Anthropic released Claude Opus 5.5 on September 22, and OpenAI followed roughly an hour later with GPT-6 Sol and GPT-6 Luna, with both companies delivering substantial API price cuts1,2,3.
Anthropic priced Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for the previous Opus, a 20% reduction in per-token rates. The company said the model works out to roughly 40% cheaper per typical task because it also runs about 30% faster than Opus 57,8. OpenAI priced Sol at $2/$10 and Luna at $0.10/$0.50 per million input/output tokens, half the cost of their GPT-5.6 counterparts, with Luna output tokens 58% cheaper.
Anthropic called Opus 5.5 the first model in a new Claude 5.5 family, saying it performs at the level of Claude Fable 5.1 for most tasks and leads on agentic coding, computer use, and knowledge work. One headline from aibase.com reported the model leading GPT-6 Astra in seven of nine tests. On third-party evals, Opus 5.5 scored 62.3% on FrontierSWE (Proximal), 65.3% on FrontierCode 1.1 (Cognition) Extended, 57.8% on CursorBench (Max), and 93.9% on ParseBench. It also reached the top position on the Vals Index, up two spots and two points from Opus 5.
Anthropic said Opus 5.5 is comparable to its Mythos model in biology and cybersecurity capabilities, so it ships with the same safeguards applied to Fable, including limits on how much the model can be used to discover exploits. The company's system card describes scaling up to 100 parallel agents. Anthropic also said the release is "sufficiently safer than its predecessors that releasing it more likely than not reduces misalignment risks". Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks.
The model also changes how it communicates: Opus 5.5 puts the most important information at the start of its messages and is less likely to use jargon.
OpenAI's caching architecture
Sol and Luna are positioned as more affordable versions of GPT-6 Astra, inheriting much of Astra's advances in coding, computer use, factuality, and alignment, though they still fall short of the flagship. OpenAI rolled both models out to ChatGPT Work, Codex, and the API; Luna is also available to Free and Go users in the desktop app.
Beyond raw token-price cuts, OpenAI emphasized inference-side efficiency. The company said it improved prompt caching for GPT-6 to deliver higher cache-hit rates by default, with cached input-token reads discounted 90%. Developers can now adjust reasoning effort and tool availability without breaking the cache, and a new Prompt Caching Dashboard lets them monitor how much context is reused. OpenAI said GitHub reported that these caching improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests.
ANALYSIS Both releases frame cost reduction as a product of compute efficiency rather than margin compression, with Anthropic citing faster inference and OpenAI citing caching architecture. The practical effect is that developers building agentic workflows, where token volumes compound across tool calls and multi-step reasoning, face materially lower marginal costs from either provider.
Latent Space, which runs its AINews digest on frontier models, said it is migrating to Opus 5.5 immediately, calling the writing-quality difference "night and day" compared with Sol.