Claude Sonnet 5.5

Anthropic does not disclose architecture type, parameter counts, expert or layer counts, attention design or training compute for any Claude model, so those fields are null. The system card says only that Sonnet 5.5 was pretrained on a proprietary mix of public web data (crawled by ClaudeBot, which follows robots.txt), public and private datasets and synthetic data generated by other models, then post-trained toward the values in Claude's constitution. What is disclosed operationally: a 1M-token context window, 128K synchronous output (300K through the Message Batches API with a beta header), adaptive thinking that is on by default and steered by five effort levels, and the same tokenizer as Sonnet 5, the one introduced with Opus 4.7, which fits roughly 555,000 words into 1M tokens. Thinking text is omitted by default and returned only as a summary when requested. Anthropic describes Sonnet 5.5 as less capable than Opus 5.5 on most of its evaluations and not at the frontier of its models' capabilities.

Coding (9.7): SWE-bench Pro 81.3%, SWE-bench Multilingual 90.3%, SWE-bench Multimodal 54.3%, DeepSWE v1.1 71.0% and Terminal-Bench 4.0 70.6% (vendor-reported, max effort). Independent data backs the jump: Artificial Analysis measured 63.6% on Terminal-Bench 4.0 at max effort, above Opus 5.5 (59.6%) in the same testing; Vals AI ranks it first on Vibe Code Bench and Code Migration; LMArena's WebDev arena places it third (1786); Cursor measured 55.5% on CursorBench 4.0, about two points behind Opus 5.5; and Proximal's FrontierSWE v2 puts it third (61.9%), behind GPT-6 Astra and Opus 5.5. Reasoning (9.3): Humanity's Last Exam 56.9% without tools and 64.5% with tools (vendor); Artificial Analysis measures 55.0% without tools against 61.4% for Opus 5.5, and its Intelligence Index ranks Sonnet 5.5 second of 224 models in its class. Math (9.2): ArXivMath (August 2026) 86.8% without tools and 95.2% with tools at max effort on Anthropic's setup, above Fable 5.1 (82.9% / 92.1%) and below Opus 5.5 (91.2% / 96.9%); no MathArena or AIME figure is published yet. Agentic (9.6): OSWorld 2.1 80.1% partial and 43.5% strict, Toolathlon-Verified 77.8% Pass@1 (level with Opus 5.5 and Fable 5.1), AutomationBench 44.7% (run by Zapier), GDPval-AA v2.1 Elo 1844 and AA-Briefcase v1.1 1811 (run by Artificial Analysis). Long context (9.2): 1M tokens at standard rates; AA-LCR 82.7% (Opus 5.5 84.7%); ProgramBench 79.7% with episodes up to the full window, against 91.2% for Opus 5.5. Multilingual (9.2): Global MMLU 92.1% across 42 languages and MILU 91.6% across 11 Indic languages, up from 89.2% and 89.3% for Sonnet 5. Vision (9.0): Chartography 61.6% without tools (Sonnet 5: 15.6%) and 90.2% with an image-cropping tool, BenchCAD Vision2Code voxel IoU 0.747 without tools and 0.963 with them, and Anthropic says it is the first Sonnet to beat Pokemon Red from screenshots alone; LMArena's vision board, however, rates it 1268, barely above Sonnet 5's 1265. Document/OCR (9.0): OfficeQA 76.9% and OfficeQA Pro 65.6%, and GDP.pdf all-pass 25.8% in Artificial Analysis testing (Sonnet 5 13.2%, Opus 5.5 26.2%). Creative writing (8.4): Anthropic and early testers report clearer, more direct writing, but the system card rates its warmth and humor slightly below Sonnet 5's. Instruction following (9.1): fewer instruction-following failures than Sonnet 5 in Anthropic's audit, but it adds unrequested tests and docs and, at low or medium effort, can stop to check in before a coding task is done. Function calling (9.3): it batches tool calls more than Sonnet 5 and had the fewest failed tool calls in Base44's comparison; forced tool_choice is gone, and Anthropic's prompting guide notes occasional wrong-case tool names or near-miss parameter names. Safety calibration (9.2): benign over-refusal of 0.02% on the API (Sonnet 5: 0.59%) and much stronger prompt-injection resistance, offset by cyber classifiers that Anthropic says will raise refusals on some benign security work. Real-time data (7.0): June 2026 cutoff; current information requires the web search tool.

Rows sourced to Anthropic are vendor-reported at max effort, averaged over five trials unless noted; competitor figures in those rows are Anthropic's compilation. Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release deployment that Anthropic says had a since-fixed structured-output bug; Artificial Analysis's current data shows 1840 and 1824. Terminal-Bench 4.0 is a different version from the Terminal-Bench 2.x scores stored under the terminal_bench key, so it stays in this table only. The hle key holds the vendor no-tools figure. LMArena figures come from the October 2, 2026 snapshot (WebDev October 1), where only the xhigh variant is listed on the text boards. GPQA Diamond, MMLU-Pro, AIME, SWE-bench Verified, MMMU, tau-bench and MRCR have no published Sonnet 5.5 figure and are null.

Anthropic rates Sonnet 5.5 "Fast" in its lineup, between Haiku 4.5 and Opus 5.5, and says it generates output 30%+ faster than Sonnet 5. Artificial Analysis confirms the direction: at the API-default high effort it measures 97.6 output tokens/sec and 12.7 seconds to the first answer token, against 66.0 tokens/sec for Sonnet 5 at high in the same data. Effort decides the wait. At medium, the default in Claude Code and the Claude apps, the first answer token arrives in about 1.2 seconds; at max effort generation runs at 134 tokens/sec, but the first answer token takes about 452 seconds because of long thinking. Artificial Analysis also rates it very verbose: 420M output tokens to run its Intelligence Index at max effort, against a median of 81M. For interactive paths, use low or medium effort, or between_tools to skip up-front thinking. Fast mode is not offered on Sonnet models.

Proprietary and API-only: no open weights, no self-hosting, no fine-tuning. It is available from launch day on the Claude API, Claude Platform on AWS and Microsoft Foundry (all as claude-sonnet-5-5), Amazon Bedrock (anthropic.claude-sonnet-5-5) and Google Cloud (claude-sonnet-5-5), and OpenRouter lists it at the same rates. First-party routing is global by default, with US-only inference at 1.1x. Bedrock and Google Cloud offer regional endpoints at a 10% premium and set their own lifecycle dates; Anthropic commits to no retirement before 2027-09-28 on the platforms it operates. Zero data retention is available, as for Opus 5.5 and Sonnet 5. Platform differences matter for agents: computer use on the Claude API and Google Cloud requires the computer_toolset_20260801 toolset, while Bedrock still accepts the older computer_20251124 tool, and server-side fallback (fallbacks: "default") is a Claude API beta, so other platforms need the SDK middleware or a custom retry.

Sonnet 5.5 is governed by Anthropic's Responsible Scaling Policy. The system card treats it as meeting the CB-1 and Autonomy-1 thresholds, applies the matching mitigations, and places it below CB-2 and Autonomy-2 by bounding it with Opus 5.5's results. It is the first Sonnet deployed with cyber classifiers that enforce the Opus 5 and Opus 5.5 policy: vulnerability discovery in source code is allowed, in compiled binaries it is blocked. Flagged cyber and frontier-LLM-development requests fall back to Sonnet 5 in Claude apps and, when opted in, on the API; biology, weapons and reasoning-extraction blocks have no fallback. Anthropic tells users to expect more refusals than on Sonnet 5 even on some benign security tasks, and the Cyber Verification Program does not cover Sonnet 5.5 yet. On harm evaluations it keeps a 95.6% harmless-response rate on the API (Sonnet 5: 96.7%) with 0.02% over-refusal on benign prompts (Sonnet 5: 0.59%). Prompt-injection resistance improved sharply: Gray Swan's Shade attacker succeeded on 3.01% of coding attempts (Sonnet 5: 19.47%) and 0.07% of computer-use attempts (2.25%), matching Opus 5.5 and Fable 5.1 in GUI settings. The alignment audit finds it improves on or matches Sonnet 5 on most measures, with Opus 5.5 still stronger overall; it hallucinates more than Opus 5.5 and is slightly more likely than Sonnet 5 to state a wrong answer. API data is not used for training without express permission, ZDR and HIPAA-ready access are available, and Anthropic lists SOC 2 Type I and II, ISO 27001:2022 and ISO/IEC 42001:2023. Its output carries Anthropic's invisible text watermark on first-party and cloud-partner surfaces.

Official SDKs cover Python, TypeScript, Java, Go, Ruby, C# and PHP, each with a named model constant for Sonnet 5.5. It runs in Claude Code (default effort medium), Claude Managed Agents (no changes beyond the model name) and every Claude app, including the Free plan, and OpenRouter lists it. Launch partners quoted by Anthropic include Epic Games, Every, CodeRabbit, Base44, Unity, Slack, Zendesk, Balyasny Asset Management, Box, Lovable and Atlassian (Rovo Agents), and the system card cites results run by Cursor (CursorBench), Cognition (FrontierCode), Proximal (FrontierSWE) and Zapier (AutomationBench). Early usage signal: 3,145 LMArena text votes and 1,531 WebDev votes for the xhigh variant within days of launch, and Artificial Analysis lists eight API providers serving it.