Claude Fable 5 is Anthropic's highest-capability widely released model, but that does not make it the automatic choice for every request. The useful question is: which model meets your quality target with acceptable speed and cost?
Anthropic currently positions Fable 5 for the hardest long-running agent work, Opus 5 for complex agentic coding and enterprise tasks, Sonnet 5 as the strongest speed–intelligence balance, and Haiku 4.5 as the fastest economical option. The comparison below reflects official documentation reviewed on August 22, 2026; models and prices can change.
Claude model comparison at a glance
| Model | Position | Best fit | API input / output per MTok | Context | Relative speed |
|---|---|---|---|---|---|
| Fable 5 | Highest capability | Long-running agents and the hardest multi-step work | $10 / $50 | 1M | Slower |
| Opus 5 | Powerful specialist | Complex agentic coding and enterprise workflows | $5 / $25 | 1M | Moderate |
| Sonnet 5 | Best general balance | Most production tasks that need speed and strong intelligence | $2 / $10 | 1M | Fast |
| Haiku 4.5 | Fastest economy option | High-volume, simpler, latency-sensitive work | $1 / $5 | 200k | Fastest |
Prices shown are standard base API input/output rates in USD per million tokens from Anthropic's documentation at the review date. Features, regional availability, batch pricing, caching, and long-context pricing may differ.
Choose Claude Fable 5 for the hardest sustained work
Fable 5 is designed for next-generation intelligence in long-running agents. Choose it when the work requires extended planning, difficult reasoning across many steps, tool use over a long horizon, or recovery from ambiguous situations—and when the quality gain is worth the highest price and slower response.
✓Long-running agents with many dependent steps.
✓The hardest coding or research tasks after cheaper models fall short.
✓High-value work where failure costs more than extra tokens.
✓Evaluations show a measurable advantage on your real workload.
Choose Claude Opus 5 for complex agentic coding
Opus 5 is Anthropic's recommended model for complex agentic coding and enterprise work. It is a strong fit for repository-scale changes, difficult debugging, multi-tool workflows, and decisions that need deep analysis but do not consistently justify Fable's higher cost.
Choose Claude Sonnet 5 as the general default
Sonnet 5 offers the best balance of speed and intelligence in Anthropic's current comparison. For product features, document analysis, routine coding, structured writing, support workflows, and interactive tools, it is the sensible model to evaluate first.
Choose Claude Haiku 4.5 for speed and volume
Haiku 4.5 is the fastest and least expensive model in this group. Use it for classification, extraction, concise transformations, routing, autocomplete-like interactions, and high-volume tasks that are easy to verify. Its context window is 200k tokens rather than the 1M listed for Fable 5, Opus 5, and Sonnet 5.
A practical model-selection ladder
Define success
Create a small evaluation set from real inputs and decide what a passing answer looks like.
Start efficient
Test Haiku for simple work or Sonnet for general work before escalating automatically.
Escalate by evidence
Try Opus or Fable only where the lower-cost model misses a measurable requirement.
Measure the system
Track accuracy, human review time, latency, retries, and total cost—not token price alone.
Use this decision rule: - Simple, repeatable, latency-sensitive task → Haiku 4.5 - Most capable general workflow → Sonnet 5 - Complex agentic coding or enterprise task → Opus 5 - Hardest long-running agent where quality dominates cost → Fable 5 Then validate the choice on your own examples. Do not select a model from a benchmark headline alone.
A larger model cannot rescue an undefined task
Model choice and prompt quality solve different problems. A stronger model may reason better, but it still needs the goal, context, constraints, evidence, and required output. Before paying for more capability, remove ambiguity and give the model a testable definition of done.
The short answer
Start with Sonnet 5 for most serious work, Haiku 4.5 for fast simple workloads, Opus 5 for complex agentic coding, and Fable 5 for the hardest long-running agents. Escalate only when your own evaluation proves the value.