GPT-5.6 just made itself CHEAPER

📋 Overview

OpenAI cut prices across the GPT-5.6 family, and the headline number is dramatic: an 80% reduction on Luna, the smallest of the three models. But Matthew Berman's argument in this video is that the price sheet is the least interesting part of the announcement.

The interesting part is how the cuts became possible. OpenAI published a blog post crediting GPT-5.6 Soul — its own flagship model — with finding the efficiency gains that paid for them. A 20% reduction in serving cost from GPU kernel work, and a 15% improvement in token generation efficiency from better speculative decoding, discovered by the model, one day after it was deployed.

Berman reads that as the opening move of a loop: a frontier model running 24/7 on frontier compute, examining production data, proposing improvements, testing them, and iterating. That is recursive self-improvement in a narrow but real sense — and it reframes the entire competitive picture, including why the flagship model conspicuously did not get cheaper.

1. The 80% Price Cut

Sam Altman announced major price cuts across the GPT-5.6 lineup. The family has three tiers, in ascending order of capability: Luna < Terra < Soul. Luna, the smallest, took the deepest cut — 80% off, bringing it to $0.20 per million input tokens and $1.20 per million output tokens.

That puts a closed frontier-lab model beneath the frontier open-source models coming out of China on raw token pricing, which is not a position anyone expected OpenAI to occupy.

2. Terra and the Fast-Mode Sleight of Hand

Terra, the middle model, received a more modest 20% cut, landing at $2 per million input tokens and $12 per million output tokens.

Soul, the flagship, received no price drop at all. What it received instead was a change to its fast mode. Previously, fast mode delivered roughly 1.5× the speed for 2× the cost. Now it delivers 2.5× the speed for that same 2× cost — with identical intelligence either way.

Berman argues this amounts to a price decrease in everything but name. If you were already paying the fast-mode premium, you are now getting substantially more throughput for the same money, which lowers your effective cost per unit of work even though the number on the pricing page never moved.

3. Reading the Intelligence-vs-Cost Chart

To make the point concrete, Berman walks through the Artificial Analysis intelligence index chart. The y-axis is the intelligence score; the x-axis is cost per task. The position you want is high and to the left — maximum capability, minimum spend.

GPT-5.6 Luna Max lands in territory that Berman describes as almost absurd. It scores slightly ahead of GLM 5.2 Max, the Chinese open-source contender, on intelligence — while costing a small fraction of it per task. Claude Opus 5 in its low-reasoning configuration scores slightly below both, at the highest cost of the three.

Same tier of intelligence, roughly a 4× to 7× spread in cost. That gap is the entire story of the announcement compressed into one scatter plot.

4. The Model That Optimized Itself

Here is the part Berman considers the actual story. OpenAI took GPT-5.6 Soul and pointed it at its own infrastructure — one day after deployment — to make itself more efficient. It worked.

OpenAI published a blog post explicitly crediting GPT-5.6 Soul with the discoveries. This is not a case of engineers using an AI assistant for autocomplete; it is the company attributing the findings to the model.

Soul in Codex "helps us analyze production traffic, identify previously overlooked sources of imbalance, test new routing strategies, and constantly tune these heuristics."

Beyond routing, the model optimized its own forward pass — identifying work that could be precomputed, avoided entirely, or parallelized — and rewrote production kernels. The phrase worth sitting with is "previously overlooked": these were inefficiencies human engineers had not caught.

5. This Is a Loop — And That Makes It Recursive

Berman's framing is that none of this is a one-time optimization pass. It is a loop: constantly examine production data → find inefficiencies → propose improvements → test them → iterate. And unlike a human engineering team, that loop runs 24/7, is highly parallelized, and is powered by frontier-model compute.

By his reading, that is recursive self-improvement. Not the science-fiction version, but the practical one: a model making itself cheaper and faster to run, which frees compute, which funds more of the same.

6. The Karpathy Parallel

One detail from OpenAI's post goes further than kernel tuning. Soul, per the company, improved its own draft model by designing and running hundreds of experiments on its architecture, testing changes in size, structure, and features.

Berman connects this directly to Andrej Karpathy's open-source auto-research repository. That project set up a loop whose objective was improving the efficiency of training a language model from scratch. The model designed its own experiments, ran them, analyzed the output, and iterated — and it surfaced real improvements that Karpathy himself had not been able to find.

The difference in scale is the point. Karpathy ran his loop on a small language model with hobbyist-scale compute. OpenAI is running the same shape of loop with a frontier model, on frontier compute, continuously. What that describes is an automated AI researcher.

7. Why Soul Didn't Get Cheaper

If GPT-5.6 Soul found 20% serving-cost savings, why did Soul's price stay flat while Luna dropped 80%?

Berman's theory: OpenAI almost certainly did find efficiency gains for Soul too — but Soul is the cash cow. The majority of revenue comes from the best models, not the workhorses. Efficiency gains on the flagship therefore go straight to margin rather than being passed to customers, while the gains on the cheap tier are handed over as a competitive weapon.

On what prompted such aggressive pricing at the bottom of the stack, his answer is Kimi K3 — the Chinese open-source model sitting near the frontier at dramatically lower cost. The pressure is coming from below.

His deeper point is structural: once an open-source model is in the wild, every inference provider and independent researcher who touches it starts finding efficiency improvements. That distributed effort surfaces discoveries faster than any small closed team can match. OpenAI, in Berman's reading, is getting ahead of a curve it cannot win by waiting.

8. The Playbook: Train Big, Serve Small

The closing chart is cost per task on the intelligence index, and the spread is startling.

All of these sit at roughly comparable intelligence — and Luna Max delivers it for about six cents.

Berman's speculation about what this implies: the frontier labs' playbook is now to train massive, inefficient models, then use those models to bake almost-as-good but far cheaper models for public consumption. The giant model is a factory, not a product.

He suspects Anthropic may stop releasing its absolute frontier models publicly at all, or delay them for a long time — noting that Anthropic had Fable internally in January, months before the public release. The loop he sketches: train the biggest model → distill cheaper models from it → earn revenue from those → fund the next big model, while the biggest model is simultaneously training its own successor.

🎯 Key Takeaways

• The cut is real but uneven. Luna dropped 80% to $0.20/$1.20 per million tokens, Terra 20% to $2/$12, and Soul got nothing on price — only a fast mode upgraded from 1.5× to 2.5× speed at the same 2× premium.
• Token price is the wrong metric. Berman insists on cost per completed task; Kimi K3 is half the per-token price of Soul but uses twice the tokens, netting out to the same cost.
• The model paid for its own discount. GPT-5.6 Soul found a 20% serving-cost reduction in production GPU kernels and a 15% token-generation efficiency gain via speculative decoding — one day after deployment.
• It is a loop, not a one-off. Examine production data, find inefficiencies, propose, test, iterate — running 24/7 on frontier compute. Berman calls this recursive self-improvement and openly worries about whether anyone can catch up.
• Soul stayed expensive because it is the cash cow. Gains on the flagship go to margin; gains on the cheap tier go to fighting Kimi K3 and the open-source pressure from below.
• Train big, serve small. Luna Max delivers comparable intelligence at $0.066/task against Claude Fable 5's $2.75 — evidence, in Berman's view, of a playbook where huge models exist to produce cheap ones, and open source is the only remaining check on it.

⏱️ Timestamp Index

• 0:25 The 80% Price Cut
• 1:41 Terra and the Fast-Mode Sleight of Hand
• 2:33 Reading the Intelligence-vs-Cost Chart
• 3:53 The Model That Optimized Itself
• 5:40 This Is a Loop — And That Makes It Recursive
• 7:53 The Karpathy Parallel
• 9:38 Why Soul Didn't Get Cheaper
• 12:47 The Playbook: Train Big, Serve Small