Best Claude Model for Coding: Fable 5 vs Opus vs Sonnet vs Haiku

As of July 18, 2026, Anthropic's coding lineup is four models across four tiers. If a ranking page you are reading mentions any other Claude model as current, it is stale, Sonnet 4.6, for example, was replaced by Sonnet 5 on June 30, 2026.

Benchmark scores above come from the sources named in each cell's benchmark, primarily the swebench.com leaderboard, re-checked on the verification date. Anthropic has not yet published comparable benchmark suites for Sonnet 5 and Haiku 4.5, so we show none rather than inventing them.

The frontier gap is real but narrow: per the swebench.com leaderboard, Fable 5 scores 95.0% on SWE-bench Verified against 88.6% for Opus 4.8, verified July 18, 2026. Whether those points are worth double the price depends entirely on the task, our full breakdown is on the Claude Fable 5 page.

The cost math makes the tiers concrete. Take a 2M-token agentic session, 1.6M tokens in, 0.4M out, typical for a meaty story with a few test iterations, at list prices from Anthropic's published per-MTok rates:

Inside Claude Code, you switch models with the /model command, mid-session if you like, the new model inherits the conversation context:

/model opus      # Opus 4.8, recommended default for complex agentic coding
/model sonnet    # Sonnet 5, routine work and cost control
/model haiku     # Haiku 4.5, subagents and quick edits
/model fable     # Fable 5, requires Claude Code v2.1.170+
/model opusplan  # Opus for plan mode, Sonnet for execution
/model best      # highest tier your org has access to (Fable 5 if available)

Three things about Fable 5 access worth knowing before you rely on it: it requires Claude Code v2.1.170 or later, it is not the default, you must select it explicitly, and it is unavailable under zero-data-retention agreements, so enterprise teams on ZDR contracts are capped at Opus 4.8 for now. The opusplan alias is the pragmatic middle path: flagship reasoning while you plan, workhorse pricing while the agent executes.

The per-task choice gets more interesting when you run agents in parallel: there is no reason every story in flight should burn tokens at the same tier. Assign Fable 5 to the hard architecture story, Sonnet 5 to the three routine ones, and let Haiku handle the subagent fan-outs inside each.

That is the workflow AIDEN, our macOS desktop app, is built around: it orchestrates your existing Claude Code CLI (and Codex, if you run both) on a kanban board, each story with an approved spec, its own git worktree, and its own model assignment, local-first, on the keys you already pay for, free for one project. The manual version of this workflow, and where it breaks down, is covered in Claude Code orchestration; the cross-vendor picture lives in the AI models for coding hub.