OpenAI Launches GPT-6 Sol and Luna with 50% Price Reduction
OpenAI announced an expansion of its product line following the release of GPT-6 Astra, introducing GPT-6 Sol and GPT-6 Luna, which decentralize the reasoning, factuality, coding, computer usage, and alignment capabilities of the Astra generation into faster and cheaper models aimed at scaling everyday tasks.
API pricing is halved from the promotional price of GPT-5.6: Sol input drops from $4 per million tokens to $2, and output from $20 to $10; Luna input drops from $0.20 to $0.10, and output from $1.20 to $0.50. The company stated that the price reduction comes from improvements in caching and reasoning efficiency, and the new prices for Sol and Luna are permanent rather than promotional. Caching reads remain at about 90% of the price, while writes are about 1.25 times the uncached input price.
Astra was launched as the flagship earlier this month, with a standard price of about $10 per million tokens for input and $50 for output, with additional charges for long contexts exceeding 272,000 input tokens, Batch/Flex at half price, and Fast mode at double. Sol is positioned for repetitive complex tasks and software development, while Luna is aimed at high-throughput tasks such as summarization, extraction, classification, and routing. No mid-tier Terra model has been introduced this generation.
OpenAI claims that in internal factuality assessments based on real conversation error correction, GPT-6 Sol makes errors about half as often as GPT-5.6 Sol, approaching Astra's reliability. Third-party evaluations show that Sol has a slight increase in coding agent index, while Luna has a slight decrease, but the cost per single task is about halved: Sol costs about $1.06 compared to $1.99 for version 5.6, and Luna costs about $0.07 compared to about $0.18. The company also stated that the high reasoning tier Luna can approach the mid-tier performance of the old Sol at about 1% of the cost.
Starting today, writing is available in ChatGPT Work and Codex, aimed at Plus, Pro, Business, Enterprise, and Edu users; the API is also launched simultaneously, with models named gpt-6-sol and gpt-6-luna. Free and Go users can try Luna on desktop, but the standard chat window has not yet integrated it. Both models support text and image input, text output, with a context of about 1.05 million tokens and a maximum output of 128,000 tokens. Amazon Bedrock has also announced general availability.
In market mechanics, this is a share competition driven by supply-side cost reduction: the buyers are developers and enterprises looking to embed intelligence into repetitive workflows, while the seller is OpenAI, which reduces unit token costs through caching hit rates and reasoning stacks, passing the price difference to users. Funding shifts from customer API bills to larger-scale calls, rather than a one-time purchase of the flagship. Beneficiaries are high-frequency coding, customer service extraction, and internal automation; those under pressure are competitors still selling mid-tier capabilities at high prices, as well as OpenAI's own Astra profit pool that has not yet seen a price reduction. The event-driven nature is clear: Astra first occupies the intelligence ceiling, then Sol/Luna captures the cost curve.
Anthropic released a new version of Opus in a similar timeframe, with both companies continuing to compete in the same price range. With improved caching hit rates, the marginal costs of long conversations and agent workflows continue to decline, making it more cost-effective for large users than single flagship calls.
Source: Public Information
ABAB AI Insight
OpenAI has shifted from selling "the strongest model" to selling a layered universe, a path already laid out in the GPT-5.6 Sol/Terra/Luna preview: flagship models capture evaluations and high customer orders, while cheaper tiers capture call frequency. In July 2026, it first slashed Luna's standard price by about 80% and Terra by about 20%, keeping Sol unchanged; at the beginning of September, it reoccupied the ceiling with Astra, and three weeks later, it inherited the names Sol and Luna to GPT-6 and halved the promotional prices. Historical actions are consistent—first using new products to elevate capability narratives, then using reasoning and caching to make mid and low tiers scalable infrastructure.
The capital path is to spread the training and alignment investments across three SKUs: Astra maintains a high price scarcity of $10/$50, Sol takes on coding and multi-step agents at $2/$10, and Luna handles massive short tasks at $0.10/$0.50. The motivation is to lock in enterprise workflow entry points, making switching costs occur internally rather than leaking to Claude or open-source local deployments. Resource mobilization relies on reasoning stacks and prompt caching, rather than another round of pure parameter competition; the simultaneous launch of Bedrock means distributing through cloud channels to increase usage.
The analogy falls on how AWS broke down computing into reserved, on-demand, and Spot, as well as how Anthropic uses Haiku/Sonnet/Opus for layering. The industry phase has shifted from "who reaches AGI first" to control and penetration: Terra's absence indicates that mid-tier can be squeezed from both ends, with positional battles occurring at the Pareto frontier—who is cheaper under the same intelligence, and who makes fewer errors at the same price. The slight increase in the Coding Agent index for Sol and slight decrease for Luna indicates that the cheaper tier is not uniformly strengthening, but designed to be "adequate and scalable."
Structural judgment belongs to the overlap of technological substitution and the transfer of pricing power. The mechanism is: caching turns repeated context from full-price input into discounted reads, and reasoning optimizations lower the service costs of the same capabilities, thus "intelligence" is no longer sold at flagship hourly rates, but at adjustable work hours. Whoever masters caching hits and routing controls the enterprise bill; the flagship still sets the ceiling, but what truly rewrites market share is whether the half-price mid-tier can replace competitors' daily workflows.
ABAB News · Cognitive Laws
Flagship sets the narrative, mid-tier sets the bill
Caching turns repetitive labor from cost into discount
Price reduction is not a concession, but a means to squeeze competitors out of workflows