OpenAI halves API prices with GPT-6 Sol and Luna, hours after Anthropic repriced Opus 5.5

OpenAI and Anthropic cut API prices within hours of each other on September 22, and the gap between them tells you where the coding-model market is heading.

OpenAI launched GPT-6 Sol and GPT-6 Luna, priced at roughly half of the GPT-5.6 models they replace. Sol, the reasoning tier, now costs $2 per million input tokens and $10 per million output, down from $4 and $20 on GPT-5.6 Sol. Luna, the fast high-volume tier, falls to $0.10 input and $0.50 output per million, down from $0.20 and $1.20. That is a 50 percent cut on Sol and about 58 percent on Luna's output price. OpenAI says the savings come from inference and caching efficiencies that it is passing through to customers. GetAIBook Metir

The flagship holds its price. GPT-6 Astra, OpenAI's top model, stays at $10 and $50 per million. OpenAI is stratifying the lineup by price rather than generation: Astra keeps the premium, Sol fights for volume, Luna goes after workloads that were experiments last month and line items this month. ByteVyte

Ninety minutes apart

Anthropic went first the same day, releasing Claude Opus 5.5 at $4 per million input and $20 per million output, 20 percent below Opus 5's $5 and $25 list price, and cutting cache reads 60 percent to $0.20 per million. Anthropic says typical workloads should cost around 40 percent less than on Opus 5. OpenAI's Sol and Luna followed within about ninety minutes, according to Value Add Pulse's tally. Value Add Pulse EveryDev

xAI had set the tone two days earlier, launching Grok 4.7 at $2 and $6 per million and declaring a price war, even as it raised rates for prompts above 200,000 tokens.

What the independent numbers say

Artificial Analysis scores GPT-6 Sol Max at 48 on its Intelligence Index with an estimated $1.06 cost per task, versus $1.99 for GPT-5.6 Sol. Luna holds at 37, with benchmark cost per task falling from $0.18 to $0.07. Independent testing found improvements in some agentic and coding evaluations and regressions in others. One caveat on OpenAI's own benchmark charts: they compare against Claude Opus 5, not the Opus 5.5 that launched the same day at a lower price, which narrows several of the headline cost-advantage claims. ExplainX

Both models shipped the same day in ChatGPT Work, Codex, and the API, according to TechCrunch and Fortune.

Why developers should care

For teams running coding agents in production, the price of the model line is becoming as important as the benchmark line. A million-token agent session, long context, heavy tool output, multiple revisions, now costs fifty cents at Luna rates versus twenty dollars at Fable 5.1-class pricing. The week's cache-read cut on Opus 5.5 matters for the same reason: cache reads are what an agent pays to reuse context it has already processed, so a session working over the same repository collects the discount on every turn.

Nobody has won the intelligence race outright. The price race now has a clear direction: down, and fast.