OpenAI cuts GPT-5.6 Luna pricing by 80% and frames it as the start of something bigger
OpenAI drops Luna API prices to $0.20/$1.20 per million tokens and trims Terra by 20%, calling it the beginning of an 'abundance flywheel' strategy.
Twenty-one days after launching the GPT-5.6 family, OpenAI has cut the API price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%. The changes went live on July 30, 2026, and they are the most significant pricing moves since the GPT-5.6 models shipped.
The new rates: Luna is now $0.20 per million input tokens and $1.20 per million output tokens, down from $1 and $6. Terra drops from $2.50/$15 to $2/$12 per million tokens. Cached input reads on Luna fall further still, to just $0.02 per million tokens.
What actually changed, and what didn’t
Sol, the top-tier model in the GPT-5.6 family, keeps its Standard price of $5/$30 per million tokens. OpenAI did add a new Sol Fast mode at $10/$60, which delivers up to 2.5 times the throughput for latency-sensitive workloads. That replaces the previous Priority Processing tier.
Subscription prices for ChatGPT Work and Codex are unchanged. But because Luna and Terra now cost less to serve internally, the credits inside those subscriptions effectively go further. It is a quiet usage-cap increase that never showed up in a pricing email.
The “abundance flywheel” framing
OpenAI published a company-level essay alongside the announcement, titled Building Abundant Intelligence. The core argument is that lower prices are not a concession but a compounding strategy:
“When the cost of useful intelligence falls, more work becomes worth doing. When models become more capable, that work creates more value. As adoption grows, we gain the revenue, real-world feedback, and visibility into demand to keep investing in the next generation of research and infrastructure.”
That framing matters because it signals intent. OpenAI is not treating this as a promotional discount. The company is publicly committing to a trajectory where prices keep falling as capability and adoption grow together.
Why now
The timing is not accidental. A few things are happening at once.
Chinese models have been pulling significant enterprise volume. Since February 2026, the share of tokens US companies route to Chinese models has stayed above 30% every week, hitting 46% in some weeks, against a 12-month average of 11% before that. On some workloads, those models carry token rates up to nine times lower than comparable US frontier systems. That is a real competitive pressure, not a theoretical one.
Anthropic released Claude Opus 5 at the same price as Opus 4.8 just days before this announcement. Google has been pushing Gemini 3.6 Flash and Flash-Lite as low-cost, fast alternatives. The market is consolidating around cost-per-task as a primary buying criterion, and OpenAI is responding directly.
OpenAI’s own benchmarking shows Luna matching Claude Opus 5 (Low)‘s intelligence score at roughly one-sixth the cost per task, putting it ahead of Gemini 3.6 Flash, GLM-5.2 Max, and Claude Sonnet 5 on that metric. That claim is worth pressure-testing with your own workloads, but it is the competitive position OpenAI is staking out.
What helped make these cuts possible
One detail worth noting: GPT-5.6 Sol was used to optimize the production software serving OpenAI’s own models. That work reduced end-to-end serving costs by 20% and improved token-generation efficiency by more than 15% through better speculative decoding. Sol effectively rewrote its own GPU kernels and speculative-decoding model inside Codex.
That is the first confirmed case of a frontier model autonomously optimizing its own inference stack in production. It also explains, at least partially, where the margin headroom for Luna’s price cut came from.
What this means for you
If you are building on the API, Luna is now genuinely worth re-evaluating for high-volume workloads. At $0.20/$1.20, it sits close to the lowest-cost commercial models available, while retaining tool use and multi-step workflow support. For applications where you were previously capping usage or routing to cheaper but less capable models, the calculus has shifted. Terra’s 20% reduction is less dramatic but still meaningful at scale.
If you are running ChatGPT Work or Codex, nothing changes in your invoice, but your effective usage ceiling has gone up. Luna and Terra consume fewer credits at the new internal rates, which matters if you have been hitting limits on agentic or long-running tasks.
If you are evaluating AI infrastructure costs more broadly, the historical comparison is useful context. GPT-5.4, the flagship from roughly four months ago, cost $2.50/$15 per million tokens. Luna, which OpenAI positions as matching that level of intelligence, now costs $0.20/$1.20. That is roughly a thirteenfold reduction in token price for comparable capability in under half a year.
The trajectory OpenAI is describing, and now demonstrating, is one where the cost of running capable AI keeps falling. For teams that have been waiting for the economics to make sense before scaling up, that moment is arriving faster than most roadmaps anticipated.