Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku

Anthropic detailed the production lineup and task economics across its model family in claude.dev's report, providing developers with comparative pricing and operational guidance. The four generally available models consist of Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, and Claude Fable 5.2, alongside Claude Mythos 5.2 for specialized research through verification programs. Anthropic recommends Opus 5.5 at default medium effort as the initial starting point for daily software engineering and general knowledge workflows.

This analysis breaks down token economics, latency trade-offs, and multi-tier agent orchestration for engineers building on the Claude Platform or using Claude Code. Readers will find pricing matrices, benchmark comparisons, effort calibration steps, and copy-paste environment commands to configure individual workloads effectively. By aligning task complexity with model tier, teams can optimize API spending without sacrificing automated verification quality.

01Model selection criteria across common workloads

Starting with Opus 5.5 handles everyday software engineering, debugging, and multi-pull-request feature branches while balancing token expenditure against complex reasoning. Developers should step down to Sonnet 5.5 for well-scoped programming tasks, bug fixes, feature iterations, and recurring administrative drafting. Haiku 5.5 fits classification, data extraction, and subagent routing loops where wrong outputs remain inexpensive to detect and verify.

Demanding refactors, architecture migrations, and quantitative financial modeling warrant escalating to Fable 5.2 at medium or high effort levels. Vision-heavy tasks and computer-use workflows run best on Opus 5.5 due to sustained agentic loops over multiple screen captures. When developers remain uncertain of precise technical requirements, Opus 5.5 serves as the default configuration across Claude Code and enterprise tiers.

Teams should calibrate effort levels as a finer operational dial before switching models outright. Pushing effort past high increases inference latency and can match the pricing tier of an upgraded model. Raising effort gives an existing model more time to call tools and verify answers, though it cannot instill missing architectural capabilities.

02Technical capabilities and benchmark comparisons

Evaluating cost per finished task reveals that token prices do not always dictate total development expense. On Terminal-Bench 3.0, Opus 5.5 operating at high effort achieved 58.9 percent accuracy using half the tokens of Fable 5.1 at max effort, which reached 58.0 percent. When tested on FrontierCode 2.0, Fable 5.2 at high effort achieved 45.3 percent accuracy, closely matching Opus 5.5 at max effort at 45.5 percent for half the cost.

Latency and throughput metrics also favor newer releases across the product catalog. Fable 5.2 generates tokens up to 30 percent faster than Fable 5.1, while Sonnet 5.5 runs 30 percent faster than Sonnet 5. On SWE-bench Pro evaluations, running Opus 5.5 at medium effort reduced accuracy by 2.5 points while cutting token costs to 70 percent compared to high effort.

03Cost optimization through automated verification loops

Structuring automated verification pipelines allows teams to cut inference costs significantly without reducing overall pass rates. Rather than running entire test suites on high effort, developers can process initial attempts at low effort and escalate only failed runs. On SWE-bench Pro subsets, running Opus 5.5 at low effort with high-effort retries passed 97 percent of tasks at $0.17 each, outperforming constant high effort at 95.3 percent and $0.29.

Prompt caching features across the model family maintain fixed read rates of $0.20 per million tokens for both Opus 5.5 and Fable 5.2. Sonnet 5.5 reduces cache reads to $0.10 per million tokens, while Haiku 5.5 offers $0.01 per million tokens on prompts up to 100,000 tokens. Concentrated benchmark spending patterns show that two difficult problems accounted for 43 percent of total expenditure on a 20-problem benchmark suite.

04CLI setup and model switching commands

Engineers working inside Claude Code can install or refresh the CLI using standard package manager commands.

npm install -g @anthropic-ai/claude-code

Active sessions support direct runtime switching between models and status inspections via terminal directives.

/model claude-fable-5-2
/effort status

Developers launching isolated workloads or pinned production environments can declare the target model and manage global defaults.

claude --model fable
claude config set model opus

05Production deployment limits and operational monitoring

Large language model pipelines require monitoring thinking parameters alongside context limits in production environments. Both Opus 5.5 and Fable 5.2 feature always-on adaptive thinking on the API and include fast mode support. In contrast, Sonnet 5.5 and Haiku 5.5 offer optional adaptive thinking that can be disabled, with Sonnet 5.5 skipping upfront thinking between tool calls.

Every model in the family provides a 1,000,000 token context window paired with a 128,000 token maximum output ceiling. High-volume systems must factor in input tier escalations, such as Haiku 5.5 rising from $0.10 to $0.50 per million tokens when prompt sizes exceed 100,000 tokens. Establishing programmatic assertions prior to upgrading models ensures that teams pay premium token rates only on genuine edge cases.