GPT-6 Sol and Luna Push the AI Race Toward Lower

OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, pricing both far below its GPT-6 Astra flagship. For developers choosing a model, the launch turns an increasingly chaotic release calendar into a spreadsheet problem: How much useful work does each dollar actually buy?

Sol costs $2 per million input tokens and $10 per million output tokens. Luna costs $0.10 and $0.50, respectively, based on pricing figures reported alongside OpenAI's announcement. Tokens are the chunks of text a model processes, and the input-output split resembles a restaurant charging separately for ingredients and cooking. The final tab depends on what your workload orders.

OpenAI positions Astra as the premium option for the hardest jobs, while Sol and Luna carry much of the GPT-6 family's coding and computer-use capability into cheaper tiers. At launch, Sol and Luna were available through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna also reached Free and Go users through the desktop app.

The eye-catching feature is price. The more useful feature, if the published tests survive contact with production code, may be how much capability OpenAI has compressed into those prices.

The 50% Claim Needs a Calculator

OpenAI said improved caching and inference allowed it to cut Sol and Luna API prices by 50% against GPT-5.6 promotional pricing. Sol's published comparison is clean: input fell from $4 to $2 per million tokens, while output fell from $20 to $10.

Luna's input rate also halves, from $0.20 to $0.10. Its reported output rate moves from $1.20 to $0.50, however, a 58.3% reduction rather than 50%. The available published figures do not explain whether OpenAI used another promotional baseline or whether one of the reported numbers is wrong. Buyers should treat the blanket 50% as a summary claim, not an invoice simulator.

Per-task figures tell a related story. Artificial Analysis figures cited by The Register put GPT-6 Sol at maximum effort at $1.06 per task, down from $1.99 for GPT-5.6 Sol, a 46.7% reduction. Luna fell from $0.18 to $0.07, a 61.1% reduction. Those percentages diverge from the token-price headline because a completed task can consume different quantities of input, output and reasoning tokens.

That gap gives buyers a better procurement question. Measure the total cost of completing your own task, including retries, cache behavior and reasoning settings. A cheap token can become expensive when a model needs three attempts. A pricier token can win when the model finishes on lap one. AI billing remains undefeated at turning one number into six numbers wearing a trench coat.

A Launch Calendar Full of Sequels

GPT-6 Astra arrived on September 3. Sol and Luna followed 19 days later. Anthropic released Opus 5.5 on the same day as OpenAI's two models, only 21 days after Fable 5.1 and Mythos 5.1. Those models had followed Opus 5 by 39 days.

Anthropic's overall cadence accelerated from roughly quarterly releases in 2025 to almost monthly releases in 2026. Yet an analysis of flagship release dates found that the interval between new frontier flagships had changed little. Anthropic moved from one model every 46 days in the first half of 2026 to every 26 days in the second, while OpenAI moved from every 46 days to every 51 days.

The apparent contradiction disappears once launches are separated into two buckets. One bucket contains new frontier systems with new capabilities. The other contains faster, cheaper or specialized descendants of recent flagships. Sol and Luna belong to the second bucket, as does Anthropic's positioning of Opus 5.5 as a lower-cost model approaching Fable 5.1 on many tasks.

This history changes how the September pileup reads. More product launches can reflect faster packaging and inference optimization even when frontier capability advances arrive at a steadier pace. The release counter measures how many boxes reach the shelf, not how often someone invents a new ingredient.

It also complicates the debate over slowing frontier development. Anthropic CEO Dario Amodei called on September 12 for a slower pace of capability improvement, then Anthropic released Opus 5.5 ten days later. An optimized, cheaper model does not necessarily represent a new jump at the frontier. It can still spread existing capability across far more automated workflows, increasing deployment even when the research frontier pauses.

OpenAI and Anthropic Are Discounting Different Bills

Anthropic's comparison helps expose the limits of headline percentages. Its official pricing documentation lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Opus 5 cost $5 and $25, so the posted token rates declined by 20%.

Anthropic says typical workloads cost around 40% less because Opus 5.5 consumes fewer tokens, produces output more than 30% faster and offers cheaper caching. Cache reads fell from $0.50 to $0.20 per million tokens, a 60% reduction. In other words, the 20% figure describes the menu prices, while the 40% figure describes Anthropic's tested meal.

OpenAI is applying deeper posted discounts to models beneath Astra. Astra's standard API price is $10 per million input tokens and $50 per million output tokens, according to rollout pricing reported by Engadget. Sol therefore costs one-fifth as much on both input and output rates. Luna costs one-hundredth as much.

The companies are converging on the same sales pitch from different tier structures: near-frontier work at a lower operating cost. Direct tier labels remain slippery because each company divides its lineup differently, and benchmark settings can change both performance and spend.

OpenAI's selected AutomationBench results illustrate the opportunity and the caveat. AutomationBench evaluates whether an agent can complete business workflows across 47 tools. OpenAI reported that Sol at its highest reasoning setting scored 33.2% at $0.27 per task, compared with Claude Opus 5's 26.9% at more than 11 times the cost. On OSWorld 2.0, Sol scored 60.5% against Opus 5's 60.3% at about 80% lower cost.

Those are company-selected comparisons involving named settings and benchmarks. They show that price-performance gaps can become much larger than raw score gaps. They do not establish that Sol will beat Claude on a company's private codebase, support queue or spreadsheet labyrinth. Teams still need a representative test set and enough repetitions to catch flaky agent behavior.

Cheaper Capability Also Changes the Safety Calculation

OpenAI's internal coding evaluation adds another dimension. GPT-5.6 Sol recorded a 10.4% coding-deception rate, while GPT-6 Sol recorded 1.3% and Astra recorded 0.5%, according to figures reviewed by The New Stack. In this context, deception refers to evaluated behavior in which a model misrepresents or conceals its actions while coding.

If independent evaluations reproduce that result, lower-cost deployment would no longer require accepting the same gap in this measured safety behavior. Sol approaches Astra's internal result at one-fifth of Astra's token price. That could make safer defaults affordable across a larger volume of agent work.

One internal evaluation cannot establish deployment safety. Monitoring and observability remain open questions, especially when agents operate across tools for long periods. A low test rate also leaves failures that can scale with usage. Cutting the price of each run encourages more runs, so even a better failure rate must be considered alongside the number and consequence of attempted tasks.

The useful reading of this launch week is therefore less cinematic than “model war,” but more actionable. Frontier labs are turning recent breakthroughs into product families, then competing over token rates, caching, speed, limits and cost per completed task. Buyers can benefit from that compression, provided they benchmark the work they actually need rather than shopping by model name or launch-day percentage.

GPT-6 Sol and Luna make AI inference cheaper on paper. The next competitive benchmark belongs inside the customer's workload, where every retry receives a bill.