Claude Sonnet 5 Launch: $2/$10 Promo Pricing Undercuts Opus 4.8 for Coding Agents

What Anthropic Shipped Today

Anthropic released Claude Sonnet 5 on July 1, 2026. The public pitch: reasoning, tool use, coding and knowledge-work performance approaching Opus 4.8, delivered at a fraction of the cost. The launch numbers to remember are the ones on the invoice, not the benchmark chart:

The promo undercuts Sonnet 4.6 by 33% flat, and undercuts Opus 4.8 by roughly 2.5× on both dimensions. Anthropic almost never launches an intelligence upgrade at a discount to the previous generation. The two-month window is the story.

Why the Promo Is Meaningful for Coding Agents

Coding agents are output-heavy relative to chat. A typical Claude Code session runs 8:1 to 12:1 input-to-output tokens over a 30-minute session, because the agent reads the codebase, holds context, and writes back diffs. When output prices drop, coding costs drop with them. Sonnet 5 promo cuts output by exactly the same 33% as input compared to Sonnet 4.6 — but versus Opus 4.8 the drop is 2.5× on both meters, which for output-heavy workloads compounds fast.

Concretely: an agent that consumed 4M input tokens and 350K output tokens per developer-day on Opus 4.8 was billing $28.75/day at Opus pricing. The same workload on Sonnet 5 promo pricing runs $11.50/day. Even if you have to burn 20% more tokens because Sonnet 5 misses some Opus-only reasoning wins, you are still at $13.80/day. The two-month promo period is 61 days; that is a $910–$1,050 delta per developer if you can route successfully.

Where Sonnet 5 Actually Replaces Opus

Anthropic's launch materials call out three capability upgrades over Sonnet 4.6: planning, browser tool use, and terminal tool use. In practice this means Sonnet 5 becomes the default for:

• Multi-file edits — Sonnet 4.6 tended to lose track around the 8-file mark; Sonnet 5 holds through mid-teens on typical monorepos.
• Terminal-driven agents — bash tool sequences with dependency between steps, where Sonnet 4.6 sometimes retried against stale state.
• Long-form refactors — the 200K context is unchanged, but tool-use stability makes it usable for the full window.

Where Opus 4.8 still earns its price: architecture design under ambiguity, adversarial code review, novel algorithm work. If your agent's job is to make judgement calls rather than execute a plan, Opus is still the right router destination. If the agent is executing a well-scoped diff, Sonnet 5 wins on total cost with acceptable quality.

Routing Strategy for the Promo Window

A defensible routing setup for the next 61 days:

1. Set Sonnet 5 as the default coding model for all agents.
2. Keep Opus 4.8 available as an escalation for /deep-think-style flagged tasks.
3. Enable prompt caching on both — cache hits price at 10% of write, so long system prompts amortize even faster.
4. Instrument per-model spend so you can measure the actual output-quality delta rather than assuming.

Update: Anthropic made the $2/$10 rate permanent. The previously announced September 1 increase to $3/$15 will not occur, so remove that increase from forecasts while continuing to verify the official rate card.

Two-Month Playbook

If your team has been budget-blocked from doing an expensive migration, refactor, or codebase reorganization, permanent $2/$10 pricing removes the artificial two-month deadline. Compare cost per accepted outcome with Opus, then schedule work by readiness rather than a canceled pricing cliff.

Rebuilding eval suites, generating documentation, and automating repetitive maintenance may all benefit from the lower standard rate, but they still need task budgets and stop conditions.