Anthropic dropped Claude Sonnet 5.5 on September 28. Six days earlier they said it would "follow in the coming weeks." The launch is notable for three reasons: the speed jump is real, the pricing did not move, and Claude Code switched its default Sonnet model to 5.5 in version 2.1.284 (September 28) and 2.1.285 (September 29).
What the numbers actually say
The headline figures from Anthropic and third-party verification:
- Terminal-Bench 4.0: 70.6% — up from Sonnet 5's 10.3%, and above Opus 5.5's 66.4%. This is an agentic coding evaluation that measures multi-step terminal task completion.
- Speed: 30%+ faster token generation than Sonnet 5.
- Cost per task: Up to 30% less because fewer tokens are burned per completed task.
- API pricing: Unchanged at $2/M input, $10/M output, $0.20/M cache reads, $2.50/M cache writes.
- Context window: 1M tokens (up from 200K on Sonnet 5).
- Max output: 128K tokens.
The Terminal-Bench jump is the eye-catcher. Sonnet 5.5 beats the flagship Opus 5.5 on this specific benchmark while costing half as much per million output tokens (Opus 5.5 is $15/M output). For developers running agentic coding loops — where the model drives a terminal, reads files, runs tests, iterates — this is the benchmark that maps closest to daily experience.
Claude Code made it default immediately
Claude Code 2.1.284 (September 28) switched the default Sonnet model to claude-sonnet-5-5. Version 2.1.285 (September 29) added --effort ultracode as a first-class toggle and a --desktop flag for the native installer. If you run claude today without pinning a model, you are on Sonnet 5.5.
This matters because most developers do not pin models. They run claude, get whatever the default is, and judge the experience by that. The default-switch is a signal: Anthropic believes Sonnet 5.5 is now the better daily driver for the majority of coding workloads.
What "30% less per task" means in practice
The per-task cost claim is not an API price cut — the per-token rates are identical to Sonnet 5. The savings come from fewer tokens spent per completed task. Faster generation means fewer thinking tokens. Higher Terminal-Bench scores mean fewer failed attempts and fewer correction loops. In agentic workflows where a single task can burn 50K-200K tokens across multiple tool calls, a 30% reduction in token burn per task compounds fast.
Cache pricing is also relevant: $0.20/M for cache reads means long-running agents that re-read the same context (codebase, specs, prior turns) pay pennies. The 1M context window makes this practical for whole-repo context.
Where it fits in the model lineup
| Model | Input $/M | Output $/M | Context | Terminal-Bench 4.0 | Positioning |
|---|---|---|---|---|---|
| Sonnet 5 | $2 | $10 | 200K | 10.3% | Previous default |
| Sonnet 5.5 | $2 | $10 | 1M | 70.6% | New default Sonnet |
| Opus 5.5 | $15 | $75 | 1M | 66.4% | Flagship, complex reasoning |
Sonnet 5.5 is now the sweet spot for agentic coding: it beats the flagship on the coding benchmark that matters most for agents, costs 1/7.5x the output price of Opus 5.5, and has the same 1M context window. Opus 5.5 still leads on GDPval-AA (a broader real-world work benchmark) by about two points, so for non-coding complex reasoning Opus keeps the crown.
What to do right now
- Update Claude Code:
npm update -g @anthropic-ai/claude-codeor reinstall the native binary. You want 2.1.284 or newer. - Don't pin Sonnet 5: If you have
ANTHROPIC_DEFAULT_SONNET_MODELor--model claude-sonnet-5-20261022(or similar) in your workflow, remove it. Let the default carry you forward. - Re-evaluate Opus usage: If you were using Opus 5.5 for coding because Sonnet 5 felt too slow or error-prone, try Sonnet 5.5 for a week. The Terminal-Bench gap suggests you may not need Opus for coding tasks.
- Watch cache costs: With 1M context, cache reads become a major lever. Structure your agent prompts to maximize cache hit rates — prefix stable context (repo map, specs, conventions) consistently.
One caveat: benchmarks are not production
Terminal-Bench 4.0 is a curated evaluation. Your codebase, your tooling, your agent prompts, and your failure modes are different. The 70.6% number is a signal, not a guarantee. What changed is that the default model for the most popular AI coding CLI is now the one that scores highest on the most relevant agentic coding benchmark — at the same price. That is a genuine step forward.
If this was useful, you can support my open-source work on Ko-fi or check out my services.