Three weeks. That is how long GPT-5.6 Luna's original price list survived.
OpenAI shipped the GPT-5.6 family on July 9, 2026. On July 30 it cut the price of Luna, the cheapest tier, by 80%, and the price of Terra, the mid-tier, by 20%. The flagship, Sol, was left exactly where it was.
Price changes at frontier labs are usually the boring part of a launch post. This one is not, for two reasons. The first is the size of it: an 80% cut on a model that has been generally available for three weeks is not a routine adjustment, and it lands in the middle of the most brutal cost war the API market has ever seen. The second is the explanation OpenAI gave for how it could afford to do this.
The company says GPT-5.6 Sol wrote the code.
The New Numbers
The change took effect July 30 across the API, with the AWS rollout following the same day.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Change |
|---|---|---|---|
| GPT-5.6 Luna (old) | $1.00 | $6.00 | |
| GPT-5.6 Luna (new) | $0.20 | $1.20 | 80% cheaper |
| GPT-5.6 Terra (old) | $2.50 | $15.00 | |
| GPT-5.6 Terra (new) | $2.00 | $12.00 | 20% cheaper |
| GPT-5.6 Sol | $5.00 | $30.00 | unchanged |
The cut travels beyond the raw API. Terra and Luna now consume fewer credits inside ChatGPT Work and Codex, while subscription prices and quota budgets stay put. In practice, subscribers get more work out of the same plan without anyone touching their invoice.
Sol got a different kind of change. OpenAI introduced Fast mode, which replaces Priority Processing in the API. Fast mode delivers up to 2.5 times the speed of standard processing at twice the price, with no change in the underlying intelligence. Requests already tagged priority route to Fast mode automatically, and the setting matches /fast in Codex.
For teams doing capacity planning, the shape of the offer is now unambiguous: the bottom of the ladder got dramatically cheaper, and the top of the ladder got an optional speed upgrade you pay extra for.
The Model Optimized the Stack That Serves the Model
OpenAI's stated reason for the cut is efficiency, and the specifics are the most interesting part of the announcement.
Inside what the company describes as a human-led process, GPT-5.6 Sol autonomously rewrote and optimized production GPU kernels, the low-level routines that decide how efficiently a model's math actually runs on the hardware. Sol also designed and ran hundreds of experiments aimed at improving token generation, and monitored training runs, stepping in when problems came up.
The reported results: the kernel work cut the end-to-end cost of serving the model by 20%, and the experiments improved token-generation efficiency by more than 15%.
OpenAI frames this as a compounding loop. Better models find the next round of efficiency gains faster, which lowers cost, which funds broader deployment. The company put it plainly in its post: as models improve and work more autonomously, "our ability to improve efficiencies accelerates."
Read that claim skeptically and it is still notable. A frontier lab is now attributing a specific, quantified chunk of its own infrastructure savings to work its model performed on the systems that run it. Whatever else is true, the marketing has moved from "our model writes code for you" to "our model writes the code that runs our model."
Customers Are Reporting Numbers, Not Adjectives
OpenAI's launch post carries testimonials, which normally means nothing. These ones carry arithmetic.
Sid Pardeshi, CTO and co-founder of Blitzy, described restructuring around Luna in concrete terms: the switch moved his team from a single structured-output call to a full tool-calling agent loop, and pushed prompt-cache reuse from 24% to 90%. Across thousands of production calls, he said, Luna handles 2.2 times more context with 8.5 times fewer output tokens, at 87% lower cost than GPT-5.4 mini.
Hoda Noorian, who works on AI product at Notion, said Terra delivered quality comparable to GPT-5.5 in the company's evaluations at half the cost per task and in 60% less time.
Stanislas Polu, co-founder and CTO of Dust, reported Luna running 40% faster and 40% cheaper than the team's previous default model on the same agentic tasks. Shaiyon Hariri, creator of Ramp's SWE-Bench work, said Luna is now the default for background agent automations.
Michele Catasta, president and head of AI at Replit, reached for the oldest phrase in the industry:
"GPT-5.6 Luna is the closest we've come to intelligence too cheap to meter. I've never seen a model this affordable be this powerful." — Michele Catasta, President and Head of AI, Replit (OpenAI, July 30, 2026)
OpenAI's own benchmark claim is the aggressive one. On Agents' Last Exam, a professional-work evaluation, the company says Luna outperforms Anthropic's Fable 5 at an estimated cost per task nearly 99% lower. That is a vendor-published comparison against a competitor, and it should be treated as one until independent testers reproduce it.
The Pressure Is Coming From China's Open Weights
The competitive backdrop makes the timing hard to read as coincidence.
Chinese-origin models have taken at least 30% of enterprise token volume on OpenRouter every single week since February 8, 2026, peaking at 46%. The twelve-month average before that stretch was 11%. LDS covered the reversal in A Startup Moved 100% of Its AI Traffic From Claude to DeepSeek. The Token Data Says It Won't Be the Last., where the AI startup Lindy told CNBC it had moved all of its traffic off Claude to save millions.
Open-weight Chinese models undercut American frontier APIs by wide margins, a gap documented in LDS coverage of DeepSeek V4 and GLM-5.2. At twenty cents per million input tokens, Luna is now in a range where that comparison is at least a conversation rather than a rout.
The contrast with OpenAI's own recent history is sharper still. In April, the company shipped GPT-5.5 six weeks after GPT-5.4 and doubled the API price. Fifteen weeks later it is cutting the entry tier by four-fifths. The direction of pricing power over that stretch is not subtle.
The Analysts Think Your Bill Goes Up
Nobody who studies enterprise AI budgets expects this to save anyone money.
Pareekh Jain, principal analyst at Pareekh Consulting, told InfoWorld that the effect will be scale, not savings. "For CIOs, the biggest impact is likely to be scaling AI adoption rather than simply cutting costs or lowering AI budgets," he said. Lower unit prices make it easier to push pilots into production and to justify agentic workflows that fire off many model calls per task. Jain expects most enterprises to reinvest the savings into higher usage rather than shrink their budgets, and named the mechanism: the Jevons paradox, where efficiency gains raise total consumption instead of reducing it.
Chandrika Dutt, research director at Avasant, made a similar point to InfoWorld, expecting teams to use the headroom to build agentic workflows that were previously hard to justify economically.
Jain also argued the trend has legs. "These price decreases are broadly sustainable in the long run," he said. "New chips, better software, and more efficient model designs will keep driving down the cost per token, and heavy competition makes big price hikes risky for any single provider and thus unlikely."
His practical advice for teams: design for model portability. If per-token costs keep falling unevenly across vendors, the architecture that wins is the one that can swap providers without a rewrite.
The Bottom Line
An 80% price cut three weeks after launch is not generosity and it is not a rounding error. It is a signal about where the floor is.
The floor is being set by open-weight competitors that most American enterprises were not routing to eighteen months ago and now send at least 30% of their OpenRouter tokens to every week, peaking at 46%. OpenAI's answer is to push its cheapest tier down to a price where that arbitrage stops being obvious, and to pay for it with efficiency gains that its own flagship model helped engineer.
For practitioners, the immediate move is unglamorous: rerun your cost model. A workflow that was uneconomical at a dollar per million input tokens may be trivially affordable at twenty cents, and the routing decisions your team made in July are already out of date. Sol for planning, Luna for execution is now a materially different calculation than it was three weeks ago.
The longer question is what happens when the loop OpenAI described actually closes. If each generation of model meaningfully lowers the cost of serving the next, the price of intelligence stops being set by chip supply and starts being set by how fast a lab's own models can optimize its stack.
Re-benchmark before you re-architect. Vendor-published cost-per-task comparisons are directional at best; run your own evaluations on your own workloads, then decide which steps genuinely need the flagship and which can be handed to the cheap tier. Terra now costs ten times what Luna does on input tokens, so the line you draw between them is worth real money at volume.
Sources
- Advancing the price-performance frontier with GPT-5.6 (OpenAI, Jul 30, 2026)
- OpenAI drops GPT-5.6 Luna and Terra API prices by up to 80% (InfoWorld, Jul 31, 2026)
- OpenAI cuts prices for two of its GPT-5.6 AI models as companies grow sensitive to costs (CNBC, Jul 30, 2026)
- AI price wars: OpenAI cuts GPT-5.6 Luna prices by 80% as model competition shifts toward cost (VentureBeat, Jul 30, 2026)
- OpenAI Cuts Prices on Select Models to Make High-Volume Work Economical (PYMNTS, Jul 30, 2026)
- Chinese AI models overtake US peers in token consumption, OpenRouter data shows (Dealroom, Jul 2026)
- GPT-5.6 (OpenAI, Jul 9, 2026)