The AI labs Anthropic and OpenAI have announced their latest large language models (LLMs), with a series of cheaper models that the firms say offer similar performance to their best at a fraction of the cost.
On Tuesday, Anthropic revealed Claude Opus 5.5, the latest version of its reasoning model which it said could rival the performance its frontier model Fable 5.1 at a reduced cost.
Claude Opus 5.5 costs 20 per cent less than Opus 5 at $4 per million input tokens and $20 per million output tokens. Anthropic said that the model is around 40 per cent cheaper than its predecessor in practice, as it uses compute more effectively and is less verbose in its responses.
In benchmarks such as FrontierCode v1.1, which tests whether AI agents can generate mergeable production-quality code, Anthropic tests found Claude Opus 5.5 scored up to 54.4 per cent versus Fable 5.1’s 50.3 per cent and Opus 5’s 48 per cent.
Also on Tuesday, OpenAI unveiled new additions to its GPT-6 family of models. GPT-6 Sol and Luna cost half as much to run as their GPT-5.6 predecessors the company said, with GPT-6 Sol available for $2 per million input tokens and $10 per million output tokens. GPT-6 Luna is OpenAI’s new cheapest model, at $0.10 in and $0.01 out.
OpenAI said the models push the frontier of cost efficiency, adding that it is using new approaches to infrastructure, caching and inference that allow it to run models more cheaply.
In FrontierCode v1.1, GPT-6 Sol on scored up to 49.3 per cent, putting it in same leagues as Claude Fable 5.1 on low effort. OpenAI tests on the benchmark AutomationBench, which tests AI agents on their ability to complete end-to-end workflows in real-world fields such as finance, operations and sales the gap narrowed further, with GPT-6 Sol scoring 33.2 per cent.
This puts it ahead of Claude Fable 5.1 for automation tasks, at nearly one ninth the cost, though behind Claude Opus 5.5.
The independent AI benchmarking platform Artificial Analysis found that GPT-6 models hallucinate less than GPT-5.6, with Sol down from 92 per cent hallucination to 60 per cent. However, the platform also noted that this is primarily from refusing to answer more queries.
Artificial Analysis also measured a regression in Luna’s coding ability, with the model scoring two points lower on the benchmarker’s in-house coding index than its predecessor.
Each release is the respective lab’s first since Anthropic chief Dario Amodei and OpenAI chief Sam Altman issued warnings over the risk and pace of advanced AI development. Last week, Dario Amodei called for a slowdown in AI development to let oversight and safety measures catch up with advanced systems, while Altman told Fortune magazine that OpenAI would delay its IPO until AI risks were addressed.
The pair’s demands have been rejected by other key figures in the AI ecosystem, with Meta chief Mark Zuckerberg and Nvidia chief Jensen Huang having argued against any new AI regulations.
