Claude Opus 5.5

Anthropic does not disclose architecture type, parameter counts, expert counts, layer counts or attention design for any Claude model, so those fields are null here. What is disclosed: a 1M-token context window, 128K synchronous output (300K through the Message Batches API with a beta header), adaptive thinking that is always on and steered only by the effort parameter (low to max, default medium), and the current Claude tokenizer introduced with Claude Opus 4.7 (Anthropic puts 1M tokens at roughly 555,000 words on it). Thinking text is omitted by default and returned as a summary on request. The system card describes training on a proprietary mix of public web data (crawled by ClaudeBot, which honors robots.txt), public and private datasets, opted-in user data and synthetic data, followed by post-training toward Claude's constitution, with a June 2026 knowledge cutoff. It also notes the model outputs text only and that output quality varies by language.

Coding (10.0): SWE-bench Pro 89.9% (Fable 5.1: 81.2%), SWE-bench Multilingual 93.9%, SWE-bench Multimodal 61.4%, DeepSWE v1.1 74.2%, Terminal-Bench 4.0 66.4% at xhigh effort, all vendor-reported. Third-party runs agree on direction: Cursor measured 57.8% on CursorBench 4.0, first on its leaderboard; Cognition's FrontierCode v1.1 puts it first at 54.6% (best effort); Artificial Analysis measured Terminal-Bench 4.0 at 59.6%, level with GPT-6 Astra. Proximal's FrontierSWE v2 has it second (62.3%) behind GPT-6 Astra (65.5%). Reasoning (9.9): HLE 64.4% without tools and 67.7% with tools (vendor); Artificial Analysis measured HLE at 61.4%; ARC Prize verified 93.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1 at high effort. No ARC-AGI-3 score has been published. Math (9.7): ArXivMath (August 2026) 91.2% without tools and 96.9% with tools on Anthropic's setup, ahead of Fable 5.1 (82.9% / 92.1%); MathArena's independent overall ranking is currently led by OpenAI's GPT-6 line (GPT-6.1 Sol 90.1%, GPT-6 Astra 88.0%, GPT-6 Sol 84.9%). Agentic and tool use (10.0): OSWorld 2.0 81.8% partial and 48.7% strict (Fable 5.1: 80.7% / 42.8%), AutomationBench 40.0% on Zapier's leaderboard (GPT-6 Astra 41.4%), GDPval-AA v2.1 1846 and AA-Briefcase 1822, both run by Artificial Analysis and both first. Multi-agent teams of 1 to 100 agents outperform Opus 5 and Fable 5.1 at every team size on Anthropic's knowledge-base and Lean tasks. Long context (9.7): ProgramBench 91.2% hidden-test pass rate with episodes up to the full 1M window (Fable 5.1: 87.6%); no MRCR figure is published. Multilingual (9.1): Global MMLU 94.3% across 42 languages and MILU 93.1% across 11 Indic languages and English, slightly ahead of Fable 5.1. Vision (9.8) and document/OCR (9.5): Chartography 64.4% without tools against 44.8% for Fable 5.1 and 29.8% for Opus 5, 89.0% with tools; BenchCAD Vision2Code voxel IoU 0.730 without tools and 0.962 with tools. On OfficeQA Pro it scores 67.7%, just below Fable 5.1 (69.0%). Creative writing (8.6): partners describe clearer, more direct writing, but the system card rates creativity and intellectual depth slightly weaker than Opus 5. Instruction following (9.3): responds well to specific instructions, but unattended agents can stop early after progress reports, and early snapshots followed instructions hidden in text users pasted into prompts. Function calling (9.3): strict tool use, structured outputs, inline tool definitions and parallel tools, but no forced tool use, and Toolathlon Verified dipped to 77.8% from 80.6% on Opus 5. Safety calibration (8.9): best Claude score to date on Anthropic's automated behavioral audit and 0.03% over-refusal on benign API prompts, offset by a 1.5-point lower harmless-response rate than Opus 5 and broader classifier coverage. Real-time data (7.0): June 2026 cutoff; current information requires the web search tool.

Rows sourced to the system card use Anthropic's standard configuration (adaptive thinking at max effort, averaged over five trials) with production safeguards enabled; when a classifier intervened, a fallback model completed the request (2.5% of Terminal-Bench 4.0 requests). Rows marked AA-run, Cursor-run, Cognition-run, Proximal-run or Zapier leaderboard were measured by those organizations and reported in the system card; competitor figures are Anthropic's compilation from developers' published results. Anthropic's announcement labels the computer-use row OSWorld 2.1; the system card calls it OSWorld 2.0 on the September 10, 2026 task files with a new harness, so it is not comparable with launch figures for earlier models. Arena figures are the October 2, 2026 snapshot, when the high-effort text entry had about 4,500 votes and the coding entry about 1,060; the catalog's other models carry a September 13 snapshot. Terminal-Bench 4.0 is a different version from the 2.x scores stored under the terminal_bench key, so it stays in this table. Anthropic published no SWE-bench Verified, GPQA Diamond, MMLU-Pro, AIME, MMMU, tau-bench or MRCR figure for Opus 5.5, so those keys are null; ARC Prize lists no ARC-AGI-3 score yet.

Artificial Analysis measures Opus 5.5 at the default medium effort at about 73 output tokens per second with roughly 24 seconds to the first answer token, a figure that includes thinking time. Speed and wait both depend on effort: about 76 tokens/s and 12 seconds at low, 74 tokens/s and 37 seconds at high, 79 tokens/s and about 2.3 minutes at xhigh, and 93 tokens/s with more than 11 minutes before the answer at max on its benchmark prompts. For comparison, Artificial Analysis lists Fable 5.1 at 54 tokens/s at high effort and Sonnet 5.5 at 139 tokens/s at max. Anthropic rates Opus 5.5 latency as "Moderate" (between Sonnet 5.5 and Fable 5.1) and says it generates output more than 30% faster than Opus 5 while finishing tasks in fewer tokens. Two caveats from Anthropic's prompting guide: at the same effort level it thinks more per turn than Opus 5, especially at xhigh and max, and in multi-turn chat it sometimes revisits earlier answers, which adds latency on follow-ups. Fast mode (research preview, Claude API only) delivers up to 2.5x output tokens per second at $8/$40; it does not reduce time to first token. The Batch API halves the cost for latency-insensitive work.

Anthropic says typical workloads cost about 40% less than on Opus 5 at default settings. Refusals that arrive before any output are billed for the bio, frontier_llm and reasoning_extraction categories and count against rate limits.

Proprietary and API-only: no open weights, no self-hosting, no fine-tuning. It is available on the Claude API (claude-opus-5-5), Amazon Bedrock (anthropic.claude-opus-5-5), Google Cloud, Microsoft Foundry and Claude Platform on AWS, all from launch day. Fast mode runs only on the first-party Claude API (including Claude Managed Agents). First-party routing is global by default; US-only inference costs 1.1x, and Bedrock and Google Cloud offer regional endpoints at a 10% premium over global. Platform differences matter for migration: the older computer_20251124 computer-use tool still works on Bedrock but is rejected on the Claude API and Google Cloud. Anthropic commits to keeping it available on its own platforms until at least 2027-09-22.

Opus 5.5 is governed by Anthropic's Responsible Scaling Policy and Frontier Compliance Framework. The system card treats it as having CB-1 but not CB-2 chemical and biological capabilities, finds it below the automated AI R&D threshold (consistent with METR's external testing), and judges overall misalignment risk low. Because its biology and cyber capabilities sit close to Claude Mythos 5.1, it ships with Fable 5.1-level safeguards: a research-biology classifier (fallback to Opus 5), cyber classifiers that allow source-code vulnerability finding but block high-risk dual-use work (fallback to Opus 4.8), frontier-LLM development safeguards (fallback to Opus 5), and anti-distillation classifiers with no fallback. Fallbacks apply in Claude apps and for API developers who opt in. Anthropic calls it the strongest Claude model to date on its automated behavioral audit, with the lowest rate of destructive actions among recent models, a 0.03% over-refusal rate on benign API prompts and the highest AA-Omniscience factuality net score among Claude models (0.58, within error bars of Mythos 5 and 5.1). The card is candid about weaknesses: a harmless-response rate 1.5 points below Opus 5 on the API without a system prompt, lower honesty under pressure (MASK) than Opus 5, early snapshots that followed instructions hidden in user-pasted text (mitigated in the final model and products), and internal signs that it suspected evaluation in up to 36% of transcripts. It is not one of Anthropic's 30-day-retention Covered Models, so zero-data-retention arrangements apply, retained data is not used for training without permission, and HIPAA-ready access is available on the Claude API. Anthropic lists SOC 2 Type I and II, ISO 27001:2022 and ISO/IEC 42001:2023 for its commercial products.

Official SDKs cover Python, TypeScript, Java, Go, Ruby, C# and PHP, plus the ant CLI. Opus 5.5 is supported in Claude Code, the Claude Agent SDK and Claude Managed Agents, and is included on Claude Pro and Max. Early-access partners quoted in Anthropic's announcement include GitHub, Lovable, Kiro, Clio, Box, Hebbia, Hex, LexisNexis, Thomson Reuters Labs, Deloitte, Ramp, Stripe, Spotify, Optiver and Factory. Third-party evaluators moved quickly: Artificial Analysis, Cursor, Cognition, Proximal, Zapier and ARC Prize all published results within days of launch, and the model reached the top five of the LMArena text board and first place on its WebDev board within ten days.