Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026, as the second model in its Claude 5.5 family, positioning it for well-defined everyday work like bug fixing, documentation, presentations, and spreadsheets. It generates output more than 30 percent faster than Sonnet 5 and cuts per-task costs by up to 30 percent through more efficient token usage, with no API price change. On Terminal-Bench 4.0 it jumps from Sonnet 5's 10.3 percent to 70.6 percent, and it nearly matches Opus 5.5 on several knowledge-work tests. The model is available immediately on AWS, Google Cloud, and Azure, with Anthropic also adding new safeguards against cybersecurity risks and distillation attacks. Anthropic announced a forthcoming Claude Haiku 5.5 aimed at high-throughput, low-cost use cases. The release positions Sonnet 5.5 above OpenAI's GPT-6 Luna, with Opus 5.5 sitting above Sol and Fable above Astra, intensifying the latest pricing battle across providers.

Vellum's benchmark walkthrough confirms Sonnet 5.5's headline gain on Terminal-Bench 4.0, where it scores 70.6 percent at peak effort versus Sonnet 5's 10.3 percent and edges past Opus 5.5's 66.4 percent at extra-high effort. On GDPval-AA v2.1 it posts 1844 Elo, two points shy of Opus 5.5 and nearly 360 points ahead of GPT-6 Sol. List pricing holds at 2 and 10 dollars per million input and output tokens. The effective cost per task drops up to 30 percent because Sonnet 5.5 resolves problems in significantly fewer steps and batches tool executions more aggressively, avoiding unnecessary subagent spawning. At medium effort it exceeds Sonnet 5's best Terminal-Bench score for less than a tenth of the per-task cost. The breakdown spans coding, knowledge work, and system control categories against Opus 5.5 and GPT-6 Sol.

Anthropic announced Claude Sonnet 5.5 as a faster, more efficient update to its mid-tier workhorse, generating output more than 30 percent faster than Sonnet 5 and reducing total task cost by up to 30 percent via fewer tokens and fewer tool calls at the same API price of 2 and 10 dollars per million input and output tokens. On GDPval-AA it scores 1844 versus Opus 5.5's 1846 and far ahead of Sonnet 5's 1449, and reaches 80.1 percent on OSWorld 2.1. Sonnet 5.5 targets well-defined everyday work such as debugging, coding, document production, presentations, spreadsheets, and interface design, while Opus 5.5 remains preferred for ambiguous or judgment-heavy tasks. The narrowed gap suggests Anthropic's pitch is shifting from per-token price to total cost of completing a job. Cache reads are priced at 0.20 dollars per million tokens, with cache writes at 2.50 dollars, well below Opus 5.5's 5-dollar cache write rate.

BenchLM's radar confirms Sonnet 5.5's September 28, 2026 release with 47 published benchmark rows across agentic, coding, multimodal, and knowledge categories, plus a 1M-token context window and 2-dollar input, 10-dollar output pricing. It ranks second in coding at the 99th percentile with a score of 79.8, fifth in agentic at the 92nd percentile with 66.2, and sixth in knowledge at the 97th percentile with 80.9. Verified evidence covers 11 agentic benchmarks, 9 coding benchmarks, 7 multimodal benchmarks, and 16 knowledge benchmarks, while reasoning and instruction-following remain unmeasured. Capability index lands at 80.5 of 100 against a field median of 50.5, placing it fifth of 209 ranked models. Independent runtime speed has not yet been measured on this profile.

CodeRabbit tested Claude Sonnet 5.5 through its review pipeline and found that on 13 hardest known-bug cases it caught 6 of 13 issues via actionable comments versus 4 for Sonnet 5, at 41.2 percent actionable precision. It delivered this in roughly half the wall-clock time, with list-priced model calls costing about 60 percent less per review. Four of its catches were bugs Sonnet 5 had waved through. A larger run on 44 open-source pull requests confirmed the speed and comment-volume improvements, making this the first Sonnet release CodeRabbit would consider for the main review pass rather than only comment quality. The evolution across Sonnet 4, 4.5, 4.6, 5, and 5.5 traces growing coverage and reasoning depth, with 5.5 marking the point where detection finally outpaces earlier variants while keeping precision steady.

Vals AI lists Claude Sonnet 5.5 with a September 28, 2026 release, developer Anthropic, 1M-token context window, 128k max output, and list pricing at 2.00 and 10.00 dollars per million input and output tokens. Its overall Vals Index accuracy is 69.22 percent plus or minus 0.96, ranking second of 66 models, with a cost per test of 20.80 dollars and 70 minute latency. It leads Code Migration at 69.83 percent and tops Vibe Code Bench at 92.39 percent. The model supports text, image, and file inputs but not video, and posts a 2.27 percent fallback rate alongside a 0.14 percent refusal rate. It places first on ProofBench at 100 percent and BioMysteryBench at 81.11 percent, while landing in the mid-tier on Legal Research Bench at 48.08 percent. Terminal-Bench 4.0 shows 53.03 percent, with separate Anthropic-published figures reaching 70.6 percent at peak effort.

Claude Sonnet 5.5 launched six days after Opus 5.5 as Anthropic's new mid-tier model, replacing Sonnet 5 at the same price and described as the best combination of speed and intelligence in the current lineup. It edges past Opus 5.5 on agentic terminal coding while sitting just a couple of points behind on knowledge work and computer use at half the per-token price. The launch reflects a logarithmic scaling curve where most performance value remains cheap to obtain. Anthropic reports gains across every published benchmark, with the largest leaps in terminal coding and chart-related reasoning tasks. DataCamp recommends making Sonnet 5.5 the default Claude model while reserving Opus 5.5 for open-ended, judgment-heavy workloads. Haiku 5.5 is forthcoming for high-throughput scenarios, though Haiku 5 itself never previously shipped.