{"id":10822,"date":"2026-09-23T13:38:19","date_gmt":"2026-09-23T13:38:19","guid":{"rendered":"https:\/\/justfineinfotech.com\/introducing-claude-opus-5-5\/"},"modified":"2026-09-23T13:38:20","modified_gmt":"2026-09-23T13:38:20","slug":"introducing-claude-opus-5-5","status":"publish","type":"post","link":"https:\/\/justfineinfotech.com\/fr\/introducing-claude-opus-5-5\/","title":{"rendered":"Introducing Claude Opus 5.5"},"content":{"rendered":"\n\n\n\n\n\n\n\n\n\n\n\n\n
Opus 5.5<\/th>\nFable 5.1<\/th>\nOpus 5<\/th>\nGPT-6 Astra<\/th>\nGPT-5.6 Sol<\/th>\n<\/tr>\n<\/thead>\n
Agentic codingTerminal-Bench 4.0\u00b9<\/th>\n66.4%<\/td>\n55.8%<\/td>\n52.3%<\/td>\n57.9%<\/td>\n37.3%<\/td>\n<\/tr>\n
Agentic codingFrontierCode v1.1 (Main)<\/th>\n54.4%<\/td>\n50.3%<\/td>\n48.0%<\/td>\n53.3%<\/td>\n47.5%<\/td>\n<\/tr>\n
Agentic codingCursorBench 4.0<\/th>\n57.8%<\/td>\n51.8%<\/td>\n46.6%<\/td>\n\u2014<\/td>\n41.7%<\/td>\n<\/tr>\n
Knowledge workGDPval-AA v2.1<\/th>\n1846<\/td>\n1735<\/td>\n1708<\/td>\n1542<\/td>\n1588<\/td>\n<\/tr>\n
Business workflowsAutomationBench\u00b2<\/th>\n40.0%<\/td>\n31.4%<\/td>\n26.9%<\/td>\n41.4%<\/td>\n28.8%<\/td>\n<\/tr>\n
Multidisciplinary reasoningHumanity’s Last Exam<\/th>\n67.7%with tools<\/td>\n65.6%with tools<\/td>\n63.6%with tools<\/td>\n57.2%with tools<\/td>\n\u2014<\/td>\n<\/tr>\n
Agentic scientific researchTerminal-Bench-Science 0.1\u00b3<\/th>\n58.7%<\/td>\n52.6%<\/td>\n29.0%<\/td>\n64.6%<\/td>\n22.4%<\/td>\n<\/tr>\n
Computer useOSWorld 2.0<\/th>\n81.8%partial<\/td>\n80.7%partial<\/td>\n74.0%partial<\/td>\n\u2014<\/td>\n\u2014<\/td>\n<\/tr>\n
Visual chart recognitionChartography<\/th>\n89.0%with tools<\/td>\n88.4%with tools<\/td>\n83.4%with tools<\/td>\n\u2014<\/td>\n\u2014<\/td>\n<\/tr>\n<\/tbody>\n<\/table>
Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort. Terminal-Bench 4.0 results are reported for Claude Opus 5.5 at xhigh effort and GPT-6 Astra at high effort, as reported by OpenAI; these represent each model\u2019s highest score. Claude Opus 5.5 was evaluated with its production safeguards enabled. When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5. This likely reduces Claude Opus 5.5\u2019s performance on these benchmarks.1 Terminal-Bench 4.0: The standard error is \u00b12.6 pts for Claude Opus 5.5 and \u00b11.6\u20132 pts for the other Claude models. The public leaderboard (5 trials\/task, Claude Code harness) reports Claude Opus 5 at 51.8%; our setup reproduces it at 52.3%, within noise. GPT-6 Astra and GPT-5.6 Sol figures are as reported by OpenAI.2 AutomationBench: AutomationBench results were run and reported by Zapier. These runs were performed without fallback models, so safeguard interventions were considered failures\u2014this resulted in a lower score than Claude Opus 5.5 would achieve in practice. Claude Opus 5.5 results come from Zapier\u2019s own evaluation during early access. Results for Opus 5, GPT-5.6 Sol, and GPT-6 Astra come from Zapier\u2019s public leaderboard.3 Terminal-Bench-Science 0.1: The standard error is \u00b13.5\u20135 pts per model. The public leaderboard (3 trials\/task, Claude Code harness) reports Claude Opus 5 at 30.0%; our setup reproduces it at 29.0%, within noise. The GPT-6 Astra figure is as reported by OpenAI.<\/figcaption>\n\n\n\n\n\n\n
Prices per 1M tokens<\/th>\nClaude Opus 5.5<\/strong><\/th>\nClaude Opus 5<\/th>\n<\/tr>\n
Cache reads<\/td>\n$0.20<\/strong><\/td>\n$0.50<\/td>\n<\/tr>\n
Input tokens<\/td>\n$4<\/strong><\/td>\n$5<\/td>\n<\/tr>\n
Output tokens<\/td>\n$20<\/strong><\/td>\n$25<\/td>\n<\/tr>\n
Cache writes<\/td>\n$5<\/strong><\/td>\n$6.25<\/td>\n<\/tr>\n<\/tbody>\n<\/table>
Terminal-Bench 4.0Accuracy vs Cost<\/figcaption>

Terminal-Bench 4.0 measures how well a model can complete complex, multi-step professional tasks within a command line interface. Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost. It matches GPT-6 Astra at about 40% of the cost<\/p>

FrontierCode v1.1, main setAccuracy vs Cost<\/figcaption>

FrontierCode measures whether an agent\u2019s code changes would be merged. At default effort (medium), Opus 5.5 scores 54.6%, higher than all other models, beating GPT-6 Astra\u2019s top score (53.3%) for about a fifth of the cost per task<\/p>

CursorBench 4.0Accuracy vs Cost<\/figcaption>

CursorBench evaluates coding agents on ambiguous, multi-file tasks taken from real Cursor sessions. At default effort (medium), Opus 5.5 scores 52.5%, compared to 51.8% for Fable 5.1 (max) and 46.6% for Opus 5 (max). It beats GPT-5.6 Sol\u2019s top score (41.7%) by 11 points for about a third of the cost per task<\/p>\n

Quote<\/p>\n

\n

\u201cDevelopers want agents that can take on real software work and finish it. In our testing across GitHub Copilot CLI and VS Code, Claude Opus 5.5 used among the fewest tokens and steps we measured. In VS Code, it solved more terminal tasks than Opus 5 in less than half the steps. More than making individual tasks more efficient, it\u2019s making developers\u2019 bigger projects more achievable.\u201d<\/p>\n<\/blockquote>\n

CompanyGitHub
\nAuthorMario Rodriguez, Chief Product Officer

GDPval-AA v2.1Elo vs Cost<\/figcaption>

Artificial Analysis\u2019s GDPval-AA v2.1 evaluates agents on real-world professional work across 44 occupations. At max effort, Opus 5.5 scores 1846 Elo, where Fable 5.1 scores 1735 and Opus 5 scores 1708. At default effort (medium), Opus 5.5 beats GPT-6 Astra at max effort for about a fifth of the cost per task<\/p>

AutomationBenchAccuracy vs Cost<\/figcaption>

AutomationBench, built by Zapier, tests whether an agent can carry out real business workflows across many connected apps. Opus 5.5 outscores Opus 5 and GPT-5.6 Sol at every effort level<\/p>

WANDRAccuracy vs Cost<\/figcaption>

Perplexity\u2019s WANDR benchmark measures agents on large data collection tasks. Opus 5.5 outperforms Fable 5.1 and Opus 5 at a lower cost per task4<\/p>\n

4WANDR: Claude models were run with offline versions of the web search and web fetch tools, programmatic tool calling, code execution, and a 980k-token task budget. This differs from Perplexity\u2019s published setup, scores are not directly comparable across the two and we only show models scored under the same conditions<\/p>\n

Quote<\/p>\n

\n

\u201cEven at its lowest effort setting, Claude Opus 5.5 caught 72% of known bugs in our code reviews to Opus 5\u2019s 56% at high effort, with fewer false alarms and a fraction of the output. On US consulting analysis, low thinking effort matched its higher thinking settings on half the output and passed our quality checks. When more lower thinking efforts are deployed in production, that\u2019s client-ready work delivered efficiently.\u201d<\/p>\n<\/blockquote>\n

CompanyDeloitte Consulting LLP
\nAuthorCarl Bennett, CIO
\nClaude Opus 5<\/p>\n

Please explain the issue to me<\/p>\n

WhatIfound<\/strong><\/p>\n

Theextradropisn\u2019tthefreetier\u2014it\u2019saregressionin0552feb(\u201cextractbillingperiodboundariesintoperiods.py\u201d,labelled\u201cNobehaviourchange\u201d)<\/p>\n

Thebug<\/strong><\/p>\n

aggregate.monthly_usage<\/code>usedtodoahalf-openinterval:<\/p>\n

start,end=month_start(year,month),next_month_start(year,month)ifstart<=e.ts<end:<\/code><\/pre>\n
Itnowdoes:<\/p>\n
ifperiod.contains(e.ts)<\/code><\/pre>\n
andinperiods.py<\/code>:<\/p>\n
@propertydeflast_day(self)->datetime:returnnext_month_start(...)-timedelta(days=1)#2026-08-3100:00:00UTCdefcontains(self,ts):returnself.first_day<=ts<=self.last_day<\/code><\/pre>\n
last_dayisadatetimeatmidnight,notawholeday.SubtractingonedayfromSep100:00givesAug3100:00,andthecomparisonisinclusive,sotheperiodeffectivelyendsatthefirstinstantofthe31st.Everyeventtimestampedaftermidnightonthelastdayofthemonthissilentlydropped\u2014roughlyonedayofusageperaccount,permonth<\/p>\n
Thenewtest_periods.pydoesn\u2019tcatchitbecauseitonlycheckslast_day.date()(whichiscorrectly2026-08-31)andtestscontainsatAug100:00,Aug1512:00,Jul3123:59andSep100:00\u2014itneverprobesanywhereinsidethe31st<\/p>\n
Claude Opus 5.5<\/p>\n
Please explain the issue to me<\/p>\n
Theextradropisabuginthebillingrefactor<\/strong><\/p>\n
Thefree-tierchangeaccountsforonly$1.50ofacme\u2019sAugustdrop.Theother$9.92comesfromabugincommit0552feb.Thatcommitwaslabelled\u201cNobehaviourchange,\u201dbutitstopscountingusagefromthelastdayofthemonth<\/p>\n
Whatchanged<\/strong><\/p>\n
Beforethemerge,aggregate.py<\/code>usedahalf-openinterval:<\/p>\n
start<=e.ts<next_month_start(year,month)#[Aug100:00,Sep100:00)<\/code><\/pre>\n
TherefactormovedthisintoPeriod.contains<\/code>andmadetheintervalclosedonthewrongboundary:<\/p>\n
last_day=next_month_start(...)-timedelta(days=1)#=2026-08-31T00:00:00Zreturnself.first_day<=ts<=self.last_day<\/code><\/pre>\n
last_dayisatimestampatmidnight,notawholeday.Soeveryeventafter2026-08-31T00:00:00Zisexcluded.TheSeptember<\/a>periodstartsSep1,soitdoesn\u2019tpicktheseeventsupeither.Theusageisn\u2019tmovedtoanothermonth;it\u2019sneverbilledatall<\/p>\n
Claude Opus 5<\/p>\n
Summarize this Slack thread for my manager in three bullet points:<\/p>\n
[09:41] Dana Osei: I\u2019ll tell them. I\u2019ll also confirm nobody else reads the Friday file.
\n[09:43] Marcus Webb: I can move the schedule, it\u2019s a one-line cron change. But not until Dana confirms.
\n[10:02] Dana Osei: Finance says the Friday file is only used Monday. Sunday night is fine<\/p>\n
Fridayordersexport\u2014#data-platform,Tuesday<\/strong><\/p>\n