Claude Opus 5.1 Benchmarks: How Much Better Is It Than Opus 5?

Anthropic's Claude Opus 5.1 model shows significant benchmark gains over its predecessor Opus 5, with scientific research capability roughly doubling to over 50% on Terminal Bench Science 1.0 and business workflow automation nearly doubling from 17.1% to 31.4% on AutomationBench. Agentic coding scores rose to 55.8% (up from 42-52.3% for Opus 5 and 37.3% for GPT-5.6), while computer use performance reached 77.9% on OSWorld 2.0, and safety over-triggering dropped by 60% on benign requests.

Claude Opus 5.1 Benchmarks: How Much Better Is It Than Opus 5? Claude Opus 5.1's benchmark gains over Opus 5 and GPT-5.6, covering coding, research, computer use, and business automation scores. What are the actual benchmark numbers for Opus 5.1 versus Opus 5? Claude Opus 5.1 posts meaningful gains over Opus 5 across every major category Anthropic tracks, with the largest jump on scientific research roughly 25-29% for Opus 5 up to over 50% for Opus 5.1 and business workflow automation, which nearly doubled from 17.1% to 31.4%. Coding, computer use, and multidisciplinary reasoning each improved by roughly 3 to 6 percentage points, and Opus 5.1 also beats GPT-5.6 on every benchmark cited. TL;DR Scientific research capability roughly doubled, with Opus 5.1 scoring over 50% on Terminal Bench Science 1.0 compared to around 24.7-29% for Opus 5, a jump described as the standout result of the release. Business workflow automation almost doubled too, going from 17.1% on Opus 5 to 31.4% on Opus 5.1 as measured by AutomationBench, which tests a model’s ability to automate real pipelines used inside companies. Agentic coding scores climbed to 55.8% on Opus 5.1 versus 42-52.3% for Opus 5 and 37.3% for GPT-5.6, and on Cursor Bench 3.2.0 Opus 5.1 hit 73.4% against 70-70.5% for the prior generation. Computer use performance , historically a weak spot for Anthropic’s models relative to OpenAI’s, rose to 77.9% on OSWorld 2.0 using a partial-credit scoring method, up from 72.9-75.4%. Cost efficiency roughly improved 2.5x per task according to Anthropic’s own cost-per-task chart, and cache-read pricing dropped by about a quarter, with up to 45% reductions claimed for highly agentic workloads. Safety over-triggering dropped sharply , with benign requests getting flagged 60% less often overall and biology/medical fallback rates down 85%, addressing a common complaint about earlier Claude models being overly cautious. Knowledge-work scoring on GDPval-AA v2 rose to 1853 from 1723-1824 for Opus 5 and 1711 for GPT-5.6, an improvement of roughly 7% on tasks meant to approximate real economic value creation. Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. How much did agentic coding actually improve? Opus 5.1 scored 55.8% on the agentic coding benchmark cited in Anthropic’s release, up from a range of 42% to 52.3% for Opus 5 depending on configuration, and well ahead of GPT-5.6’s 37.3%. On Cursor Bench 3.2.0, a separate coding-agent benchmark, Opus 5.1 reached 73.4% against roughly 70% for Opus 5. That’s not a leap on the scale of the research benchmark, but it’s a consistent, real improvement rather than noise. For teams running autonomous coding agents in CI pipelines or using Claude inside IDE integrations, this translates to fewer failed task completions and less human intervention needed to unblock stuck agents, though the exact real-world impact depends heavily on codebase complexity and task length. Why did the scientific research score jump so much? The research benchmark, Terminal Bench Science 1.0, tests a model’s ability to run open-ended scientific investigation loops, the kind of “auto-research” workflow where an agent formulates hypotheses, runs experiments in a terminal environment, and iterates without a human in the loop. Opus 5 and GPT-5.6 both scored in the mid-to-high 20s percentage-wise on this benchmark. Opus 5.1 crossed the 50% mark, which is more than double the prior generation’s performance. That kind of gap suggests a structural improvement in how the model handles long-horizon, low-supervision tasks rather than a marginal tuning update. It’s also one of the newer benchmarks in the industry, so there’s less historical data to compare against, but the size of the jump makes it the most notable single number in this release. What does the AutomationBench score mean for business use cases? AutomationBench evaluates how reliably a model can automate workflows and systems that resemble what’s actually installed inside companies, things like connecting APIs, scripting repetitive processes, and building pipelines that don’t fall over on edge cases. Opus 5 automated these workflows reliably about 17.1% of the time. Opus 5.1 does it 31.4% of the time, which is close to double. For anyone building automation tooling or offering workflow-as-a-service products, that’s a meaningful reliability improvement, since a model that succeeds on a third of representative tasks instead of a sixth changes the economics of how much manual QA and retry logic you need to wrap around it. Is Opus 5.1 actually cheaper to run, or just more capable? Anthropic published a cost-per-task comparison rather than a straight per-token price comparison, and on that measure Opus 5.1 comes out ahead by roughly 2.5x. The logic is that per-token pricing doesn’t tell you much on its own, because more capable models often complete tasks in fewer total tokens, while older or less efficient models can rack up token usage even at a lower headline rate. Comparing dollars spent for a given quality of output is a more honest measure of efficiency. Anthropic also changed how cache reads are priced, with reported cost reductions of about 25% generally and up to 45% for heavily agentic workloads that make repeated calls in quick succession. Combined with the intelligence gains, this points to a model that’s both more capable and cheaper to operate per completed task, which is a different and more useful claim than just “smarter.” Did Anthropic fix the over-cautious safety behavior? One of the most common complaints about earlier Claude models was excessive refusal or fallback behavior, where a benign question about biology, cybersecurity, or hardening an application would get flagged and routed to a less capable model instead of being answered directly. Anthropic says Opus 5.1 reduces these false-positive safety triggers by 60% overall, and specifically cuts fallback rates on biology and medical questions by 85%. That’s a significant shift for anyone using Claude for legitimate technical or scientific work who previously had to rephrase prompts or switch models to get a straight answer. Do these benchmark gains matter for how the model actually feels to use? Benchmarks measure standardized tasks, and standardized tasks don’t always predict how a model performs in open-ended, real conversation. It’s worth noting that model comparisons based on benchmark scores alone can diverge from subjective quality, since a model that tests well on structured tasks doesn’t always produce the most useful or well-judged output in freeform use. That’s a real limitation of any benchmark-first analysis, including this one. The more informal comparisons that circulate after a model release, people testing creative tasks, novel interaction designs, or just vibes, often surface differences that don’t show up in the official numbers at all. Treat the benchmark gains described here as a floor on capability improvement, not the whole picture. Frequently Asked Questions How much better is Claude Opus 5.1 than Opus 5 on coding tasks? On agentic coding benchmarks, Opus 5.1 scored 55.8% compared to a range of 42-52.3% for Opus 5. On Cursor Bench 3.2.0, it scored 73.4% against roughly 70% for Opus 5, a smaller but consistent improvement. What is AutomationBench and why does it matter? AutomationBench measures a model’s ability to automate workflows and systems resembling real business pipelines. Opus 5.1 scored 31.4% versus 17.1% for Opus 5, nearly doubling reliability on this type of task. Does Opus 5.1 beat GPT-5.6 on these benchmarks? Yes, based on the figures cited, Opus 5.1 outperforms GPT-5.6 across every category mentioned, including research, coding, computer use, and knowledge work. Is Claude Opus 5.1 cheaper to run than Opus 5? Anthropic’s cost-per-task comparison suggests roughly 2.5x better cost efficiency for equivalent output quality, plus cache-read price cuts of about 25%, rising to 45% for highly agentic workloads. Did Opus 5.1 fix Claude’s tendency to over-flag safe requests? Anthropic reports a 60% reduction in false-positive safety flags overall, with an 85% reduction specifically for biology and medical questions, addressing a frequent user complaint about earlier versions.