When GPT-5.6 shipped at half the price of Claude Fable 5 while landing two points behind it on the intelligence index, the obvious question was how OpenAI hit that price. Part of the answer is a discount. The more interesting part is that GPT-5.6 spends fewer tokens to reach the same answer, because it does its internal reasoning in a dense, model-native notation rather than in verbose English. This is the leading edge of a research direction that has been building in the open for two years, and it is worth understanding properly — because it explains where model pricing is heading, and it comes with a real trade-off nobody should ignore.
To see why "thinking in a private language" is a cost lever and not a gimmick, you have to start with how models learned to reason out loud in the first place.
How reasoning became visible
In 2022, Google researchers published a deceptively simple finding: if you ask a large model to "think step by step" before answering, it gets much better at hard problems. This was chain-of-thought prompting, and it changed how everyone used these models. Instead of jumping straight to an answer, the model would write out its reasoning — "first I need to find X, then Y, therefore Z" — and that visible scratch-work dramatically improved accuracy on maths, logic, and multi-step tasks.
Chain-of-thought worked because it gave the model room to compute. Each token a model generates is a small amount of computation, so a model that writes 500 tokens of reasoning before answering has literally done more thinking than one that answers immediately. This insight became the foundation of the reasoning-model era: OpenAI's o1 in late 2024, DeepSeek-R1 in early 2025, and every frontier reasoning model since scales up exactly this — spend more computation at inference time, in the form of more reasoning tokens, to solve harder problems. The industry calls it test-time compute, and it has been one of the biggest drivers of capability gains since raw model scaling started to plateau.
But there is a catch, and it is a big one.
The catch: English is an expensive medium for thought
Chain-of-thought reasoning is written in natural language, and natural language is enormously redundant. When a model writes "first, let me carefully consider what the problem is actually asking me to find here," most of those tokens carry almost no information — they are grammatical scaffolding, politeness, and human-readability that the model does not need in order to compute. It is writing an essay to itself when a few symbols would do.
And you pay for every one of those tokens. As we cover in the hidden reasoning-token tax, reasoning models bill the hidden thinking at the output rate, so a verbose chain of thought is not just slow — it is the single largest cost driver on a reasoning workload. A model that needs 4,000 tokens of flowery English reasoning to solve a problem is, in a real sense, wasting most of your money on grammar.
So the obvious question researchers started asking around 2024 was: does the model actually need to think in English at all?
Thinking without words
The answer, increasingly, is no — and this is where the "private language" comes from.
The clearest public marker is Meta's "Chain of Continuous Thought" work from late 2024, nicknamed Coconut. The idea is to let the model reason in its own continuous internal representation — the high-dimensional vectors it computes with — instead of forcing each reasoning step to become a discrete word. Reasoning never leaves the model's native space; it does not have to be "rendered" into English and re-read. Strikingly, the researchers found that on some problems this reasons better, not just cheaper, because the model can hold several possible lines of thought at once rather than committing to one word at a time. Natural language, it turns out, was a bottleneck as well as a cost.
A parallel line of work goes by names like compressed chain-of-thought and latent reasoning, and the thread connecting all of it is the same: the model's reasoning does not need to be human-readable to be correct, and stripping out the human-readability makes it faster and cheaper. Whether the internal representation is a compressed token vocabulary the model invented for itself or a fully continuous latent trajectory, the effect is the same — the same logical work, in a fraction of the footprint.
GPT-5.6 is the first frontier product to make this a headline cost lever rather than a research demo. You send an ordinary prompt and get an ordinary answer; behind the scenes, the model reasons in its compressed notation and generates far fewer hidden tokens to get there. The advance is not a bigger brain. It is a more efficient inner monologue.
What this actually buys
Three things, in order of how much they matter to a buyer.
Lower cost per answer. This is the direct one. Fewer reasoning tokens per problem means a lower effective price for the same quality, independent of the sticker rate. It is why GPT-5.6 can sit near the top of the board at a mid-tier price.
Lower latency. Tokens are generated one at a time, so fewer reasoning tokens also means faster answers. A model that thinks in shorthand returns its response sooner, which matters for anything interactive.
Headroom for harder problems. Because compressed reasoning is cheaper per step, the model can afford more steps within the same budget. Efficiency does not just save money on easy problems; it lets the same price buy deeper thinking on hard ones. This is the compounding effect — every gain in reasoning efficiency raises the ceiling of what is affordable.
Does the compressed reasoning stay correct?
The obvious worry about all of this is that compressing reasoning must lose something — that a model thinking in shorthand is cutting corners and will be wrong more often. It is the right question to ask, and the research answer is more encouraging than intuition suggests.
The Coconut results and related latent-reasoning work found that reasoning in the model's native representation was not just cheaper but, on several problem types, more accurate. The reason is subtle: when a model reasons in discrete language, it has to commit to one specific word at each step, which forces it to collapse many possible lines of thought into a single path early. Reasoning in a continuous internal space lets it keep several possibilities alive at once and commit later, which is closer to how good problem-solving actually works. Verbose natural-language chain-of-thought was never the optimal way for a model to think; it was just the way we could read.
That said, "more accurate on some benchmarks" is not "always better," and the honest position is that this is still maturing. Compressed and latent reasoning are strongest on well-structured problems — maths, logic, planning — where the reasoning has a clear shape. On open-ended tasks the picture is less settled, and vendors are conservative about where they deploy the most aggressive compression. The practical reality in a shipping model like GPT-5.6 is a blend: compressed reasoning where it is proven, more conventional reasoning where it is not, tuned so the quality you observe is unchanged while the token count drops. You are not trading accuracy for cost; you are removing tokens that were not buying accuracy in the first place.
What it means for how you build
For anyone building on these models, the shift from visible to compressed reasoning has a few concrete consequences worth planning around.
First, stop treating the pricing page as the cost. As reasoning moves into a compressed form, the number of tokens a model spends per answer becomes the dominant cost variable, and it varies more between models than the sticker rate does. Log your reasoning-token usage and compare models on cost-per-answer, as we detail in the hidden reasoning-token tax.
Second, decide early whether you need auditable reasoning. If you are in a regulated domain or building anything where you may have to explain a decision, treat visible chain-of-thought as a feature and weigh it against the token savings of a compressed reasoner. This is a design decision, not a runtime tweak — it influences which model you build on.
Third, expect the models to keep pulling ahead of your intuitions about what is affordable. Work you shelved because reasoning was too expensive may be viable on the current generation, because the same problem now costs a fraction of what it did a year ago. It is worth periodically re-testing the ideas you rejected on cost grounds; the ground has moved.
The trade-off nobody should skip
There is a real cost to this, and it is not financial: interpretability.
A visible, English chain of thought is auditable. When a model reasons in words, you can read its work, catch a flawed assumption, and understand why it reached an answer. This matters enormously in regulated settings — a bank or a hospital often needs to explain a decision, not just make it. When reasoning moves into a compressed private notation or a continuous latent space, that transparency is largely gone. The model's "thinking" becomes a black box even to the people running it. You get the answer and the confidence, but not a legible trail.
This is an active and unresolved tension in the field. Some of the same researchers advancing latent reasoning are also working on ways to selectively decode it back into human-readable form for auditing, and safety teams have flagged that a model whose reasoning we cannot read is harder to oversee. For most commercial workloads the efficiency is worth it, but "we cannot see how it thought" is a genuine consideration for high-stakes and regulated use, and it should be a conscious choice rather than a default you discover later. If auditability is a hard requirement, a model with visible chain-of-thought is still the safer pick even at a higher token cost.
Where this is heading
The direction is clear even if the details are not settled. Natural-language reasoning was a brilliant accident — it made models better and, as a side effect, made their thinking readable. The next phase keeps the capability gain and drops the readability where it is not needed, because English was never the efficient medium for machine thought; it was just the one we could observe.
For anyone building on these models, the practical takeaway is simple. The price of intelligence is increasingly set not by how smart a model is but by how efficiently it thinks, and efficiency is improving fast. Expect the gap between sticker price and real cost to keep widening in favour of the models that have learned to say more to themselves in fewer words. And expect a parallel, quieter push to keep at least some of that reasoning legible — because a frontier we cannot read is a frontier we cannot fully trust.
For the concrete cost mechanics, see the hidden reasoning-token tax; for the model-versus-model view, see Fable 5 vs GPT-5.6; and for the broader shift, the efficiency race.