OpenAI Cut GPT-5.6 Prices 80% by Having the Model Optimize Itself
OpenAI slashed GPT-5.6 Luna's price 80% after using GPT-5.6 Soul to find its own efficiency gains, a real step toward recursive self-improvement.
What did OpenAI actually announce?
OpenAI cut the price of GPT-5.6 Luna, the smallest model in its GPT-5.6 family, by 80%. It now costs 20 cents per million input tokens and $1.20 per million output tokens, undercutting most frontier open-source models coming out of China on a per-token basis. GPT-5.6 Terra, the mid-tier model, got a 20% price cut, landing at $2 per million input tokens and $12 per million output tokens. The flagship model, GPT-5.6 Soul, kept its list price but gained a faster inference mode: 2.5x the speed for 2x the cost, up from the previous 1.5x speed for the same 2x cost. Same price increase, more speed, which functions as a quiet price cut for anyone who values throughput.
The more interesting part of the announcement is how OpenAI found the room to cut prices at all. According to the company’s own writeup, it used GPT-5.6 Soul itself, running inside its Codex coding agent, to analyze production traffic, rewrite GPU kernels, improve routing heuristics, and optimize the model’s speculative decoding pipeline. The result: 20% lower serving costs from kernel improvements and 15% better token generation efficiency from improved decoding. A frontier model was turned loose on its own infrastructure and told to make itself cheaper to run.
TL;DR
- OpenAI cut GPT-5.6 Luna’s price by 80%, bringing it to 20 cents per million input tokens and $1.20 per million output tokens, cheaper per token than most open-source Chinese frontier models.
- The savings came from self-optimization, with GPT-5.6 Soul running inside Codex to find production inefficiencies, rewrite GPU kernels, and improve speculative decoding.
- Cost per completed task matters more than price per token, since models like Kimi K3 are half the listed price but use roughly twice the tokens to finish the same job, netting out to a similar real cost.
- On the artificial analysis intelligence index, GPT-5.6 Luna Max lands near the top left of the chart, matching or slightly beating models like GLM 5.2 Max and Claude Opus 5 on intelligence while costing a fraction as much per task.
- Soul’s price stayed flat despite likely efficiency gains, which suggests OpenAI is using those gains to widen margins on its highest-revenue model rather than pass savings along.
- This looks like the emerging playbook for frontier labs: train an expensive, maximally capable model, then use it to compress cheaper, nearly-as-capable versions for public release, funding the next big training run.
- The self-optimization loop echoes Andrej Karpathy’s “nanochat” style auto-research experiments, where a model designs and runs its own experiments to improve a smaller model’s architecture, but scaled to production infrastructure at a major lab.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How did GPT-5.6 make itself cheaper?
The mechanism is straightforward in concept even if the engineering underneath is not. OpenAI pointed GPT-5.6 Soul, paired with its Codex coding agent, at its own production systems. The model analyzed real traffic patterns and found load-balancing inefficiencies nobody had flagged before. It tested new routing strategies and tuned the heuristics that decide how requests get distributed across GPUs. It also worked on the forward pass, the core computation that turns an input into a next-token prediction, looking for steps that could be precomputed, skipped, or run in parallel instead of sequentially.
Two of the more concrete outcomes: a 20% reduction in serving cost from GPU kernel rewrites, and a 15% improvement in token generation efficiency through better speculative decoding (a technique where a smaller draft model proposes tokens that the larger model verifies, speeding up generation without sacrificing quality). OpenAI also said Soul improved its own draft model by running hundreds of experiments on architecture choices like size and structure.
None of this requires a philosophical leap to “AI improving AI” in the sci-fi sense. It’s a capable coding and analysis model applied to a well-defined optimization problem, running continuously against live production data instead of being called once and set aside. The scale is what stands out: millions of daily prompts generate a constant stream of data that a frontier model can mine for inefficiencies, around the clock, without needing sleep or a meeting on the calendar.
Why does cost per task matter more than cost per token?
Price-per-token comparisons are easy to headline but often misleading. A cheaper model that needs twice as many tokens to solve the same problem isn’t actually cheaper. Kimi K3, the open-source Chinese model referenced in OpenAI’s comparisons, is roughly half the price of GPT-5.6 Soul per token, but tends to use about twice as many tokens to complete equivalent tasks. Net cost: roughly a wash.
This is why cost-per-completed-task benchmarks, like the ones tracked by artificial analysis, are more useful than raw pricing tables. On that metric, GPT-5.6 Luna Max comes in around 6 cents per task, compared to about 26 cents for GLM 5.2 Max, roughly 50 cents for Grok 4.5, and around 40 cents for Claude Opus 5 at low settings, all while landing in a similar intelligence tier. Claude Sonnet 5 Max, by comparison, runs about $1.80 per task for comparable capability. The efficiency gap isn’t marginal. It’s the difference between a model you can run constantly in production workflows and one you reserve for occasional high-value queries.
Is this actually recursive self-improvement?
It depends on how strict a definition you use. Full recursive self-improvement, in the sense AI safety researchers worry about, implies a system that improves its own core capabilities in a runaway loop with no human oversight. What OpenAI describes is narrower: a model optimizing the infrastructure and efficiency of serving itself and related models, with humans reviewing and shipping the changes. It’s optimization of deployment, not a redesign of the model’s underlying intelligence.
Still, it’s a real, working example of a loop where a frontier model identifies inefficiencies, proposes fixes, tests them, and iterates, largely on its own. That pattern resembles Andrej Karpathy’s open-source auto-research experiments, where a model was tasked with improving the training efficiency of a small language model from scratch, ran its own experiments, and found gains a human researcher had missed. OpenAI’s version runs the same kind of loop against a production-scale frontier model with a full company’s compute behind it. The difference in scale between a hobbyist experiment and a top lab’s infrastructure is significant, and that gap is exactly what raises questions about competitive moats.
What does this mean for competition between AI labs?
If OpenAI and Anthropic are both far enough ahead to have their best models improving the systems that serve their next-best models, the gap between leading labs and everyone else could widen rather than close. A frontier model that optimizes its own kernels, routing, and decoding is also, indirectly, funding and accelerating whatever comes next: lower serving costs mean better margins, and better margins mean more capital for the next training run.
There’s a plausible strategy forming here: train the largest, most expensive model you can afford. Use it to compress smaller, nearly-as-capable, far cheaper models for public and developer use. Generate revenue from those workhorse models. Reinvest that revenue into the next flagship. Meanwhile, the flagship itself may not even ship to the public quickly. Anthropic reportedly had its top model internally for months before public release, which suggests labs may increasingly hold their true frontier capability back while monetizing the compressed version.
Notably, GPT-5.6 Soul’s own price didn’t drop, even though it almost certainly saw similar efficiency gains internally. The likely explanation is that Soul is where the bulk of OpenAI’s revenue comes from, not the smaller workhorse models, so any efficiency gains there get absorbed as margin rather than passed to customers. Competitive pressure, particularly from open-source releases like Kimi K3 out of China, may be part of what pushed OpenAI to move quickly on Luna and Terra’s pricing instead.
Frequently Asked Questions
Why did GPT-5.6 Luna get an 80% price cut but Soul got none?
Luna and Terra are the workhorse models most developers use at scale, so cutting their price expands the number of tasks OpenAI can profitably serve. Soul is the primary revenue driver among power users and enterprise customers, so any efficiency gains there likely went toward improving margins instead of lowering the sticker price.
Does using GPT-5.6 to optimize itself count as recursive self-improvement?
It’s a real, working instance of a model improving the systems that serve it: kernels, routing, decoding. It’s not the same as a model rewriting its own core training or capabilities autonomously, but it’s a concrete step in that direction and worth watching closely.
How is “cost per task” different from “price per token”?
Price per token measures the sticker cost of input and output tokens. Cost per task accounts for how many tokens a model actually needs to solve a given problem. A cheap-per-token model that uses twice the tokens can end up costing the same, or more, than a pricier model that’s more token-efficient.
Is GPT-5.6 Luna actually competitive with open-source models like Kimi K3 or GLM 5.2?
On intelligence benchmarks tracked by artificial analysis, GPT-5.6 Luna Max performs comparably to or slightly ahead of models like GLM 5.2 Max, while costing a fraction as much per completed task, based on OpenAI’s cited comparisons.
What does this mean for developers choosing between OpenAI and Anthropic?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The price gap and efficiency gains shift the calculation on cost-sensitive workloads. For high-volume, latency-tolerant tasks, GPT-5.6 Luna’s new pricing makes it markedly cheaper to run at scale, which may pull developers who previously defaulted to more expensive frontier models for routine work.
