Best Claude Model for Coding: A Task-by-Task Decision Guide

#The best Claude model for coding depends on the task — start with Sonnet 5

For ordinary coding work, default to Sonnet 5. It is fast, permanently priced at $2/$10 per million input/output tokens since 2026-08-10 (not an introductory rate), and gives a 1M-token context window. Switch up to Opus 5 ($5/$25) when a task is genuinely hard or long-running: multi-file refactors, debugging that crosses layers, or work against fast-moving frameworks where its May 2026 knowledge cutoff matters. Switch down to Haiku 4.5 ($1/$5) for bounded, checkable, high-volume steps. Fable 5 is almost never the right coding spend (details below).

Basis: Anthropic's model-selection guidance (C-013), internal cost/intelligence sweeps (C-015), Claude Code team's model-vs-effort heuristic (C-025), and Haiku 4.5 agentic-coding results (C-026). Verified 2026-08-26.

This guide covers coding only; the broader "best Claude model" decision weighs the same menu differently.

#The current Claude lineup and what it costs (August 2026)

Four current models matter for coding; API IDs, prices, context windows, effort support, and knowledge cutoffs are below.

#Models, API IDs, and pricing

#Context, output, effort, and knowledge cutoff

All figures from Anthropic's models overview and pricing page, verified 2026-08-26. Sonnet 5's $2/$10 launched as introductory pricing; per the 2026-08-10 edit to the announcement it is permanent — the scheduled 2026-09-01 rise to $3/$15 will not occur (C-005). Haiku 4.5 supports extended thinking only; the other three support adaptive thinking with an effort parameter, default high (C-001).

#Did it not know enough, or not try hard enough?

Model choice and reasoning effort are two separate controls. Reasoning effort is how much work the model does per turn — how many files it reads, how many tool calls it makes, whether it verifies before reporting back. The model sets the capability ceiling; effort sets thoroughness.

Anthropic's Claude Code team's heuristic: if Claude had the context, clearly tried, and still got it wrong, the task outgrew the model — switch models. If it skipped a file, never ran the tests, or bailed partway, it did not try hard enough — raise effort instead. Effort has five levels (low, medium, high, xhigh, max) and defaults to high; changing it mid-conversation invalidates the prompt cache, so set it at session start.

Migrating from Opus 4.7 or 4.8? Do not carry old effort settings over: their recommended coding start was xhigh; on Opus 5 it is the default high. Re-run a fresh effort sweep and delete "verify twice" prompt text — Opus 5 self-verifies, and Anthropic measured stale 4.8-era prompts costing 36% more per ticket.

#When to switch up to Opus 5 (and why Fable 5 is rarely the answer)

Anthropic's own positioning starts complex agentic coding — work where the model runs a loop of reading files, calling tools, running tests, and editing over many turns — on Opus 5. Signals a task has outgrown Sonnet 5:

• A refactor spanning many files or layers, where a wrong early assumption compounds.
• Debugging where cause and symptom sit in different layers.
• Long-running agent sessions where Sonnet 5 keeps re-deriving context or abandoning subtasks.
• Work against a fast-moving framework — Opus 5's cutoff is May 2026 versus January 2026 for Sonnet 5 and Fable 5, so it is less likely to hallucinate deprecated APIs.

The premium is real: Opus 5 runs $5/$25 against Sonnet 5's $2/$10 — the same price as Opus 4.8, whose retirement horizon runs into 2027 — and Anthropic rates its latency moderate versus fast. Most "I need Opus" moments are "didn't try hard enough" moments that higher effort fixes for less money.

Fable 5 is the capability ceiling, not the coding default. On Anthropic's internal 482-problem SWE-bench Pro subset (selected for harness compatibility, not comparable to the public leaderboard), Opus 5 scored 91.7% versus Fable 5's 91.3% — inside run-to-run noise — at about 60% of the cost. Anthropic publishes no SWE-bench Verified or Pro leaderboard numbers for Opus 5; treat third-party percentages as unverified. Access friction too: Fable 5 is never the Claude Code default, and safety classifiers reroute some flagged requests to Opus models mid-session. Unless you have measured a task Opus 5 at xhigh cannot do, skip it.

#When a cheaper or faster model is enough

Haiku 4.5 is the fastest and cheapest current model at $1/$5; Anthropic positions it as the starting point for high-volume, latency-sensitive work and sub-agent tasks. It fits steps with a cheap, automatic check — a script you will run, a test you will execute, a diff you will review. It has no effort parameter (extended thinking only); vendor-reported figures put it near 90% of Sonnet 4.5 on an agentic coding evaluation but much further behind on long tasks — keep it off long-running loops.

Or keep Sonnet 5 and step effort down: in Anthropic's internal sweeps, medium on long-running coding work gave up about 2 points for half the cost; Sonnet 5 at medium is comparable to Sonnet 4.6 at high, and low suits high-volume, latency-sensitive work. Budgeting caveat: Claude 4.7+ models use a newer tokenizer emitting roughly 30% more tokens for the same text — recount usage before comparing costs across generations.

#Where the choice happens in Claude Code

Claude Code resolves default to Opus 5 on Max, Team Premium, Enterprise pay-as-you-go, and API accounts, and to Sonnet 5 on Pro and Team Standard. Switching is a one-line change, with version floors — Opus 5 needs Claude Code v2.1.219+, Sonnet 5 v2.1.197+:

# --- Switch the model mid-session (inside Claude Code) ---
# Saved as your default since Claude Code v2.1.153 (C-020).
> /model claude-opus-5
> /model claude-sonnet-5

# --- Switch the model at launch (from your shell) ---
claude --model claude-opus-5
claude --model claude-sonnet-5

# Or via environment variable (C-020):
#   export ANTHROPIC_MODEL=claude-opus-5

# --- Set reasoning effort (C-021) ---
# low | medium | high | xhigh | max
# Default is high on every effort-capable model except Opus 4.7 (xhigh).
# low/medium/high/xhigh persist across sessions; max is session-only
# unless set via CLAUDE_CODE_EFFORT_LEVEL.
> /effort high
claude --effort xhigh

# --- Version floors (C-020) ---
#   Opus 5    requires Claude Code v2.1.219+
#   Sonnet 5  requires Claude Code v2.1.197+
#   Fable 5   requires Claude Code v2.1.170+
# Check your version with: claude --version

/effort and --effort settings persist across sessions, except session-only max. For the local test below, pin full model IDs (or set ANTHROPIC_DEFAULT_OPUS_MODEL / ANTHROPIC_DEFAULT_SONNET_MODEL): aliases like sonnet and opus resolve per provider and drift over time.

This page does not cover plan limits or slowness: check Claude usage limits or diagnose Claude Code slowness.

#Still not sure? Run a small test on your own tasks

When the Sonnet 5 versus Opus 5 call is close, no benchmark settles it — your workload does. Anthropic's own method for weighing cost against intelligence, scaled to one developer (it baselines across effort levels, not just defaults):

1. Pull three to five real tasks from recent work — tickets or commits, not toy prompts.
2. Run each candidate on the same tasks in the same coding agent at default effort; record pass or fail.
3. Record tokens and latency for each run alongside the verdict.
4. Compare cost per completed task, not per token — a more capable model often finishes in fewer turns, and a failed cheap attempt still bills its tokens plus the retry.

To run this without touching your Anthropic subscription: route Claude Code to a selected, separately billed catalog model through Fast Inference (fast claude on --model <catalog-id> maps Claude Code's model slots to one catalog model; fast claude off restores your settings), then verify per-model tokens and spend in the usage dashboard. It bills a separate Fast balance — no added Anthropic plan quota, no speedup to an unchanged Anthropic model. Setup mechanics live on the Claude Code integration docs.

Step-by-step setup: Claude Code on Fast Inference guide.