New AI ranking: Claude Opus 5.5 leads, but OpenAI is closing in

What's new

On October 27, 2026, Artificial Analysis released version 5 of its Intelligence Index, a combined score used to compare how capable AI models are. Artificial Analysis describes it as an update with two new tests and three upgraded ones. According to AGTP on X, it is the biggest upgrade since v4 in January.

The big news is that 60 percent of the index is now based on private test sets or held-out answers. It is the first time the majority relies on private data. That makes it harder for AI companies to train their models on the exact questions being tested.

The new tests are FrontierCode 2.0, which has 132 coding tasks from real open source projects and judges whether the answer is worth merging into the codebase, and Terminal-Bench-Science, with 70 science tasks in which the AI has to work on its own in a terminal environment.

What's smart

The index has nine tests across four areas, each weighted 25 percent: scientific work, coding, knowledge work (e.g., office tasks) and general knowledge. That gives a broader picture than a single test.

Three tests have been upgraded: HLE-Diamond is a cleaned-up version with 1,000 questions, AA-Omniscience v1.1 has had questions removed that the models already find too easy, and CritPt v2 is a hard, private set with 40 questions and code execution.

Note that Artificial Analysis runs the measurements itself, not the model makers. And figures from v4.3 and v5 must not be mixed, because they consist of different tests.

Cheaper or better?

In v5 the ranking looks like this: Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55 and Gemini 4 Argon (high) 50. Anthropic and OpenAI models hold the top five spots. No model wins every test: Opus 5.5 leads on three of the nine, while five other models lead the remaining six.

Opus 5.5 also led the old index with 58 points, but back then GPT-6 Astra was 5 points behind (53). Now the gap is just 1 point, so the OpenAI models do relatively better in v5.

On price, the differences are large. The cost is per task in the Intelligence Index: $10.04 for Opus 5.5, $6.24 for GPT-6 Astra, $1.27 for GPT-6.1 Sol, $0.96 for GPT-6.1 Terra and $13.04 for Claude Fable 5.1. GPT-6.1 Sol is thus only 3 points below the top at a fraction of the price.

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens in the API (tokens are small chunks of text). Opus 5 cost $5 and $25. Monthly prices for regular users are not stated in the sources.

Open weights models, whose foundation is publicly available, are falling further behind: the gap to the leader has grown from 3 to 12 points. DeepSeek V4.1 Flash and GLM-5.3 score 32, while Claude Haiku 5.5 scores 36.

What it's good for

Opus 5.5 was released in September 2026 and is available on Claude.ai, via the API and from AWS, Google Cloud and Azure. It can handle up to 1 million input tokens and 128,000 output tokens, and Zero Data Retention (no data storage) is available. The sources do not say whether special pricing or terms apply in Denmark or the EU.

In v5, Opus 5.5 leads on HLE-Diamond with 68 percent versus GPT-6 Astra's 64 percent, so it is strong on hard knowledge questions. GPT-6 Astra leads on Terminal-Bench-Science with 63 percent versus 59 percent.

In practice: if you want the absolute best, Opus 5.5 is still out front. If you need to keep the bill down, GPT-6.1 Sol or Terra is an obvious place to look.