OpenAI cut GPT-5.6 Luna prices by 80% and Terra by 20%, added a Fast mode running 2.5x quicker for twice the price, and claimed Luna beats Claude Fable 5 on Agents’ Last Exam at 99% lower cost per task (announcement). The saving is credited to an efficiency campaign in which Sol autonomously rewrote production GPU kernels and its own speculative-decoding drafts (efficiency post); the newsletter’s gloss is that the optimizer is now optimizing its own invoice. It closes the arc opened when GPT-5.5 topped KernelBench for writing the kernels it runs on, and follows Sol post-training Luna three weeks earlier. The same issue carries Dwarkesh Patel’s counter-argument that compute could get ten times more expensive, and two days later OpenAI’s finance chief folded the cut into an abundant-intelligence strategy.
Development
A model autonomously rewrites the kernels behind an 80% price cut
The point where the recursive loop reaches the rate card: an efficiency gain the model engineered for itself is passed straight through as a price cut, collapsing cost and capability into a single process.
Claim
OpenAI cut GPT-5.6 Luna prices by 80% and trimmed Terra by 20%, crediting an efficiency campaign in which Sol autonomously rewrote production GPU kernels and its own speculative-decoding drafts, and claiming Luna beats Claude Fable 5 on one agent benchmark at 99% lower cost per task.
Reported metric: -80% / -99% per task
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to July 30, 2026 (theinnermostloop.substack.com)
- Advancing the price-performance frontier with GPT-5.6 (openai.com)
- GPT-5.6: frontier intelligence, efficiency (openai.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- An agent cuts its own cost by 98% overnight
- Two API settings triple a benchmark score on six times fewer tokens
- A 276B model with 12B active sets a new open cost-performance frontier
- A lab touts an 80% price cut, a billion users, and agents writing 99.8% of tokens
- A training speedrun record falls to 75.4 seconds
- Self-improving agents rebuild the inference stack serving two open models
- Welcome to July 30, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.