Six point five. That is the number that traveled.
An item crossed the crypto wires last week — Crypto Briefing, no byline, headline length — reporting that Perplexity had built a training method called JOLT that burns 6.5x fewer tokens than GRPO while matching its accuracy. No benchmark was named. No dataset. No paper link. No ablation table. No statement about whether the code exists, whether it is open, or whether a single person outside the building has run it.
I have a reflex about announcements shaped like this. In 2017 I spent three months on Etherscan tracking wallet clusters that were moving tokens they had minted nine days earlier, and the tell was never in the whitepaper. It was always in what the announcement declined to specify. A precise number attached to a vague claim is not a technical result. It is a positioning statement with a decimal point bolted on.
None of that makes JOLT false. It makes JOLT unpriced. And in a bear market, unpriced things get bought on narrative and sold on the first filing.
GRPO — Group Relative Policy Optimization — came out of DeepSeek and became the load-bearing wall of the R1 line of reasoning models. Its contribution was narrow and enormous: instead of training a separate critic or value network to estimate the baseline for advantage, it samples a group of completions for the same prompt, scores them, and normalizes the rewards within the group. The group becomes its own baseline. The critic disappears. Memory falls, the pipeline simplifies, and the whole thing trains on hardware a mid-sized lab can actually rent.
That matters for reading the JOLT claim, because GRPO was already an efficiency innovation. It was not a capability breakthrough in the architectural sense. It was a way to buy the same reasoning behavior with fewer resources. Everything downstream of it, JOLT included, is competing inside an already-optimized envelope.
Now the accounting. A frontier model's cost stack has four large buckets: pre-training compute, post-training compute, inference, and the human labor that touches all three. Pre-training dominates absolute FLOPs. Post-training is where marginal reasoning capability now gets made, and it is also where measurement gets sloppy, because RL pipelines are heterogeneous — rollouts, reward scoring, reference-model log-probabilities, KL penalties, evaluation harnesses, data curation — and each consumes a different resource denominated in a different unit.
Perplexity is an answer engine. Subscription plus API, competing against OpenAI Search and Google's AI Overviews, with a brand built on being the technically serious alternative. A research-team narrative is worth real money to that company — in hiring, in enterprise sales, in the next round. I do not say that cynically. I say it because it tells you which task distribution a method like JOLT was probably tuned on: short, factual, citation-grounded answers. That is not the same distribution as long-horizon competition math, and the transfer assumption is doing a lot of unpaid work.
There is a broader reason this item landed where it did. Over the past eighteen months the retail capital pools for crypto and for AI have merged in the mind of the marginal buyer. When the crypto bid evaporates, the AI bid is the only one left standing in the same feed. Crypto media did not become an AI outlet. It became the fastest pipe into a market that needs something to price.
Here is the problem, and it is a units problem.
Tokens are a unit of text. They are not a unit of work. In a GRPO-style loop you sample a group of G completions per prompt, score each, and backpropagate. Each completion is a forward pass. Token count and forward-pass count correlate, but they are not the same number, and the gap between them is exactly where efficiency claims go to hide. Shorten average completion length and tokens fall. Sample more completions per prompt and filter aggressively and tokens-per-useful-sample falls while total generation work rises. Report the first number. Bury the second.
An efficiency claim is a statement about a denominator, and someone always chooses the denominator. Six point five is a ratio. A ratio without a stated numerator and denominator is a mood.
Four mechanisms could plausibly deliver that ratio, and they are not equally interesting.
Reward-signal density — extracting more gradient per rollout through shaping or reweighting — is real work and also the mechanism most likely to trade against alignment quality, because a denser proxy reward is a stronger invitation to reward hacking.
Sample filtering or a curriculum — spending rollouts only where the current policy sits near the decision boundary — is where the biggest genuine wins usually live, and also where the savings are most misleading, because filtering happens after generation. You paid for the discarded samples. They simply do not appear in the count of what you trained on.
Variance reduction in the advantage estimate is incremental and mathematically respectable and probably worth a factor of two on a good day. Not 6.5. GRPO's group normalization is already a variance-reduction trick; squeezing it further is polishing, not rebuilding.
And a stronger reward model does not eliminate cost. It relocates it. You move spend from the policy's rollout budget to the scorer's inference budget, and if the scorer is larger than the policy — which it increasingly is — the total may get worse while the headline metric gets better.
I learned that the expensive way. In 2022, interning at a Beijing fund, I watched a book lose fifteen percent of capital on positions whose risk model had counted all the right variables in the wrong units. The model was not wrong about volatility. It was wrong about what volatility was denominated in. Post-training budgets fail identically. I have since watched a team celebrate a rollout-token reduction that was fully offset by reward-model forward passes and a nightly evaluation suite nobody had entered into the cost sheet. The tokens went down. The invoice did not.
The benchmark is the next hole. "Matches GRPO accuracy" is a sentence with a gap in the middle where the task should be. Grade-school math is saturated — parity there means the method did not break anything, which is a low bar dressed as a high one. The benchmarks that discriminate now are competition math, competition code, graduate-level science, the ones where a handful of points separate a usable reasoning model from a demo. Efficiency papers have a well-worn habit of reporting parity on the saturated set and going quiet on the hard one. That is not fraud. It is selection, and selection is how honest people produce misleading headlines.
Then variance. RL runs are noisy, and a single-seed comparison reported as "matches" is a coin flip wearing a lab coat. Without error bars across seeds and without a stated evaluation protocol, the phrase conveys no information at all. I have signed off on enough internal benchmarks to know that the number people quote is usually the best of four runs, and the number they omit is the median.
Then memory. Group-relative methods compare within a batch. If JOLT needs a larger group, or multiple sampling rounds to do its filtering, activation memory and all-reduce traffic climb even as the token counter drops. For a frontier lab with a fat cluster this is an accounting footnote. For a three-person team on eight rented A100s it is the entire decision, because the binding constraint was never tokens — it was whether the job fits and whether the interconnect keeps up. A method that is 6.5x cheaper per token and 1.4x more expensive per GPU-hour is not a gift to small teams. It is a gift to large ones.
And then the reference class. JOLT is being benchmarked against GRPO. You do not benchmark against a straw man. The moment Perplexity chose GRPO as its comparison, it confirmed GRPO as the de facto standard of open post-training — something the ecosystem already knew from the TRL integration and the Open-R1 reproductions. It also means JOLT is a delta on a delta. GRPO eliminated the critic. JOLT, if real, shaves the rollout bill. Neither touches the fact that pre-training remains the dominant cost line for anyone chasing the capability frontier. The efficiency frontier and the capability frontier are different places, and this industry keeps pricing them as though they shared an address.
Which brings me to the metric that would actually settle this. Not tokens per update. Cost per solved problem — dollars and FLOPs and wall-clock, divided by the number of benchmark items the model newly gets right. That denominator cannot be gamed by shortening completions or discarding samples, because discarded samples still cost money and solved problems are binary. If Perplexity published that number against GRPO on competition math, the argument would be over in a paragraph. Nobody publishes that number. That is not an accident of formatting.
The most informative thing about this story is not the method. It is the channel.
An AI training technique was announced through a crypto outlet. That is not an accident of distribution; it is a routing decision, and it tells you who the message was built for. Crypto media publishes fast, asks few questions, and serves an audience structurally long on cost-decline narratives — because every decentralized compute marketplace, every GPU-token project, every training network with a governance forum has a chart that depends on two beliefs held simultaneously: that compute demand is infinite and that compute cost is collapsing. Both can be true over a decade. Neither can be verified by anyone reading three hundred words at two in the morning.
Here is the asymmetry that makes the routing work. The people who could check a 6.5x claim — the ones who have run a GRPO loop, who know what a group sampler costs in FLOPs, who have watched a reward model drift — do not read Crypto Briefing for AI research. The people who read Crypto Briefing cannot check the claim. Narratives live precisely in that gap, and the gap is not closing. Liquidity is a ghost, not a foundation, and narrative liquidity is the ghostliest kind there is.
There is a second decoupling worth naming, and it runs against the obvious read. Efficiency innovations do not lower the ceiling; they lower the floor. LoRA and QLoRA made fine-tuning accessible and did nothing whatsoever to slow the frontier labs' compute buildout, because the competition is over capability, not cost. Jevons applies with unusual force here: cheaper training produces more training. So even in the best case — JOLT is real, open, and independently reproduced — the effect on aggregate GPU demand is ambiguous and the effect on leading-lab spending is approximately zero.
What changes is who can afford to compete in the vertical layer. Legal reasoning, clinical triage, legacy code migration, the unglamorous domains where a domain-specific reward model beats a general one. That is a real shift with real margin consequences. It is just not the shift the headline implies, and it does not rescue a single token chart.
So the number to watch is not 6.5. It is whether anyone publishes wall-clock and FLOPs beside it, and whether the code lands on GitHub within thirty days of the paper.
If it does, the winners are the small teams doing domain RL, who get a cheaper post-training loop and a genuinely lower floor. If it does not, we learned something about the distribution channel and nothing about the method. Smart contracts don't validate accuracy claims; they validate collateral, and the only collateral here is a press item. Which of those two outcomes is the market currently pricing — look at the compute-token charts and answer honestly.
