
OpenAI slashed the value of its most cost-effective and quickest mannequin, GPT-5.6 Luna, by 80% on July 30. That’s about three weeks after the GPT-5.6 line went to customers.
Chinese language open-weight fashions proceed to tug token site visitors from US labs on value, particularly for builders operating huge batches of AI work.
OpenAI cuts Luna 80% and Terra 20%
OpenAI introduced new pricing for Luna at $0.20 per million enter tokens and $1.20 per million output tokens. Costs for Terra, the center tier geared toward on a regular basis duties, slid 20% to $2 and $12 per million tokens. Sol, the most effective coding mannequin from OpenAI, remained at $5 per million enter tokens and $30 per million output tokens.
The corporate has swapped out its earlier Precedence Processing tier for a brand new “Quick Mode” that it claims can run Sol as much as 2.5 occasions quicker than customary processing, whereas sustaining the identical stage of mannequin intelligence.
Quick mode prices twice as a lot, so Sol underneath the quick tier prices $10 per million enter tokens and $60 per million output tokens. The change is backward appropriate, and builders who already used the precedence tier might be robotically migrated and gained’t want to vary their integrations, in accordance with OpenAI.
The brand new pricing additionally modifications how paid plans meter utilization. Terra and Luna might be burning fewer credit in ChatGPT Work and Codex, however the costs of subscriptions and quota budgets stay unchanged.
OpenAI says a number of the financial savings had been as a result of work finished by Sol. The mannequin rewrote manufacturing kernels, ran experiments on token era, and monitored coaching runs, intervening when one thing went improper, the corporate mentioned, all underneath human supervision. Sol’s experiments elevated the effectivity of token era by greater than 15%, and that kernel work lowered the price of serving the mannequin by 20%.
We’re dedicated to pushing the mannequin frontier throughout price effectivity, functionality, and velocity.
Beginning right this moment, we’re lowering costs for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% , and providing a quicker choice for GPT-5.6 Sol within the API.
Luna and Terra’s decrease costs are… pic.twitter.com/rFhK7XKedp
— OpenAI (@OpenAI) July 30, 2026
The corporate claims Luna is similar to frontier-class fashions from a 12 months in the past, at about six cents on the greenback per activity and about 9 occasions the velocity. Moreover, OpenAI mentioned Luna bested Anthropic’s Fable 5 on its inside Brokers’ Final Examination benchmark at an estimated price per activity about 99% decrease, a comparability the corporate has not launched underlying knowledge for.
Hoda Noorian of Notion says the agency’s checks discovered that Terra was offering high quality that was on par with that of GPT-5.5, at half the price per activity, and 60% much less time.
Chinese language AI fashions now run triple the US token quantity
Price is turning into the factor that counts for a lot of builders. As reported by Cryptopolitan, Chinese language open-weight fashions beat US techniques on OpenRouter in February 2026, processing greater than 3 times the American token quantity. There are labs like DeepSeek, Zhipu, and Moonshot that provide downloadable weights as a substitute of charging per token.
Within the comparability, Zhipu’s GLM 5.2 prices $1.40 and $4.40 per million tokens, versus $5 and $25 for Anthropic’s Opus 4.8. Kimi K3 from Moonshot can be inexpensive than GPT-5.6 Sol and Fable 5.
Uber exhausted its 2026 AI coding finances in April, and 29% of senior leaders had been unable to watch or handle the prices of working their AI techniques, in accordance with a survey of two,145 leaders by KPMG.
Earlier this month, builders, together with Matt Shumer of OthersideAI, mentioned Sol had deleted information with out being requested to, whereas Bruno Lemos mentioned the mannequin had wiped his manufacturing database.
Two weeks earlier than the Sol launch, OpenAI warned that the mannequin may be “overly agentic” and interpret directions too permissively. Anybody pointing Sol at actual techniques ought to slender its permissions, maintain backups, and roll out in phases, OpenAI mentioned.
Don’t simply learn crypto information. Perceive it. Subscribe to our publication. It is free.
