OpenAI set its own flagship model to work on improving its serving infrastructure after deployment — and today passed the results to customers as the largest price cut the GPT-5.6 family has seen. GPT-5.6 Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from the $1.00 and $6.00 it carried at launch on July 9 — an 80% reduction across both rates, effective immediately, confirmed by CNBC. GPT-5.6 Terra falls 20%, to $2.00 input and $12.00 output from $2.50 and $15.00. GPT-5.6 Sol, the flagship, is unchanged at $5.00 and $30.00.
For any developer whose production traffic runs through Luna — classification, document extraction, request routing, the first-pass drafts and long agent loops where one user instruction fans out into dozens of model calls — the arithmetic changed today. The same calls that cost $1.00 to input yesterday cost $0.20 this morning.
Sol Rewrote Its Own Code. Here Is What That Means for Pricing
The engineering story behind the cut is what separates this announcement from a routine pricing move. In a post published yesterday, OpenAI engineers described a self-optimization loop they ran after GPT-5.6 Sol reached general availability.
The company set Sol to work inside Codex — OpenAI’s own coding environment — on the GPU kernels that underpin its inference infrastructure. Kernels are the low-level programs that execute the mathematical operations at the core of every transformer model: matrix multiplications, attention computations, activation functions. Sol rewrote and optimized those kernels using Triton and Gluon, two open-source GPU programming languages that OpenAI maintains. Combined with broader kernel improvements that Sol identified through the same process, the effort cut end-to-end serving costs by 20%. OpenAI used its own FpSan (Floating-Point Sanitizer) tool — open-source — to validate that the model-generated kernel code was mathematically correct.
Sol also redesigned its own speculative-decoding draft model. Speculative decoding is a standard inference optimization: a smaller, faster draft model predicts candidate output tokens, and the larger target model verifies them in parallel rather than generating one token at a time. When a draft model improves its prediction accuracy, the verification acceptance rate rises and effective throughput increases — without new hardware. Sol ran hundreds of experiments to improve its draft model, raising token-generation efficiency by more than 15%.
These gains compound. Better kernels reduce the compute required per forward pass. Better speculative decoding reduces the number of forward passes required per output. Both improvements reduce serving cost simultaneously, which is how a 20% and a 15%+ gain translate into the headline 80% price reduction for a tier whose margins were already thinner than Sol’s.
OpenAI also announced that the price reductions for Luna and Terra are reflected in how usage is counted in Codex and ChatGPT Work subscriptions — meaning subscription customers whose plans route to the cheaper models will see their allotted usage stretch further at no additional cost.
What Changed at Every Pricing Tier
Luna’s new standard rate — $0.20 input / $1.20 output per million tokens — cascades through every processing tier. Batch and Flex processing, already discounted at half the standard rate, now price Luna at $0.10 input and $0.60 output per million tokens. Cached input reads, discounted 90% from standard, fall to $0.02 per million tokens on Luna. For teams building agent loops with large stable system prompts, cached-input costs under the new rate are negligible.
Terra’s revised rate — $2.00 input / $12.00 output — positions it below Anthropic’s Claude Sonnet 4.6, which currently carries $3.00 input and $15.00 output per million tokens.
At the low end, Luna now undercuts Anthropic’s cheapest published model, Claude Haiku 4.5, by a factor of five on input and roughly four on output, according to Anthropic’s pricing page.
The competitive picture at the mid-tier is more dynamic. Anthropic’s Claude Sonnet 5 carries an introductory rate of $2.00 input and $10.00 output per million tokens — pricing that runs until August 31, 2026, after which it is scheduled to rise to $3.00 and $15.00. During that window, Terra and Sonnet 5 trade near-parity. After August 31, the gap could widen substantially.
OpenAI also announced a new Fast mode for Sol in the API, replacing the prior Priority Processing offering. Fast mode delivers up to 2.5 times the speed of standard processing at twice the standard price, with no change in model intelligence. Existing API calls tagged for priority automatically route to Fast mode without requiring code changes.
OpenAI additionally announced that Auto-review in the ChatGPT app and Codex CLI is being upgraded from GPT-5.4 to GPT-5.6 Luna. Combined with Luna’s new price, OpenAI says this should make Auto-review approximately ten times less expensive, reducing the cost of agentic workflows for teams that rely on it.
Luna at $0.20: Which Workloads Can Developers Run Now?
Luna’s original price was already OpenAI’s lowest, but $1.00 per million input tokens still put certain high-volume automation workloads at the margin of commercial viability. At $0.20, the arithmetic changes for applications that produce large token volumes per task.
Blitzy’s CTO and co-founder Sid Pardeshi said Luna processes 2.2 times more context with 8.5 times fewer output tokens compared with GPT-5.4 mini, at 87% lower cost. Dust co-founder and CTO Stanislas Polu reported that Luna runs 40% faster and 40% cheaper than their previous default model on the same agentic tasks. Replit’s president and head of AI, Michele Catasta, described Luna as "the closest we’ve come to intelligence too cheap to meter" — a phrase that signals a qualitative shift in how the company thinks about which use cases are worth building.
The practical consequence for development teams is that the three-tier routing architecture OpenAI recommends — Sol for the hardest reasoning and coding work, Terra for steady production tasks, Luna for high-volume routine calls — now carries a larger price gap between tiers. A workload that was borderline for Luna at $1.00 input is decisively economical at $0.20.
Why Does OpenAI Cut a Three-Week-Old Model’s Price?
The GPT-5.6 family had been on the market for 21 days when OpenAI cut it. That timeline reflects two pressures converging simultaneously.
The first is enterprise cost fatigue. Enterprises have been applying mounting pressure on AI vendors to demonstrate measurable return on investment. Uber burned through its entire annual AI budget in four months before introducing spending tiers on some internal tools. Amazon’s engineering organization separately moved to cap AI spending after cost overruns in the same period. OpenAI CFO Sarah Friar told employees at an internal all-hands meeting on Wednesday that July’s annualized recurring revenue alone exceeded the entire second quarter — crediting the GPT-5.6 model family, ChatGPT Work, and wider adoption of Codex. No specific dollar figures were disclosed. Board chair Bret Taylor joined the meeting and acknowledged that developers who leaned hard on Claude Code eventually ran up large bills and began shopping for alternatives.
Those alternatives increasingly come from China. Moonshot AI, the Beijing-based lab backed by Alibaba, released Kimi K3 on July 16 — a 2.8-trillion-parameter open-weight model that placed near the top of multiple third-party benchmarks. Ion Stoica, a computer science professor at UC Berkeley and co-founder of both Databricks and the Arena benchmark, said the gap between Chinese open-weight models and US frontier labs had narrowed from six to nine months to roughly two to three months.
The OpenRouter usage data shows where that narrowing translates into actual production decisions. Since February 8, 2026, the share of tokens US companies route to Chinese models on the platform has stayed above 30% every week, reaching 46% in some weeks, compared with a 12-month average of just 11% before that date. Chinese models on some workloads carry token rates up to nine times lower than equivalent US frontier systems.
OpenAI’s strategic response is legible: defend the high-volume, cost-sensitive workloads at the Luna and Terra tiers — the classification tasks, document extraction, request routing, and long agent loops that represent the bulk of production traffic — while preserving Sol’s positioning as the premium unchallenged flagship. For developers routing bulk traffic to Luna, the same calls now cost one-fifth what they did three weeks ago.
IPO Context and the Efficiency Flywheel
Both OpenAI and Anthropic filed confidential S-1 prospectuses for public listings in June 2026. Anthropic disclosed in May that its revenue run rate had reached $47 billion, up sharply from approximately $10 billion across all of 2025, driven in large part by Claude Code. OpenAI is separately negotiating with Nvidia on a compute backstop potentially worth up to $250 billion, tied to a planned 10-gigawatt data center campus in southern Ohio. Against those capital commitments, efficiency gains in serving costs are directly relevant to the company’s margin trajectory.
The engineering post published yesterday frames Sol’s self-optimization work as the beginning of a feedback loop rather than a one-time exercise. OpenAI engineers wrote that as models improve and can work more autonomously, the company’s ability to accelerate efficiency gains itself accelerates — a compound dynamic. If that loop holds, the pace at which future price reductions become feasible may be set in part by how well the current generation of models can improve their own infrastructure — not only by hardware procurement cycles or human engineering investment. This is the first publicly documented instance of a production frontier model autonomously rewriting its own serving-stack code and having that rewrite verified, accepted, and translated into published customer pricing.
Frequently Asked Questions
What are the new prices for GPT-5.6 Luna and Terra?
Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from $1.00 and $6.00. Terra drops to $2.00 input and $12.00 output from $2.50 and $15.00. Sol is unchanged at $5.00 input and $30.00 output. Batch and Flex processing, discounted at half the standard rates, now price Luna at $0.10 input and $0.60 output. Cached input reads on Luna fall to $0.02 per million tokens. Changes are live in the OpenAI API as of today, July 30, 2026, with AWS Bedrock pricing beginning to roll out during the day.
How did OpenAI cut Luna’s price by 80% just three weeks after launch?
OpenAI set GPT-5.6 Sol to work inside Codex on its own production GPU kernels after the model reached general availability. Sol rewrote those kernels in Triton and Gluon, OpenAI’s open-source GPU programming languages — reducing end-to-end serving costs by 20%. Sol also redesigned its own speculative-decoding draft model through hundreds of autonomous experiments, raising token-generation efficiency by more than 15%. Speculative decoding works by having a smaller draft model predict output tokens that the larger model verifies in parallel, so an improved draft model means more verified tokens per forward pass and lower per-token cost. These two gains, compounding in the same serving stack, created the margin headroom OpenAI passed to customers as lower prices.
Does this change how I should architect agentic workflows?
Yes, for teams running high-volume agent loops. At $0.20 input per million tokens, Luna is no longer a niche choice for simple, repetitive calls — it is economical for workflows where each user request fans out into dozens of model calls before returning an answer. Blitzy reports that Luna processes 2.2 times more context with 8.5 times fewer output tokens compared with GPT-5.4 mini, at 87% lower cost. Teams that have been routing all such traffic to Terra or Sol for quality reasons should re-evaluate: Luna today delivers performance that was frontier-class one year ago, according to OpenAI’s own benchmarking, at approximately six cents on the dollar per task compared with that prior generation.
What does it mean that Sol optimized its own infrastructure?
This is the first publicly documented case of a production frontier model autonomously rewriting its own serving-stack code — the kernels that run the model’s own computations — and having that rewrite verified, accepted into production, and translated into actual customer pricing. Prior AI optimization work of this kind was done by human engineers or applied to simpler serving configurations. If the efficiency loop compounds as OpenAI’s engineers describe — where each generation of models helps optimize the infrastructure that runs the next — it raises the possibility that future AI price trajectories may be partially determined by how well a model can improve its own stack, not only by new hardware or human engineering time. That is a shift in how AI infrastructure cost curves have historically worked.
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.