OpenAI’s GPT-6 Sol Price Cut Tied to Removed Terra Tier as Anthropic and DeepSeek Reshape AI Pricing

Anthropic’s Claude Opus 5.5 costs 40% less than Opus 5 and delivers Fable-class performance, Leiphone reported. The report said Opus 5.5 matches the previous top model on most daily tasks, with gaps only in very deep logical reasoning and very large code engineering. It also cut the cost of cache reads by 60%, reducing the expense of long tasks that involve document analysis, multi-turn tool calls, self-correction and deeper reasoning.

OpenAI introduced GPT-6 Sol and Luna. Leiphone reported that GPT-6 Sol was presented as matching GPT-5.6 Sol while halving the price, to $2 per million tokens for input and $10 per million tokens for output. But the report said the comparison depends on a tier shift. OpenAI’s earlier lineup was flagship Sol, mid-tier Terra and lightweight Luna. The new lineup is Astra, Sol and Luna. Terra has disappeared, and online observers joked that Terra had become Sol, moving the whole ladder up one rung.

Compared with GPT-5.6 Terra, which cost $2 per million tokens for input and $12 for output, GPT-6 Sol at $2 input and $10 output is almost unchanged, according to the report. Leiphone described Anthropic’s reduction as the most genuine among the three, DeepSeek as still holding an absolute cost advantage, and OpenAI’s move as more of a commercial pricing narrative supported by engineering optimization.

In the lightweight segment, the report said DeepSeek-V4.1-Flash appears more expensive than GPT-6 Luna at peak hours. DeepSeek-V4.1-Flash’s input fee is 1.5 to 3 times Luna’s, and its output fee is 1.2 to 2.4 times Luna’s. The two are not directly comparable in size or capability, Leiphone said. GPT-6 Luna is a true small model, possibly with only tens of billions of parameters, used for simple short-text filtering and classification. DeepSeek-V4.1-Flash is a 552B-parameter mixture-of-experts model that activates only about 16B expert parameters through its asymmetric causal encoding-decoding architecture. DeepSeek’s low prices also come with conditions, including off-peak discounts and a heavy reliance on high prefix repetition. Luna’s lower price has no such threshold, which the report said can matter to smaller developers.

OpenAI’s pricing rules are central to its strategy, Leiphone reported. Cache-hit input is priced at 90% off, or $0.20 per million tokens. The report argued that this encourages developers to keep large enterprise prompts and data attached to OpenAI’s cache, raising switching costs. For long context, GPT-6 raises prices once input exceeds 272K tokens: the entire request’s input fee doubles and output fee rises by 1.5 times. Leiphone said this uses price to limit high-consumption long-context use and reflects weaker long-text efficiency than rivals that rely on stronger compression.

DeepSeek and OpenAI have taken different paths to lowering costs, according to the report. DeepSeek began with mixture-of-experts designs and pushes activation efficiency further in V4.1-Flash, where prefill and decode use different modules and only about 8B to 16B parameters are active. The report cited the YOCO approach of caching once, so the prefill stage generates a global cache and the decode stage reuses it. DeepSeek also compressed KV cache through multi-head latent attention and later compressed sparse attention. Leiphone said V4 reduced KV cache to about 3,560 bytes per token, and V4.1’s second-generation compressed sparse attention cut it to about 890 bytes per token, roughly 2% of a standard Transformer architecture. DeepSeek also tries to use idle GPU time and space, as well as host memory and data buses, in large cluster scheduling.

OpenAI took a more pragmatic route, the report said, avoiding fine-grained MoE changes that could disrupt its scaled training graph. It relied on parameter distillation and sparsification, splitting some redundant dense layers into lighter small MoE branches to reduce floating-point operations per token. For long context, it uses tiered pricing and routes requests beyond 272K tokens to B200 nodes with more memory. Its main cost lever is prompt caching. Leiphone reported that OpenAI extended cache lifetime to 30 minutes, supports dynamic mode changes without invalidating old caches, offers explicit cache breakpoints and diagnostic tools, and prewarms caches for system prompts, tool definitions and reference documents. In a Copilot interactive coding scenario, the number of tokens requiring reprocessing was cut by more than half. In enterprise agent deployments such as Manus, cache hit rates were pushed from 85% to above 90%, according to the report.

For the past two years, the cost-performance narrative in large models belonged to DeepSeek, and its founder Liang Wenfeng was nicknamed Liang Sheng by online commenters. After OpenAI’s cuts, some joked that a new Ao Sheng had risen. Leiphone wrote that OpenAI’s value story is half label change and half cache discount, both of which depend on users not examining the price table closely. It also asked what players without ten-thousand-GPU clusters or self-developed architectures will bring to the next round.