Claude Opus 4.5

Claude Opus 4.5 is a frontier large language model released by Anthropic on November 24, 2025, the first model in the Opus 4.x line. Anthropic positioned it as its most capable model at release for coding, agentic tasks, and computer use, and reported state-of-the-art results on several software-engineering and agentic benchmarks (Source: anthropic.com). It was the model underlying Anthropic's revenue growth from late 2025 through early 2026, and is the reference point for what Dario Amodei later described as an "80x revenue surprise."

Lineage and history

Claude Opus 4.5 opened the Opus 4.x line, succeeding the Opus 4 and Opus 4.1 models of the preceding generation. Anthropic released it alongside updates to the Claude Developer Platform, Claude Code, and the consumer Claude apps (Source: anthropic.com). Within the Opus line it was followed by Claude Opus 4.6 (February 5, 2026), Claude Opus 4.7 (April 16, 2026), and Claude Opus 4.8 (May 28, 2026). The AGI Timelines page tracks the progression from Opus 4.5 to Opus 4.6 to Opus 4.7 as a measurable indicator of frontier capability change.

Architecture and training

Anthropic has not disclosed the parameter count of Opus 4.5, consistent with its practice across the Opus 4.x family. The model operates with a 200,000-token context window (Source: anthropic.com). It is multimodal over text and images, with Anthropic reporting improved vision, reasoning, and mathematics relative to its predecessors (Source: anthropic.com).

Opus 4.5 introduced an effort parameter on the Claude API, a control that lets developers trade output-token usage against capability: at a medium effort level Anthropic reported that Opus 4.5 matched Sonnet 4.5's best SWE-bench Verified score while using 76% fewer output tokens, and at its highest effort level exceeded Sonnet 4.5 by 4.3 percentage points while using 48% fewer tokens (Source: anthropic.com). The release was accompanied by platform features for longer-running agents, including context compaction and a memory tool, which Anthropic reported raised performance on a deep-research evaluation by almost 15 percentage points (Source: anthropic.com). Anthropic's published benchmark methodology used a 64K thinking budget, a 200K context window, and default high effort, averaged over five trials (Source: anthropic.com). As with other Claude models, Anthropic's alignment approach combines reinforcement learning from human feedback with its Constitutional AI / RLAIF methods, detailed in the Opus 4.5 system card (Source: anthropic.com).

Capabilities and benchmarks

Anthropic reported that Opus 4.5 was the first model to exceed 80% on SWE-bench Verified, a benchmark of real-world GitHub bug fixes (Source: technology.org). The figures below are Anthropic-reported unless otherwise noted; the SWE-bench Verified leaderboard maintained at swebench.com lists a separate high-reasoning configuration.

On the ARC-AGI-2 abstract-reasoning test, the 37.6% figure was reported as more than double GPT-5.1's 17.6% and above Gemini 3 Pro's 31.1%; on GPQA Diamond, Opus 4.5's 87.0% trailed Gemini 3 Pro (91.9%) and GPT-5.1 (88.1%) (Source: vellum.ai). On the τ2-bench agentic benchmark, Anthropic described a case in which Opus 4.5 found a policy-compliant path to a customer request that the benchmark had not anticipated and scored as a failure, which Anthropic cited as an example of creative problem solving (Source: anthropic.com).

On METR's HCAST-style time-horizon evaluation, Opus 4.5 measured a task time horizon of approximately 5 hours, and was the first frontier model reported to clear that threshold (Source: blog.aifutures.org). This figure is load-bearing for the time-horizon progression tracked on AGI Timelines.

Availability and pricing

Opus 4.5 was made available on November 24, 2025 through the Claude consumer apps, the Claude API (model ID claude-opus-4-5-20251101), and all three major cloud platforms — Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Azure (Source: x.com). API pricing was set at $5 per million input tokens and $25 per million output tokens, the same structure subsequently applied to later Opus 4.x releases (Source: anthropic.com; platform.claude.com).

At release Anthropic shipped product updates tied to Opus 4.5: a revised Plan Mode and a desktop version of Claude Code capable of running parallel local and remote sessions; automatic summarization of earlier context in long Claude-app conversations; broader access to Claude for Chrome for Max users; and expanded beta access to Claude for Excel for Max, Team, and Enterprise users (Source: anthropic.com). Anthropic also removed Opus-specific usage caps and raised overall usage limits for Max and Team Premium subscribers (Source: anthropic.com).

Safety and evaluations

Opus 4.5's safety case is governed by Anthropic RSP v3.1. Anthropic released the model under ASL-3 (AI Safety Level 3) protections, based on the model's demonstrated capabilities as assessed in the Claude Opus 4.5 system card (Source: anthropic.com; anthropic.com). In the announcement Anthropic described Opus 4.5 as the most robustly aligned model it had released to that point and reported industry-leading robustness against prompt injection: on a combined direct-and-indirect prompt-injection test developed and run by Gray Swan, Anthropic reported an attack success rate of 4.7% for Opus 4.5, against 12.5% for Gemini 3 Pro and 21.9% for GPT-5.1 (Source: anthropic.com; vellum.ai). The model is not released with open weights. A detailed independent review of the system card and its alignment evaluations was published by Zvi Mowshowitz (Source: thezvi.substack.com).

Role in Anthropic's revenue growth

Opus 4.5 is the model most often credited with Anthropic's end-2025 revenue inflection. In what Dario Amodei described at Code With Claude on May 6, 2026 as an "80x revenue surprise," Anthropic's annualized revenue rose from under $1 billion at the end of 2024 to $9 billion at the end of 2025 and to $30 billion by early April 2026 (Dario Amodei; Source: theinformation.com). Opus 4.5 was the model powering Claude Code, Anthropic's agentic-coding product, during the enterprise revenue ramp through Q4 2025 and Q1 2026; Claude Code reached a run-rate of more than $500 million by September 2025 (Source: theinformation.com). The AI Bubble Debate page and the Anthropic Snapshot tables use Opus 4.5 as the reference point for this period. Whether the inflection was driven primarily by Opus 4.5's model capability or by Claude Code's product positioning is not settled in independent analysis.

Reception

Anthropic's announcement collected customer statements describing Opus 4.5 as stronger than Sonnet 4.5 on internal coding and agentic benchmarks while using fewer tokens, with several testers citing token-usage reductions of roughly half on equivalent tasks and improved long-horizon autonomous performance (Source: anthropic.com). Independent coverage emphasized the SWE-bench Verified result as the first published score above 80% on that benchmark (Source: technology.org). Vellum's benchmark write-up characterized Opus 4.5 as state-of-the-art on coding and abstract reasoning while noting that Gemini 3 Pro led on GPQA Diamond, MMMLU, and Vending-Bench 2 and GPT-5.1 led on MMMU, describing the overall profile as among the strongest and most well-rounded in the field rather than dominant on every benchmark (Source: vellum.ai).

Related models

Opus 4.5 was succeeded by Claude Opus 4.6 (February 2026), Claude Opus 4.7 (April 2026), and Claude Opus 4.8 (May 2026).

The system card rules out both AI R&D-4 and CBRN-4, but qualifies each. On autonomy, none of 18 internal survey participants — "some of the most prolific users of the model in Claude Code" — believed it could fully automate an entry-level remote-only research role, while the card notes the model "has just barely reached our pre-defined benchmark rule-out thresholds, rather than greatly exceeded them." On CBRN, an expert uplift trial found Opus 4.5 "meaningfully more helpful to participants than previous models, leading to substantially higher scores and fewer critical errors, but still produced critical errors that yielded non-viable protocols" — from which Anthropic draws a framework-level conclusion, that "a clear rule-out of the next capability threshold may soon be difficult or impossible under the current regime," and states plainly that "the CBRN-4 rule-out is less clear for Claude Opus 4.5 than we would like." A November 24 changelog corrects the card's own ARC-AGI-1 disclosure: training used "a reshuffled train/test split comprising both the public train and test sets," not the public training set alone as originally reported.

Relationships

• superseded-by: Claude Opus 4.6 (Feb 2026), Claude Opus 4.7 (Apr 2026), Claude Opus 4.8 (May 2026).
• deployed-by: Claude Code, Anthropic (consumer Claude).
• depends-on: Anthropic's Responsible Scaling Policy (Version 3.1) — safety case (released under ASL-3).
• related: AI Bubble Debate (revenue-inflection anchor), AGI Timelines (METR time-horizon data point).
• related: Dario Amodei — Code With Claude May 6 80x-revenue-surprise framing.
• related: Track Record — resolution tracking for the time-horizon and deprecation forecasts above.

Sources

• Anthropic, Introducing Claude Opus 4.5 (Nov 24, 2025) — release, capabilities, effort parameter, pricing, availability, product updates, safety. (Source: anthropic.com)
• Anthropic, Claude Opus 4.5 system card — ASL-3 classification, evaluations. (Source: anthropic.com)
• Vellum, Claude Opus 4.5 Benchmarks (Explained) (Dec 3, 2025) — benchmark table compiled from the system card. (Source: vellum.ai)
• Amodei, Code With Claude (May 6, 2026) — 80x revenue surprise framing. (Source: theinformation.com)
• METR HCAST time-horizon framework — Opus 4.5 ~5h baseline. (Source: blog.aifutures.org)

The system card is summarized at System Card: Claude Opus 4.5 (Anthropic, November 2025).