New data point that stopped me cold: an 80% price drop on a frontier-class AI model, and the model itself did the work to make it happen. That’s not a typo. The smallest model in the GPT-5.6 family just got 80% cheaper because a bigger model found the savings.
I caught this in a breakdown from Matthew Berman, the creator who walks through OpenAI’s latest announcement and the artificial analysis charts behind it. He lays out something wild: OpenAI pointed its top model, GPT-5.6 Soul, at its own infrastructure to hunt for efficiency gains. The result is a price cut that reshapes the math for everyday builders.
Here’s the part I keep rereading. The expert explains that Soul, running inside Codex, kept analyzing production traffic, testing new routing strategies, and rewriting production GPU kernels on a loop. A model improving itself, around the clock. That’s recursive self-improvement in plain sight.
🔹 The numbers that matter
- GPT-5.6 Luna dropped 80%: now 20 cents per million input tokens, $1.20 per million output.
- GPT-5.6 Terra dropped 20%: $2 input, $12 output per million.
- Soul got no price cut, but a faster API mode: 2.5x speed for 2x cost, same intelligence.
The creator stresses a point I think too many people miss. It’s not the price per token that counts. It’s the cost per task completed. He shows that Kimi K3 looks half the price of Soul, but it burns twice the tokens to finish the same job. Effectively the same cost.
🔹 Why Luna is the sleeper hit
On the cost-per-task chart, the original poster shows Luna Max landing at roughly 6 cents per completed task. Compare that to GLM 5.2 Max near 26 cents, Grok 4.5 around 50 cents, and Claude Opus 5 low near 40 cents, all at similar intelligence. Luna sits in its own quadrant: high capability, tiny cost.
A few practical ways to use that gap:
- Move high-volume, repetitive tasks (classification, extraction, summarizing) to Luna Max and pocket the savings.
- Reserve the pricier Soul or Fable-class models for the genuinely hard reasoning jobs.
- Benchmark on cost-per-task, not sticker price, before you commit to any provider.
🔹 The strategy Berman spots
This is the insight I found most useful. The creator speculates that the big closed labs now run a repeatable playbook: train a massive, expensive model, then use it to bake smaller, near-as-good, far cheaper models for the rest of us. Soul stays the cash cow, so its price held while margins quietly improved. He even suggests labs may hold their absolute frontier models back, noting Anthropic reportedly had Fable months before releasing it.
Tips and pitfalls from the breakdown:
- Tip: Watch open-source pressure. The person who posted it credits models like Kimi K3 and GLM as the main force pushing these prices down.
- Pitfall: Don’t confuse cheaper-per-token with cheaper-per-task. Token-hungry models erase their own discount.
- Pitfall: Speed changes cost math too. A faster mode at the same multiplier is effectively a discount.
The bigger worry the expert raises is momentum. If the best models are already training the next best models, how does anyone without that compute catch up? He argues open-source is the only real competitive pressure left, and I think he’s onto something there.
Want the full chart walkthrough and his take on which subscription to pay for? Check out the complete video for the details.
🤖