at medium and increase to high if it misses things. The creator personally favors high for many tasks, uses extra high for particularly complex audits, testing, or large investigations, and advises against max, which can encourage needless overthinking. Anthropic’s guidance, as presented in the video, also recommends comparing a higher effort level with a more capable model at lower effort before committing to the extra cost.

How the creator distinguishes Opus and Fable

The creator characterizes Opus as the careful “teacher’s pet”: it follows instructions closely, checks its work, and is reluctant to leave possible issues unaddressed. That makes it strong for thorough implementation and risk-sensitive tasks, but can lead to feature creep, excessive nitpicking, or prolonged cycles of review. The creator suggests prompting Opus to offer an alternative when it thinks the requested approach is not a good idea.

Fable is portrayed as more intuitive and willing to challenge assumptions. It may overlook minor issues or make a less conventional choice, but it can find clever solutions that Opus misses and may be better at deciding when something is ready to ship.

The creator’s shorthand is:

• Verifiable work: Favor Opus when tests, specifications, or other objective checks can establish correctness.
• Open-ended work: Consider Fable when creativity, exploration, or judgment matter more than a single provable answer.

The creator stresses that these are tendencies, not guarantees. They describe cases where Opus outperformed Fable and recommend switching models when the first choice is not producing useful results.

Examples and practical findings

• In a performance investigation involving long-running threads, Opus built a useful test environment but missed a costly history-handling issue. Fable identified that an unresolved task list could cause the entire conversation history to be included, and proposed separating the task state from that history. The creator says this reduced data sent in certain edge cases by about 150×.
• The creator also found Fable useful for reviewing stalled work and judging whether it was ready to merge, while Opus sometimes continued proposing additional changes.
• On speed, the creator reports that Opus 5.5 and Fable 5.2 had similar observed output rates, around 100–110 tokens per second in their measurements. They cite higher rates for Sonnet and Haiku, with Haiku the fastest of the three. These are the creator’s observations and referenced benchmark figures, not universal guarantees.
• The video discusses subscription economics as well as API pricing. The creator argues that Claude subscriptions can provide substantially more usable work per dollar than paying API rates or using other model subscriptions. They also note that, under the plan described, Fable usage is capped at half of the weekly allowance, while the remaining allowance can be used for Opus or other eligible models. These cost comparisons depend on the plan and usage assumptions discussed in the video.

Product and tool mention

The sponsored segment describes Browserbase as browser infrastructure and APIs that let agents search the web, interact with sites, and access data from URLs. It also highlights Stagehand, an open-source SDK intended to make browser use by LLMs faster and more reliable.

Guides and analysis covered

1. Choosing a model: Start with Opus for general engineering; switch to Fable for open-ended exploration, alternative ideas, or a fresh assessment of stalled work.
2. Tuning effort: Begin at medium, increase when needed, and avoid assuming that maximum effort is always better.
3. Choosing lightweight models: Use Haiku for bulk analysis and simple information-processing tasks; reserve Sonnet for appropriately scoped work if it fits the task.
4. Evaluating costs and speed: Consider subscription limits and task-level results, not just API prices or model rankings.

Main speakers and sources

• Theo, the video’s presenter, sharing personal model comparisons, coding examples, and benchmark observations.
• Addy Osmani, via an Anthropic model-selection article discussed in the video.
• Artificial Analysis, cited for model-performance and speed comparisons.
• Browserbase, the sponsor featured in the browser-agent segment.