Claude Sonnet 5 for Coding: The Workhorse With a 1M Context

Claude Sonnet 5 went generally available on June 30, 2026, replacing Sonnet 4.6 as the middle tier of the Claude lineup, below Opus 4.8 and Fable 5, above Haiku 4.5. The API identifier is claude-sonnet-5, and the headline change is the window: 1M tokens (default), a size previous generations reserved for the expensive tiers, now standard at mid-tier pricing.

Sonnet has always been the tier most code actually ships on, and Sonnet 5 leans into that role: fast enough for tight edit-test loops, big enough to hold a real repo in context, and priced so that running several of them in parallel is not a budget conversation. Where it sits against the rest of the family is mapped in the best Claude model for coding.

Data verified July 18, 2026 against Anthropic's published pricing and docs. Anthropic has not yet published SWE-bench scores for Sonnet 5 that we can verify, so this page carries no benchmark table, when verified numbers land, they will appear here and on the models hub.

Sonnet 5 uses a new tokenizer, which sounds like a footnote and is not. The same source file tokenizes to a different count than it did under Sonnet 4.6, which means three things drift: cost projections built on 4.6-era token counts, context-fit heuristics ("this repo fits in N tokens"), and any per-story token budgets you enforce in CI or in an orchestrator.

The case for Sonnet 5 is not that it is the smartest Claude, it is that most stories in a sprint do not need the smartest Claude. Where it is the right pick:

The fleet case is where the economics get interesting. Running multiple Claude Code sessions in parallel multiplies every per-token price by the fleet size, which is why AIDEN boards commonly route routine story cards to Sonnet by default and reserve Opus for the cards that are flagged hard.

Sonnet 5 is noticeably behind Opus and Fable on hard multi-step reasoning, that is the honest trade for the price. The tells are consistent: stories that require holding a subtle invariant across many files, debugging where the cause is far from the symptom, and long autonomous runs that must recover from their own wrong turns. When a Sonnet session starts circling, re-attempting the same fix or contradicting its own earlier reasoning, that is the signal to move the story up a tier, not to re-prompt harder.

The next tier up is Opus 4.8 at $5 in / $25 out per MTok, Anthropic's recommended starting point for complex agentic coding, and above that Fable 5 at $10 in / $50 out per MTok for the genuinely hard work. The full decision rule across all three tiers is in Opus vs Sonnet vs Haiku.