Anthropic Introduces Claude Sonnet 5.5

Anthropic has released Claude Sonnet 5.5, the second foundation model in its Claude 5.5 family. Positioned as a faster, lower-cost companion to Claude Opus 5.5, the model generates outputs more than 30% faster than Sonnet 5 while reducing compute costs by up to 30% on everyday workflows. According to the official announcement from Anthropic, the system specializes in software engineering, multi-file code editing, and structured office tasks.

While Opus 5.5 handles open-ended technical architecture requiring deep deliberation, Sonnet 5.5 targets high-frequency engineering loops, bug fixing, and rapid document drafting. Anthropic plans to complete the model tier with Claude Haiku 5.5 in upcoming updates.

Token pricing and operational speed

Token rates match Sonnet 5 at $2.00 per million input tokens, $10.00 per million output tokens, and $0.20 per million tokens for prompt cache reads. Because Sonnet 5.5 completes tasks in fewer total reasoning steps, production costs per workload decrease by up to 30%. Output generation speeds exceed Sonnet 5 by more than 30%.

Pricing Tier (per 1M tokens)	Claude Sonnet 5.5	Claude Opus 5.5	Shift vs Predecessor
Prompt Cache Reads	$0.20	$0.20	Parity with Sonnet 5
Prompt Cache Writes	$2.50	$5.00	50% lower than Opus
Input Tokens	$2.00	$4.00	50% lower than Opus
Output Tokens	$10.00	$20.00	50% lower than Opus
Generation Throughput	30%+ faster	Standard pace	Fastest Sonnet release

Agentic software engineering and coding evaluations

The model displays sharp gains in command-line software tasks. On Terminal-Bench 4.0, which tracks autonomous execution in Linux container shells, Sonnet 5.5 scored 70.6% at Max effort, surpassing Sonnet 5 (10.3%) and leading Opus 5.5 (66.4%). At Medium effort, the default setting across Claude consumer apps, Sonnet 5.5 exceeds Sonnet 5’s top score for less than a tenth of the compute cost per run.

On CursorBench 4.0, which tests real multi-file refactoring cases collected from developer sessions, Sonnet 5.5 scored 55.5%, coming within two points of Opus 5.5 (57.8%) and surpassing Sonnet 5 (34.1%). On FrontierCode 1.1 Main, which checks whether pull requests meet production repository conventions, the model reached 52.1% at Xhigh effort, outperforming Sonnet 5 (42.4%) and GPT-6 Sol (49.3%).

Early industrial trials confirmed these efficiencies. Engineering leadership at Epic Games reported that Sonnet 5.5 managed tens of thousands of lines of gameplay system architecture during system design audits with minimal guidance. Automated pull request review platform CodeRabbit noted that Sonnet 5.5 eliminated unnecessary web searches and reduced excessive token consumption compared to Sonnet 5.

Professional knowledge work and computer use benchmarks

Evaluations conducted by independent research platform Artificial Analysis show Sonnet 5.5 scoring 1844 on GDPval-AA v2.1 across 44 occupational categories, nearly matching Opus 5.5 (1846) and gaining roughly 400 points over Sonnet 5 (1449). On AA-Briefcase v1.1, which measures real-world enterprise assignments, Sonnet 5.5 scored 1811, surpassing GPT-6 Sol (1483).

In desktop computer automation on OSWorld 2.1, Sonnet 5.5 reached 80.1% partial task completion, outpacing Sonnet 5 (57.0%) and approaching Opus 5.5 (81.8%). On visual chart recognition via Chartography, it scored 61.6% without external tools, beating Sonnet 5 (15.6%) and GPT-6 Sol (53.6%). In addition, Sonnet 5.5 became the first Sonnet model to complete Pokémon Red relying exclusively on visual screenshots.

Benchmark	Claude Sonnet 5.5	Claude Sonnet 5	Claude Opus 5.5	GPT-6 Sol
Terminal-Bench 4.0	70.6% (Max)	10.3%	66.4% (Xhigh)	(not reported)
FrontierCode 1.1 (Main)	52.1% (Xhigh)	42.4%	54.4%	49.3%
CursorBench 4.0	55.5%	34.1%	57.8%	(not reported)
GDPval-AA v2.1 (Elo)	1,844	1,449	1,846	1,487
AA-Briefcase v1.1 (Elo)	1,811	1,359	1,822	1,483
Humanity’s Last Exam	64.5% (tools)	54.9% (tools)	67.7% (tools)	(not reported)
OSWorld 2.1 (Computer Use)	80.1% (partial)	57.0% (partial)	81.8% (partial)	(not reported)
Chartography (Visual)	61.6% (no tools)	15.6% (no tools)	64.4% (no tools)	53.6% (no tools)

Safety, cybersecurity fallbacks, and anti-distillation

On Anthropic’s automated behavioral audit across approximately 1,850 scenarios, Sonnet 5.5 matched or improved upon Sonnet 5 on misuse resistance and sandbox containment. Because the model demonstrates strong cybersecurity proficiency, Anthropic deployed safeguards with visible fallbacks to Sonnet 5 on high-risk exploit requests. Defensive security analysts can apply for tiered permissions through the expanded Cyber Verification Program.

To protect intellectual property from automated extraction, Sonnet 5.5 includes classifiers that counter model distillation attempts. The model expands preserved thinking, preventing third parties from separating reasoning traces from the original account. Zero data retention policies apply to commercial deployments.

Availability and migration guide

Claude Sonnet 5.5 is available on Claude apps, the Claude API under the identifier claude-sonnet-5-5, and through cloud infrastructure providers including Amazon Web Services Bedrock, Google Cloud Vertex AI, and Microsoft Azure. Teams transitioning from Sonnet 5 using thinking-off modes must configure the between_tools setting to maintain compatibility.