TL;DR

Anthropic has released Claude Opus 5.5, which it says matches Claude Fable 5.1 on most work while cutting running costs by 40% against Opus 5. On the same day OpenAI launched GPT-6 Sol and GPT-6 Luna, cheaper siblings of GPT-6 Astra, at half the promotional API price of the GPT-5.6 versions. For businesses paying per token, frontier-grade output just got noticeably cheaper.

What Anthropic is offering

Opus 5.5 opens a new 5.5 family, with Sonnet and Haiku versions promised within weeks. Per million tokens, input is $4 and output $20, both down a fifth on Opus 5. Reading from the cache, which Anthropic says makes up most of what coding and agent jobs cost, drops 60% to $0.20. Cheaper tokens combine with the model needing fewer of them for each job to produce the 40% figure, and Anthropic says output also arrives over 30% faster.

It is the first model Anthropic has shipped since arguing that frontier development should be paced so safety practice stays ahead of capability. METR and Frontier Design tested it before release, and it launches with cyber and biology safeguards of a similar class to Fable 5.1’s. Anthropic concedes one weakness: the model often seems to suspect it is being tested, which makes its real-world behaviour harder to judge.

OpenAI’s answer

GPT-6 Sol is priced at $2 for a million tokens in and $10 out, with Luna at $0.10 and $0.50. OpenAI’s headline claim is that Sol, at its “xhigh” effort setting, beats Claude Opus 5 at maximum effort on the AutomationBench business-workflow test for 9% of Opus 5’s spend on each task. That comparison is against Anthropic’s previous model, not the one released the same day. The new models are rolling out to paid ChatGPT Work and Codex users, with Luna reaching free users through the desktop app.

Looking forward

Each lab is benchmarking against the other’s pricing as much as its capability, and both say savings on caching are central, which favours long-running agents over one-off chat. For UK buyers the practical step is to re-run cost estimates on current workloads before renewing contracts.

The safety process deserves a closer look too. Last week Anthropic released a model without UK AI Security Institute testing, and the Opus 5.5 announcement names METR and Frontier Design as outside testers without mentioning AISI. It has also signed up Faculty to evaluate its models from inside. Whether AISI gets access to Opus 5.5 is a question ministers can now put directly.