GPT-6 Sol and Luna Launch: Half the Price, Matching Opus Performance

OpenAI has once again triggered a dual wave of model iteration and price competition. According to analysis from a Bilibili content creator, the newly released GPT-6 Sol and Luna models deliver intelligence on par with the previous generation at half the cost — putting competitors Anthropic's Opus series and DeepSeek in an awkward position. The defining theme of this update isn't raw "intelligence" anymore; it's "lighter, faster, and cheaper."

Price War: Luna Now Undercuts DeepSeek

The most immediate impact of this release is the pricing. GPT-6 Sol matches the intelligence of the previous generation while cutting its price in half. Luna goes even further — its pricing has dropped below DeepSeek^Ad V4 Pro's off-peak, cache-miss rate.

GPT-6 Luna Max takes things to an extreme: its per-task cost is reportedly just one-quarter that of DeepSeek V4.1 Flash. Keep in mind that V4.1 Flash had long been regarded as the gold standard for price-to-performance ratio — that crown has now been directly challenged. This restructuring of cost fundamentals means OpenAI is no longer competing solely in the premium flagship segment; it's actively encroaching on the budget tier that domestic Chinese models have historically dominated.

For enterprise users making large-scale API calls, a four-fold reduction in cost essentially multiplies the scale of tasks they can affordably run — and that's where the real impact of this price war lies.

Background: DeepSeek V4.1 Flash is a lightweight inference model from DeepSeek, designed for high-concurrency, low-latency API use cases. It had earned a reputation in the developer community as the "peak price-performance" option thanks to extremely low input/output token pricing. DeepSeek's pricing typically splits into two tiers — cache hit and cache miss — with the former significantly cheaper than the latter. The article's comparison of Luna undercutting DeepSeek V4 Pro even at off-peak cache-miss rates means Luna holds a price advantage even in DeepSeek's least favorable pricing scenario, which makes this a particularly relevant benchmark for enterprise procurement decisions.

Real-World Task Performance Reaches the Top

Being cheap alone isn't enough to claim a decisive win. What really puts competitors in a difficult spot is that GPT-6 also delivers strong performance on real-world tasks.

In the Automation Bench evaluation, GPT-6 Luna at higher reasoning intensity matched GPT-5.6 Sol at approximately 1% of the cost. Across automated workflow evaluations covering sales, operations, customer service, finance, and HR, GPT-6 Sol High scored 33.2% — surpassing Claude Opus 5 at maximum reasoning intensity, which scored 26.9%.

What's worth examining more closely is the cost comparison: GPT-6 Sol High achieved a higher score while costing only one-tenth as much per task as Opus 5. Better performance at a fraction of the price is a near-total domination of enterprise automation use cases. This explains why some have described this as "targeting Opus 5.5" — OpenAI is squarely aiming at the enterprise automation market that Anthropic values most.

Background: Automation Bench is a benchmark designed to evaluate AI performance on real enterprise automation tasks, distinct from traditional academic benchmarks like MMLU or HumanEval. It simulates multi-step workflows in actual business processes, including sales lead handling, operations data integration, customer service ticket routing, financial report generation, and HR screening. The core metric is end-to-end task completion rate rather than single-step accuracy, making it a more meaningful measure of a model's practical value in complex workflows. The benchmark also tracks token consumption per task, making "price-to-performance" (completion rate per unit cost) a quantifiable comparison dimension — which is the context behind the "one-tenth the cost of Opus 5" figure.

Lighter and Faster: Real Gains in Coding Efficiency

The key differentiator of this update, as the content creator puts it, is "not intelligence — it's being lighter and faster."

On real-codebase software engineering benchmarks, GPT-6 Sol nearly matches Claude Fable 5's top score. Developer Flavio Damo, who had early access, shared a compelling data point: tasks that previously took about an hour on GPT-5.6 now regularly complete in around 20 minutes, using less than half the tokens.

Three times faster, half the cost — for developers relying on AI-assisted coding, this is a concrete productivity shift. When the model is no longer the bottleneck and call costs drop significantly, both the frequency and the scope of AI-assisted programming expand naturally.

Additionally, OpenAI product lead Tibo Sotio announced that Plus, Pro, and Business accounts will receive a one-time usage quota reset, allowing subscribers to get an extra round of usage within their current billing cycle. This move is clearly designed to accompany the new model launch and encourage users to try the new models quickly.

Availability and Subscription Options

GPT-6 Sol and Luna are now available to Plus, Pro, Business, Enterprise, and Azure users. API users can access the new models directly using the GPT-6 Sol and GPT-6 Luna model identifiers.

For users in mainland China, two approaches were outlined. The first is registering and subscribing via the web — a relatively straightforward process involving switching language settings, registering an email, and completing verification. The second involves purchasing through a mobile app store using gift cards, though this requires more technical know-how and a suitable network environment.

The content creator noted that the Android method has a success rate of only around 40% due to environmental and device constraints, and is not recommended. After a successful gift card payment, it is normal for the balance to take 5 to 10 hours to be credited, so leave enough time.

Note: Third-party subscription channels carry account and payment risks. Readers should assess compliance and security implications independently and proceed with caution.

Final Thoughts

The launch of GPT-6 Sol and Luna reflects a clear strategic direction: build a competitive moat through cost and speed while maintaining intelligence. Half the price, enterprise task performance surpassing Opus, and coding efficiency doubled — these three cards played simultaneously bring both the price war and the performance war directly to competitors.

As model intelligence approaches saturation, cost efficiency and response speed are becoming the defining competitive variables of the next phase. For developers and enterprise users, this may mean accessing stronger automation capabilities at a lower budget than ever before.