Claude Opus 5

Claude Opus 5, released on 24 July 2026, is the reasoning-focused member of Anthropic's Claude 5 family. It takes a context window of one million tokens and can return up to 128,000 tokens in a single response, or 300,000 through the Batch API under a beta header. Input is text and images; output is text, and the model is multilingual. Its knowledge cutoff is May 2026. The defining change in this generation is adaptive thinking, which replaces the explicit extended-thinking toggle of earlier Claude models: the model decides how much internal reasoning a request warrants, while an effort parameter with five levels lets the caller bias that decision. It supports tool use, agentic workflows and computer use, and is benchmarked on OSWorld 2.0 for the latter. Anthropic reports state-of-the-art results on Frontier-Bench v0.1 and a score three times the next-best model on ARC-AGI 3, but the same announcement records two rows it does not win: Opus 5 stays behind Claude Mythos 5 on cybersecurity tasks, and on CursorBench 3.2 it lands just under Claude Fable 5, within 0.5% of that model's peak score. The Frontier-Bench figure is a vendor-internal run on the mini-SWE-agent harness, mean reward over five attempts per task, with Opus 4.8 serving as fallback whenever a safety classifier refused a request. Artificial Analysis measured it at 61 points on the Intelligence Index at maximum effort against a median of 32 — but that was version 4.1 of the index, the one in force at launch; under the current version 4.3 the same configuration scores 51 against a median of 24. Opus 5 is not the top of Anthropic's price list — Claude Fable 5 sits above it — but at five dollars per million input tokens and twenty-five per million output it is the company's main choice for hard reasoning and long-horizon agentic work. It runs on the Claude API, Amazon Bedrock, Google Vertex AI and Microsoft Foundry, and powers Claude.ai for paid tiers, Claude Code and Claude Cowork. Anthropic deployed it under ASL-3 of its Responsible Scaling Policy. Earlier Opus releases — 4.8, 4.7, 4.6, 4.5 — remain documented by the vendor; Opus 4.1 is deprecated and retires in August 2026.

▸News

The chart that made Opus 5 state of the art has a footnote: on refusals, the answer came from Opus 4.8

9/13/2026 · Anthropic

Five Opus releases, one price: Anthropic's flagship has not changed its rate since November — and the dearest Claude on sale is not a flagship

8/23/2026 · Anthropic — cennik API i noty wydawnicze kolejnych modeli Claude; zestawienie własne wujec.ai

Claude's watermark covers models launched from 2 August — Anthropic has not launched one since

8/17/2026 · Anthropic

▸Videos

No videos yet.

▸Reviews