Anthropic drops Claude Sonnet 5.5 with 30% speed boost and massive coding leap

While everyone was waiting to see who would claim the top benchmark spot next, Anthropic quietly supercharged its workhorse.

Usually, upgrading an AI model means bracing for fatter bills and sluggish latency. That wasn't the case on September 27, 2026, when Anthropic officially unveiled Claude Sonnet 5.5, the second entry in its 5.5 generation. Built for daily workflows, it runs over 30% faster and slashes per-task costs by up to 30% for routine jobs, all while breathing down the neck of the flagship Opus 5.5.

The model is available immediately across the Claude Platform, AWS Bedrock, Google Cloud, and Microsoft Azure, keeping zero data retention in place. A lighter Claude Haiku 5.5 will follow in the coming weeks.

Coding leaps and flagship parity

On paper, Sonnet 5.5 keeps the familiar sticker price of $2 per million input tokens and $10 per million output tokens. But cache reads have dropped by half, down to $0.10 per million tokens.

The bigger story is raw execution efficiency. On the agentic coding benchmark Terminal-Bench 4.0, Sonnet 5.5 skyrocketed to 70.6%, leaving Sonnet 5's 10.3% in the dust and eclipsing Opus 5.5 at 66.4%. In real-world IDE coding on CursorBench, it hit 55.5%, trailing Opus 5.5 by just over two points.

Early testers pointed out that Sonnet 5.5 grasps large codebases much quicker. It frequently batches tool calls into single steps, wrapping up multi-step tasks in fewer turns.

It matches that stride in white-collar work. On the GDPval-AA evaluation spanning 44 occupations, Sonnet 5.5 scored 1844, roughly 400 points ahead of Sonnet 5 and virtually tied with Opus 5.5 at 1846. It is also the first Sonnet model capable of beating Pokémon Red using only raw screenshots.

Faster outputs, but watch the effort dial

Speed gains are obvious right away. Anthropic says generation speeds are up by more than 30%, making interactive HTML creation, UI designs, and slides feel nearly instantaneous.

Because its cybersecurity skills now match older flagship tiers, Sonnet 5.5 becomes the first Sonnet model equipped with advanced safety fallbacks, automatically routing risky requests back to Sonnet 5. It also introduces defenses against distillation attacks to protect chain-of-thought data.

Independent testing, however, offers a crucial caveat. Artificial Analysis placed Sonnet 5.5 at 56 on its Intelligence Index, right behind Opus 5.5 at 58 and beating GPT-6 Astra. But at maximum effort, AA clocked average task costs around $7.60, roughly 50% higher than Sonnet 5.

Reviewers online noted that at top effort settings, Sonnet 5.5 tends to generate exceptionally verbose output tokens. Pushing the dial to the max can wipe out the savings.

Why it matters

For developers and office workers tackling scoped problems, Sonnet 5.5 is an instant no-brainer. Day-to-day bug fixes, feature patches, financial summaries, and presentation decks run smoothly at medium effort, delivering flagship-grade quality at a fraction of Opus pricing.

Still, automated pipelines and agentic builders should run careful cost audits before rolling it out wholesale. The lower token pricing and cheaper cache reads only save money if output tokens stay bounded. Left on max effort, runaway reasoning chains could surprise your finance team.

From an industry standpoint, Anthropic timed this move brilliantly. Landing right ahead of OpenAI's DevDay keynote, Sonnet 5.5 sets a demanding baseline for practical enterprise agents. It proves that mid-tier models can match last month's flagship intelligence while running faster.

OpenAI now faces heavy pressure to demonstrate that its lineup can deliver comparable per-task efficiency, not just raw benchmark vanity scores.