Opus 5.5
Claude Opus 5.5 is Anthropic's September 2026 flagship Opus-class model for long-horizon agentic coding, computer use and knowledge work — 1M-token context, native image input, adaptive (non-disableable) reasoning at four effort tiers, and $4/$20 per 1M tokens (40% cheaper per typical task than Opus 5).
Empirical Evaluation Results
Architectural Profile & Capabilities
Claude Opus 5.5 is Anthropic's flagship successor to Claude Opus 5, released around 22 September 2026 — roughly two months after Opus 5 (July 2026) and three weeks after Claude Fable 5.1. Anthropic's own platform documentation confirms the model, its positioning ("our most capable Opus model yet for coding, agents, and knowledge work"), its $4 / $20 per-million input/output pricing, and a $0.20/M cache-read rate (5% of standard input price). The company states Opus 5.5 "costs less per token than Opus 5 and uses fewer tokens per task," and that it requires less compute to serve than Opus 5, with the resulting efficiency passed through as a ~40% lower cost per typical workload at default settings. A 2x batch tier ($2/$10) and a Fast mode (~2.5x speed at $8/$40) round out the serving options. Capability envelope: Opus 5.5 is a 1,000,000-token-context, image-input, text-output model optimized for multi-step changes in large codebases, terminal/agent harnesses, and computer-use loops. On the independent side, Artificial Analysis places Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) at the top of its Intelligence Index v4.3.2 with a score of 58, and the model appears as the reference configuration on the GDPval-AA v2.1 economics-oriented leaderboard (reported ~1,846 Elo in secondary coverage). Vals AI's suite evaluation reports Opus 5.5 among the leaders on SWE-bench Verified, Terminal-Bench and LiveCodeBench, and Vibe Code Bench v1.1 ranks it #2 (behind Claude Fable 5 at 90.35%). Secondary reporting circulates Terminal-Bench 4.0 66.4%, FrontierCode v1.1 54.4%, CursorBench 4.0 57.8%, OSWorld 2.0 81.8%, Terminal-Bench-Science 0.1 58.7% and AutomationBench 40.0%; these figures are vendor-adjacent, not Anthropic-published, and should be re-verified against the official model card before procurement decisions. Notably, Anthropic's own team has publicly flagged likely SWE-bench test-set contamination, and Opus-class scores on SWE-bench Verified have saturated in the 96–97% band (Vals AI run: Opus 5 at 97.00%), so treat that benchmark as a weak discriminator at this generation. Reasoning and control surface: Opus 5.5 inherits the Opus 5 family's adaptive thinking plus an `effort` parameter (medium / high / xhigh / max) that trades reasoning-token spend against thoroughness; thinking cannot be cleanly disabled and reasoning tokens are billed as output at the expensive rate. The published Opus 5 specification (1M context, 128K max output, 300K max output via Batch API beta, adaptive thinking, default effort high) describes the platform family into which 5.5 fits; long-context premium tiers materially change effective unit economics (>200K-token prompts have historically moved to $10/$37.50 on Opus-class). Architecture and parameter count remain undisclosed by Anthropic — no MoE, expert-routing, attention-variant (GQA/MQA/MLA) or positional-encoding disclosures exist for this model, so any Mixture-of-Experts or active-parameter claims are speculation. Operational summary: pick Opus 5.5 as the default Opus-class workhorse for agentic engineering and document-heavy knowledge work; step up to Claude Fable 5.1 only when maximum reasoning depth justifies the higher per-task cost. Budget for reasoning-token amplification (2–4x billed output versus visible answer length at high/max effort), verify long-context pricing tiers and interleaved-thinking beta headers on your target surface (Anthropic API vs Bedrock/Vertex vs OpenRouter routing), and re-benchmark on your own harness rather than trusting saturated public suites.
Anthropic does not disclose Opus 5.5's parameter count, layer configuration, or attention internals; no primary source confirms sparse Mixture-of-Experts, expert count, top-k routing, or active-parameter budget, so the defensible classification is a proprietary dense autoregressive Transformer family (Claude 5.x/5.5 lineage) rather than MoE or a diffusion LLM. Confirmed platform behavior: 1,000,000-token context window (single 1M variant, no smaller-context SKU), 128K max output tokens (300K via Batch API beta), adaptive thinking that is enabled by default and not fully disableable, an `effort` control with medium/high/xhigh/max tiers (default high), interleaved/extended-thinking blocks on agent turns, and native image input with text output. Anthropic states Opus 5.5 requires less compute per token to serve than Opus 5, implying an inference-efficiency improvement (distillation, quantization, or re-architected serving) rather than a parameter-count disclosure. Tokenizer/vocabulary, positional encoding, and KV-cache/attention implementation (GQA vs MLA vs sliding-window hybrid) are all undisclosed; 'less compute to serve' plus the 1M window is consistent with a KV-cache-efficiency mechanism such as grouped-query or latent attention, but this is inference, not verified spec.
Recommended Workloads & Primary Use Cases
- •Benchmark opacity and saturation: most circulating Opus 5.5 numbers (Terminal-Bench 4.0 66.4%, FrontierCode 54.4%, OSWorld 2.0 81.8%, GDPval-AA 1,846 Elo) originate from secondary reporting, not an Anthropic model card; meanwhile Anthropic has itself flagged SWE-bench test memorization and the suite is saturated at 96-97%, so treat headline coding scores as weakest-evidence, highest-inflation signals.
- •Reasoning tokens are always-on and billed at output rates: adaptive thinking cannot be disabled, and effort=high/xhigh/max can multiply billed output 2-4x over visible answer length. A 'cheap' $20/M output rate can therefore produce a per-task cost far above Sonnet-class or Fable-class alternatives at lower effort settings; pin effort and max_tokens explicitly per workload class.
- •Long-context economics and effective-context decay: the 1M window is real at the API level, but Opus-class long-context pricing tiers historically switch above ~200K tokens ($10/$37.50 on Opus 4.6) and retrieval accuracy degrades well before the nominal ceiling. Cache reads are the only cheap path ($0.20/M at 5% of input); uncached 1M-token prompts are expensive on every turn.
- •Output-length truncation and beta-feature drift: 128K max output (300K only on the Batch API beta) will truncate long agent transcripts and generated specs; interleaved thinking, effort tiers and extended output ceilings require specific headers/surfaces, and OpenRouter third-party routing may not honor effort, thinking or caching semantics identically to the first-party Anthropic API.
- •Prompt-format and behavioral sensitivities: post-Opus 5 Claude models respond differently to heavy prescriptive prompting than 4.x-era models — Anthropic now recommends lighter scaffolding and explicit model/effort selection. Expect effort-dependent verbosity drift, occasional overconfidence in self-verification, and refusal-pattern changes on cyber/bio-adjacent prompts versus Opus 4.x; do not port 4.x system prompts unchanged.
- •Versioning churn risk: Opus 5.5 shipped within a dense release cadence (Opus 4.6 Feb 2026, Opus 5 Jul 2026, Fable 5.1 early Sep 2026, Opus 5.5 late Sep 2026), so deprecation and pricing-tier shifts are plausible within a single quarter; pin model IDs and re-validate before contract renewal.
Calls to Opus 5.5 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Anthropic.
More Models from Anthropic
Compare other engines available in this laboratory.
Anthropic's flagship 1M-context frontier multimodal foundation model, engineered for deep reasoning, autonomous coding agents, and complex long-horizon system execution.
Anthropic's most capable model (released May 28, 2026; knowledge cutoff Jan 2026).
Successor to Opus 4.6 at the same price (released Apr 16, 2026).