OpenAI Cut Luna's Price 80%. Sol Wrote the Kernels That Paid for It.

TL;DR

On July 30, OpenAI cut API prices on two of the three GPT-5.6 tiers. Luna drops 80%, from $1.00 / $6.00 to $0.20 in and $1.20 out per million tokens. Terra drops 20%, from $2.50 / $15.00 to $2.00 and $12.00. Sol, the flagship, does not move: still $5 / $30. The reason OpenAI gives in its announcement is the part worth your attention. It says Sol autonomously rewrote production kernels and ran hundreds of token-generation experiments, cutting the end-to-end cost of serving the family by 20% and improving token-generation efficiency by more than 15%. The model optimized the thing that serves the model, and the savings landed on your invoice.

The new numbers

This is a straight repricing of existing models, not a new release. Same weights, same endpoints, same names. If you have Luna wired into a classification job or a subagent fleet, you woke up to a bill that is one fifth of yesterday's for identical work.

Input prices moved the same percentages: Luna $1.00 to $0.20, Terra $2.50 to $2.00. OpenAI's pricing page now lists all three tiers at the new rates. The tier structure we broke down when the family went generally available is intact; only the middle and bottom of it got cheaper.

Why: OpenAI says the model did the optimizing

Vendors cut prices all the time and usually mumble something about "efficiency gains." OpenAI got specific, and the specifics are strange. Per iClarified's readout of the announcement, GPT-5.6 Sol "autonomously rewrote production kernels" and "designed and ran hundreds of token-generation experiments." Those two workstreams reportedly reduced end-to-end serving cost by 20% and improved token-generation efficiency by more than 15%. The Decoder ties the token-generation gain specifically to speculative decoding.

What speculative decoding actually is

A transformer normally emits one token per forward pass, and each pass drags the whole weight set through memory. Speculative decoding pairs the big model with a small fast draft model that guesses several tokens ahead, then has the big model verify the whole guessed run in a single pass, keeping the prefix that matches and discarding the rest. It is a court stenographer who types what they expect you to say and only backspaces when you surprise them: when the guess is good you got several tokens for the price of one, and when it is bad you paid a rounding error to find out.

Kernel rewriting is the less glamorous half and probably the bigger number. Serving a model means thousands of hand-tuned GPU routines for attention, matrix multiplies, quantization, and memory movement. Squeezing them is exactly the kind of narrow, verifiable, brutally repetitive optimization work an agent can grind at in a loop with a benchmark as the scoring function. If OpenAI's account holds up, this is the first time a frontier lab has publicly credited its own model with a line-item change to what customers pay.

Cheap does not mean dumb, but check the receipts

The obvious question is what you give up at $1.20 per million output tokens. Artificial Analysis, which benchmarks independently of the labs, puts Luna at max reasoning at 51 on its Intelligence Index, against 55 for Terra, 59 for Sol, and 60 for Anthropic's Claude Fable 5. That is a real gap at the top, and a narrower one than the 25x price spread between Luna and Sol would suggest.

The more useful metric is what a full index run costs. Artificial Analysis measured that at pre-cut prices.

OpenAI's own framing is bolder. It says Luna matches frontier-class models from a year ago at roughly six cents on the dollar per task and close to nine times the speed, and that on Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, Luna beats Claude Fable 5 at an estimated cost per task nearly 99% lower. Read that as a vendor claim on a vendor-favored board, because it is one. Artificial Analysis independently notes that Luna at max reasoning matches or exceeds Gemini 3.5 Flash and GLM-5.2 at max reasoning while costing less, which is the more sober version of the same story.

What this changes for your routing

If you run anything at volume, the arithmetic just moved under you:

• Subagent fleets get cheap. The pattern where a Sol-class planner farms out dozens of narrow tasks was previously bounded by what the workers cost. At $1.20 out, the workers are close to free relative to the planner.
• Terra is now the boring default. At $2 / $12 it costs less than Luna's launch price did three weeks ago, with a four-point higher index score. If you downgraded to Luna purely on budget, re-check that decision.
• Batch and classification work should be re-benchmarked, not re-priced. The tempting move is to cut your spend by 80% on the same volume. The better move is to spend the same and buy five times the volume, or five times the reasoning effort, whichever your eval says helps.
• Sol did not get cheaper. The gap between the flagship and everything below it widened by design. Routing, not loyalty, is the strategy.

The caveats

Three things stay straight-faced here.

The self-optimization claim is unaudited. "Sol rewrote our kernels" is OpenAI describing OpenAI's internal engineering, with no third-party verification of how much of the work was autonomous versus agent-assisted-with-a-human-reviewing-every-diff. It is a plausible claim about exactly the kind of task agents are good at, and it is still a claim.

The competitive read matters as much as the technical one. CNBC frames the cuts as a response to customers growing cost-sensitive, and The Decoder puts it more bluntly: OpenAI went to China pricing. Cheap Chinese open-weight models and Microsoft pushing its own MAI line as a budget alternative are both real pressure. A price cut can be an engineering result and a defensive move at the same time.

Cost per task is not cost per outcome. A model that is 5x cheaper and needs 3x the retries is 1.7x cheaper, and only your eval knows the multiplier. Run it on your own workload before you rewire the router.

Key Takeaways

• Effective July 30, GPT-5.6 Luna costs $0.20 / $1.20 per million input/output tokens, down 80% from $1.00 / $6.00. Terra fell 20% to $2.00 / $12.00. Sol is unchanged at $5 / $30.
• OpenAI attributes the cut to efficiency gains it credits to GPT-5.6 Sol, which it says autonomously rewrote production kernels and ran hundreds of token-generation experiments.
• The claimed result: 20% lower end-to-end serving cost and more than 15% better token-generation efficiency, with The Decoder attributing the latter to speculative decoding.
• Artificial Analysis independently scores Luna at max reasoning at 51 on its Intelligence Index versus 55 for Terra, 59 for Sol, and 60 for Claude Fable 5, at $0.21 per index run under pre-cut pricing.
• OpenAI's sharper claims, including beating Fable 5 on Agents' Last Exam at nearly 99% lower cost per task, are self-reported on a board it selected.
• The practical move is to re-benchmark your routing rather than just pocket the discount, and to remember that cheaper per token is not cheaper per correct answer.

Sources: OpenAI, "Advancing the price-performance frontier with GPT-5.6", OpenAI API pricing, CNBC, The Decoder, iClarified, Notebookcheck, Artificial Analysis: GPT-5.6 Luna, Artificial Analysis: GPT-5.6 has landed