Since I’ve been using Claude, I always wonder which model to choose and what “effort” level to give it: medium, high, or xhigh? Here’s what I’ve found.
For months, I used the most powerful model available by reflex. Opus by default. Maximum effort. Result: slow, expensive, and often unnecessary for what I was actually doing.
The reality is simpler: the right model isn’t the most powerful — it’s the one that matches the real complexity of your task.
The Hierarchy in June 2026
Anthropic now offers four model tiers, from fastest to most powerful:
Haiku 4.5: The sprinter. Fast, economical, no extended reasoning.
Sonnet 4.6: The workhorse. Quality/speed balance for 80% of tasks.
Opus 4.8: The specialist. Deep reasoning, long contexts, complex problems.
Fable 5: The new peak, launched June 9, 2026. A class above Opus, with built-in guardrails for certain sensitive domains.
The Five-Question Test
Before choosing a model, ask yourself these five questions. One “yes” is enough to move up a level.
1. Do I need to analyze more than 50,000 words in a single pass? No → Haiku or Sonnet is enough. Yes → Opus or Fable 5.
2. Does the task require comparing, synthesizing, or reasoning across multiple documents at once? No → Sonnet. Yes → Opus.
3. Will I repeat this operation dozens or hundreds of times per day? Yes → Haiku. Speed and cost matter more than depth.
4. Would an error in the response have real consequences (business decision, legal analysis, production code)? Yes → Opus minimum.
5. Am I working on a problem Sonnet or Opus didn’t solve correctly after two attempts? Yes → Fable 5.
Concrete Examples by Task
Summarize a 200-word email → Haiku 4.5
Draft a response to a request for proposal → Sonnet 4.6
Analyze 10 distinct data files → Sonnet 4.6 (high effort)
Review an 80-page contract → Opus 4.8
Design a system architecture → Opus 4.8
Debug a complex multi-component problem → Opus 4.8 or Fable 5
Write 500 similar product descriptions → Haiku 4.5
Produce a risk analysis with contradictory sources → Fable 5
The Case of 10 Data Files
Frequent question: if I have 10 files to analyze, what do I use?
It depends on what you’re doing with them.
Extracting numbers or normalizing data? Haiku or Sonnet, medium effort.
Looking for patterns, anomalies, or correlations between files? Sonnet, high effort.
Need to synthesize the files into a strategic recommendation with explicit reasoning? Opus, high effort.
Complexity isn’t in the volume. It’s in the level of reasoning required.
The Effort Parameter: What It Actually Changes
Choosing the model is one decision. Choosing the effort level is another, distinct one.
Effort controls the depth of internal reasoning before Claude produces a response. Concretely: the higher the effort, the more Claude “thinks” before answering — which costs more time and processing tokens.
In the claude.ai interface
Low: Fast response, minimal reasoning. For factual questions, rewording, simple repetitive tasks.
Medium: The recommended balance for most uses. Writing, standard analyses, routine technical questions.
High: Claude systematically takes time to reason before answering. For multi-step problems, analyses requiring nuance, consequential decisions.
xhigh / Max: Reserved for difficult problems where high isn’t enough. Complex debugging, system architecture, advanced scientific or legal reasoning.
Ultra Code (Opus 4.8, API only): Specifically optimized for high-complexity software development tasks.
In the API
The effort parameter offers finer control. Anthropic documentation explicitly recommends medium as the default for Sonnet 4.6: it avoids unnecessary latency caused by high as the default.
Practical rule: reducing effort by one level on simple tasks can reduce latency and cost by 30 to 50% with no perceptible difference in response quality.
Note on Fable 5
Fable 5 is the first public instance of a new model class at Anthropic, called Mythos. Launched June 9, 2026, it exceeds Opus in reasoning, vision, and autonomous work on long tasks.
I had the chance to test it a few days before it was withdrawn. General impression: the difference with Opus is perceptible on long, ambiguous tasks. Where Opus sometimes needs follow-ups to maintain the thread of complex reasoning, Fable stayed on course across several steps without intervention. On routine tasks (writing, standard analysis, technical questions), the difference was marginal. That confirms what the grid says: power only justifies the cost if the task truly demands it.
It is currently unavailable. On June 12, Anthropic withdrew it for all users following a U.S. export control directive. No return date has been announced as I write these lines.
When it returns, it will likely be accessible via a pay-per-use credit system (token billing) rather than included in standard subscriptions. Anthropic indicated it wants to reintegrate it into subscriptions once capacity and regulatory constraints allow.
In the meantime: Opus 4.8 remains the reference model for the most demanding tasks. Fable 5 itself was designed to automatically fall back to Opus 4.8 on refusal, which confirms Opus is a solid base for the vast majority of real use cases.
The Decision Grid in Summary
Repetitive task, high volume, little nuance required → Haiku 4.5 / low or medium effort
Writing, standard analysis, routine technical questions → Sonnet 4.6 / medium effort
Multi-document analysis, consequential decisions, nuance required → Sonnet 4.6 high effort OR Opus 4.8 medium effort
Complex problems, architecture, deep reasoning → Opus 4.8 / high or xhigh effort
Extreme cases, long autonomous tasks, problems others haven’t solved → Fable 5 (currently unavailable, see note above) → In the meantime: Opus 4.8 / xhigh or max effort
The reflex to always take the most powerful model is understandable. But it’s a bit like using a chainsaw to cut bread. The right tool for the right task remains the most effective rule — even in artificial intelligence.
And you? Do you use other criteria than these to choose between Haiku, Sonnet, and Opus? A signal in response quality, an empirical rule you’ve developed, a task where your intuition proved wrong? I’d be curious to read your experiences in the comments.