OpenAI cut GPT-5.6 Luna's price by 80%, to $0.20 per million input tokens and $1.20 per million output, and Terra by 20%, to $2.00/$12.00. Sol held its base price but gained a Fast Mode running 2.5x faster for 2x the cost. The mechanism behind the discount is the actual story: OpenAI says its flagship model, Sol, analyzed its own production traffic and rewrote its own serving kernels to fund the cut.
How a model paid for its own price cut
Sol analyzed production traffic patterns, identified load imbalances, tested routing strategies, and rewrote its own production GPU kernels in Triton/Gluon. OpenAI credits that work with a 20% reduction in end-to-end serving costs and a 15% improvement in token generation efficiency from speculative decoding. It's being described as the first publicly documented case of a frontier model self-optimizing its own runtime, with the savings passed directly to customer pricing.
The math has a gap worth naming. A 20% infrastructure saving can't mathematically fund an 80% price cut on its own. Roughly 60 percentage points of that discount came from somewhere else, almost certainly OpenAI eating margin to defend Luna's position against cheap Chinese open-weight models.
Why cost-per-token is the wrong metric
Different models need different volumes of tokens to complete the same task, which makes price-per-token comparisons misleading on their own. Kimi K3, for example, is nominally half the price of Sol, but requires roughly twice the tokens to execute the same task, making the effective cost identical. Under a cost-per-task measurement, GPT-5.6 Luna Max completes tasks for roughly 6 cents, against GLM 5.2 Max at about 26 cents and Claude Sonnet 5 Max at $1.80. That's the number that actually determines whether a cheaper model saves you money.
The competitive pressure behind the timing
This cut lands against two threats at once. Anthropic has expanded usage 80x over three months and dominates parts of the coding market. Chinese open-weight models (GLM 5.2, Kimi K3, DeepSeek V4) are commoditizing inference at a fraction of frontier pricing. Undercutting on price-per-task lets OpenAI hold onto developers before open-source alternatives pull them away. DeepSeek's response was immediate: V4 Flash 0731 launched the same week at $0.28/M output, and one head-to-head test found DeepSeek beating Luna on a Rubik's-cube benchmark at a comparable price point.
Anthropic didn't chase the discount directly. It held Claude Opus 5 at Opus 4.8's rate while delivering near-Fable-5 performance, and launched Claude Sonnet 5 at an introductory $2/$10 per million tokens through August 31, near-parity with Terra only until that window closes.
What this means if you're running high-volume workflows
At $0.20/$1.20 per million tokens, Luna is now cheap enough to make previously cost-prohibitive workflows practical: large-scale batch content generation, massive document processing, elaborate multi-agent delegation chains. If your workflow is latency-critical rather than volume-critical, Sol's Fast Mode gives you 2.5x generation speed for a predictable 2x cost premium. Either way, stop comparing models on the pricing table alone. Build a simple harness that measures total token consumption per finished output for your specific prompts, since that's the number that tells you your actual operating margin.
Price cuts are also showing up in subscription usage limits, not just API rates, so Codex and ChatGPT Work quotas stretch further under the new pricing even if you're not calling the API directly.
FAQ
How much cheaper is GPT-5.6 Luna now?
80% cheaper: $0.20 per million input tokens and $1.20 per million output, down from $1.00/$6.00.
Did the price cut come entirely from efficiency gains?
No. OpenAI's own kernel rewrites accounted for roughly 20% of the cost reduction; the remaining ~60 percentage points of the 80% cut likely came from OpenAI absorbing margin to compete with cheap open-weight alternatives.
Why does "cost per task" matter more than price per token?
Different models require different token volumes to complete the same task. A model with a lower price per token can cost the same or more overall if it needs twice as many tokens to finish the job.
Should I switch my high-volume workflows to Luna?
If your workflow is volume-sensitive and token-heavy, Luna's new pricing likely makes previously cost-prohibitive batch work viable. Build a cost-per-task benchmark on your specific prompts before committing, since raw pricing tables don't capture actual token consumption per output.
This piece is the companion writeup to the Daily AI Pulse episode "OpenAI GPT-5.6 Price Cuts: Luna 80% Off, Terra 20% Off, Sol Fast Mode." Watch it on YouTube, or dig into the source links above for the full picture.
More AI breakdowns for solo builders and small teams 👉 joebuildsai.com