OpenAI cut the API prices of its new GPT-6 Sol and GPT-6 Luna models by roughly half on September 22, 2026, positioning both below rival flagship pricing on the same day Anthropic shipped Claude Opus 5.5. The timing was not a coincidence. GPT-6 Sol now costs less per million tokens than Claude Opus 5.5, and GPT-6 Luna undercuts DeepSeek’s V4.1 Flash on both input and output pricing, according to Techmeme’s aggregation of the launch coverage and OpenAI’s own announcement. The aggressive pricing did not stop there: by October 2026, OpenAI’s own API pricing page was already listing a follow-up GPT-6.1 Sol model with cached-input tokens billed at just $0.10 per million, a sign this round of cuts was the opening move in a longer campaign rather than a one-off promotion.
The move reopens a pricing fight that has defined the AI industry since DeepSeek’s earlier releases forced every major lab to defend its margins. This time, the story is not an open-weight upstart undercutting closed labs. It is OpenAI, the most valuable private AI company in the world, undercutting both a premium closed competitor (Anthropic) and a low-cost open-weight one (DeepSeek) in the same announcement. That combination is what makes this week’s pricing news different from the price cuts that came before it.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI Actually Changed With GPT-6 Sol and GPT-6 Luna
GPT-6 Sol and GPT-6 Luna launched on September 22, 2026, replacing the promotional pricing OpenAI had attached to the GPT-5.6 lineup. Sol dropped from $4 input and $20 output per million tokens to $2 input and $10 output, a straight 50% reduction on both sides of the ledger. Luna fell from $0.20 input and $1.20 output to $0.10 input and $0.50 output, a 50% cut on input and roughly a 58.3% cut on output, per figures reported by The Next Web and confirmed directly by OpenAI. The rollout itself was staggered rather than universal: Sol and Luna initially shipped into two ChatGPT workspaces, Work and Codex, in September 2026, while Luna separately reached Free-tier and Go desktop users, meaning the new pricing reached different slices of OpenAI’s user base on different tracks rather than all at once.
Both models carry a standard context window of 272,000 tokens, with some coverage indicating that prompts exceeding that ceiling may be billed at roughly 2x the input rate and 1.5x the output rate. That detail has circulated less consistently across outlets than the headline pricing, so it should be treated as a secondary claim rather than a confirmed policy. What is consistent across every report is the core price table itself, repeated nearly verbatim by 9to5Google, Techmeme, and SiliconANGLE.
GPT-6 Sol and GPT-6 Luna Pricing, Old vs New
| Model | Previous Input | New Input | Previous Output | New Output | Cache Read |
|---|---|---|---|---|---|
| GPT-6 Sol | $4.00 / M tokens | $2.00 / M tokens | $20.00 / M tokens | $10.00 / M tokens | $0.20 / M tokens |
| GPT-6 Luna | $0.20 / M tokens | $0.10 / M tokens | $1.20 / M tokens | $0.50 / M tokens | $0.01 / M tokens |
Cache pricing tells its own story. Sol’s cache write rate sits at $2.50 per million tokens against a $0.20 cache read, while Luna’s cache write is $0.125 against a $0.01 read, according to Techmeme’s breakdown of the launch. For developers running coding agents that repeatedly reuse long system prompts, cache reads at those levels change the real-world cost of a production deployment far more than the headline input price does. That dynamic only sharpened in the weeks after launch: as of October 2026, OpenAI’s official API pricing page listed the newer GPT-6.1 Sol variant’s cached-input rate at $0.10 per million tokens, just five percent of Sol’s $2.00 uncached input price, which underscores how much of the real savings in this pricing fight now live in cache behavior rather than in the headline input and output numbers. OpenAI framed the entire package as a byproduct of better caching and inference efficiency, with the savings passed to customers rather than absorbed as extra margin, a rationale 9to5Google’s coverage attributed directly to the company.
Claude Opus 5.5’s New Price Sits Right in the Crosshairs
Anthropic’s Claude Opus 5.5 launched at $4.00 per million input tokens and $20.00 per million output tokens, down from Claude Opus 5’s $5 and $25, a 20% cut. That is a meaningful reduction in isolation. It becomes a lot less impressive next to GPT-6 Sol’s $2/$10 rate, which lands at exactly half of Opus 5.5’s price on both input and output. Readers tracking the earlier Claude Opus 5.5 cost comparison against Fable 5.1 already know Anthropic has been positioning Opus as the premium-but-efficient option in its lineup. OpenAI’s Sol pricing directly challenges that positioning.
The gap widens further against Luna. At $0.10 input and $0.50 output, GPT-6 Luna is roughly 40 times cheaper than Claude Opus 5.5 on both metrics. That is not a fair fight in capability terms, since Luna is OpenAI’s lightweight, high-throughput tier rather than a reasoning flagship. But for the large share of production workloads that do not need frontier reasoning, cost differences of that magnitude are the entire decision. SiliconANGLE’s report described the sequence plainly: Anthropic released Opus 5.5, and OpenAI countered with two cheaper GPT-6 models on the same day.
DeepSeek V4.1 Flash’s Peak and Off-Peak Pricing, Explained
DeepSeek V4.1 Flash uses a variable pricing schedule rather than a flat rate. Off-peak, it charges $0.15 per million input tokens and $0.60 per million output tokens. During peak hours, those rates roughly double, to $0.30 input and $1.20 output. That structure, reported in detail by VentureBeat’s coverage of the launch, was designed to smooth demand across DeepSeek’s infrastructure by giving developers a financial incentive to shift batch workloads to quieter hours.
GPT-6 Luna beats DeepSeek V4.1 Flash on every one of those numbers. Luna’s flat $0.10 input rate undercuts DeepSeek’s best off-peak input price of $0.15. Luna’s $0.50 output rate beats DeepSeek’s off-peak output price of $0.60. And because Luna does not have a peak surcharge, the gap balloons during DeepSeek’s peak window, where Luna’s $0.10/$0.50 sits at roughly a third of DeepSeek’s $0.30/$1.20. Chinese outlet 36Kr reported that Luna’s pricing was widely read as a direct challenge to DeepSeek’s position as the default low-cost model for high-volume inference.
Side-by-Side: All Four Models on One Price Table
| Model | Input / M tokens | Output / M tokens | Positioning |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | OpenAI flagship, coding agents |
| GPT-6 Luna | $0.10 | $0.50 | OpenAI economy / high-throughput tier |
| Claude Opus 5.5 | $4.00 | $20.00 | Anthropic flagship reasoning model |
| DeepSeek V4.1 Flash (off-peak) | $0.15 | $0.60 | Open-weight economy tier |
| DeepSeek V4.1 Flash (peak) | $0.30 | $1.20 | Open-weight economy tier, peak surcharge |
Laid out this way, the pattern is hard to miss. OpenAI now has an entry in almost every price tier, and in two of those tiers it is now the cheapest option on the table, not just the cheapest closed-weight option. That is the detail driving most of the analysis around this launch, including comparisons to the broader model landscape covered in GPT-6 Astra’s earlier three-way comparison against Gemini 3.8 and DeepSeek V4.1 Flash.
Why OpenAI Says It Made the Cut
OpenAI’s stated explanation, per 9to5Google’s reporting, is straightforward: improved caching and more efficient inference lowered the actual cost of serving these models, and the company chose to pass those savings on rather than keep them as margin. That is a plausible, unremarkable explanation for a routine price adjustment. It does not, on its own, explain the timing.
The timing is the story. Anthropic and OpenAI shipped their respective flagship updates within a day of each other, and OpenAI’s pricing came out lower on the model most directly comparable to Opus 5.5. Whether that reflects a genuine cost breakthrough or a calculated response to Anthropic’s roadmap is not something either company has confirmed on the record. What is confirmed is that the two launches landed close enough together that outlets covering both, including SiliconANGLE and 9to5Google, treated them as a single competitive event rather than two unrelated announcements.
Does This Undercut the Case for Open-Weight Models?
Open-weight models built their market position on a simple pitch: better price-to-performance than closed labs, plus the freedom to self-host. DeepSeek V4.1 Flash’s off-peak pricing of $0.15/$0.60 was, until this week, one of the strongest arguments for that pitch among serious production users. GPT-6 Luna’s $0.10/$0.50 flat rate weakens that argument in a very specific way. If a developer’s primary reason to reach for an open-weight economy model was raw API cost, that reason has partly evaporated, at least for workloads that do not require self-hosting, data residency, or fine-tuning control that closed APIs cannot offer.
That does not eliminate the case for open-weight models. Self-hosting, licensing flexibility, and the ability to fine-tune or fully audit a model remain reasons that have nothing to do with per-token API pricing. But the argument that open-weight models are simply cheaper at the API level, which has circulated widely since DeepSeek’s earlier releases including the uncensored variants covered in our look at abliterated DeepSeek V4.1 Flash builds on Hugging Face, is now harder to defend with a straight face when a closed-weight competitor is cheaper on both input and output tokens.
Market and Industry Reaction So Far
Coverage of the dual launch has been consistent in framing, if not always in emphasis. The Next Web described OpenAI’s move as effectively halving GPT-6 pricing against the prior promotional rates. SiliconANGLE framed it as a direct counter to Anthropic’s Opus 5.5 release. VentureBeat’s analysis leaned toward the DeepSeek comparison, treating Luna’s flat-rate undercut of DeepSeek’s variable pricing as the more consequential story for high-volume developers. 36Kr’s coverage in the Chinese tech press emphasized the same angle, noting explicitly that Luna’s input price beat DeepSeek’s best available rate.
What none of the available reporting establishes is a confirmed public market reaction in the traditional sense. OpenAI and DeepSeek are both privately held, so there is no stock ticker to move. No verified public-market data on Anthropic’s valuation shifting in response to this specific announcement has surfaced in the coverage reviewed. Any claim tying this pricing news to a specific stock price move would be speculation dressed up as fact, and readers should treat any such claim elsewhere with skepticism until a named source confirms it.
The Benchmark Question Nobody Has Fully Answered Yet
Pricing is easy to verify. Performance is not, at least not yet. One report characterized GPT-6 Sol as surpassing the full GPT-5.6 series on unspecified measures, but the underlying benchmark, scoring methodology, and whether the claim originated from OpenAI or an independent evaluator were not disclosed in that coverage. No outlet reviewed for this article has published a complete, independently verified benchmark table placing GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, and DeepSeek V4.1 Flash on the same standardized tests under the same conditions.
That gap matters more than it might seem. Price comparisons only tell half the story if the cheaper model also does meaningfully worse work, forcing developers to spend more tokens on retries or longer prompts to hit the same output quality. Until a neutral third party runs Sol, Luna, Opus 5.5, and V4.1 Flash through the same coding, reasoning, and agentic benchmarks, any claim that one model “wins” on value is really only a claim about price, not about value. Readers comparing reasoning quality across current models may find the earlier breakdown in ChatGPT vs Claude vs Gemini vs Grok’s reasoning gap analysis useful context, though that piece predates this week’s Sol and Luna launch.
How the 2025-2026 AI Price War Got Here
This week’s cuts did not happen in a vacuum. Over the past year, every major lab has leaned on lower inference pricing as a competitive lever alongside, and sometimes instead of, raw benchmark gains. Caching became one of the most important tools in that fight, since repeated system prompts and long-running agent sessions can be served far more cheaply once a provider builds efficient cache infrastructure. Peak and off-peak pricing, of the kind DeepSeek uses for V4.1 Flash, emerged as a way to smooth infrastructure demand rather than simply cut prices across the board.
Model families have also become more segmented during this period, with labs increasingly selling a premium reasoning tier alongside a cheap, fast, high-throughput tier rather than a single one-size-fits-all model. GPT-6 Sol and Luna are a textbook example of that segmentation, mirroring the split OpenAI already ran with earlier releases. The pattern also shows up elsewhere in the market, including xAI’s Grok 4.7 launch at $2 input and $6 output, which arrived with its own claim to a competitive price-to-benchmark ratio against Claude.
What This Means for Developers Building on These APIs
For teams running coding agents or high-volume production pipelines, the practical takeaway is that per-token math changed meaningfully this week, and it is worth re-running cost projections rather than assuming last month’s pricing still holds. A workload that consumed a few million tokens a day on GPT-5.6-era pricing now costs roughly half as much on Sol, and dramatically less on Luna if the task does not require flagship-level reasoning.
Cache strategy now matters more than ever. With Sol’s cache read at $0.20 per million tokens against a $2.00 uncached input rate, and Luna’s cache read at a mere $0.01, structuring prompts to maximize cache hits is the difference between a marginal savings and a genuinely different cost profile for an agent-heavy application, and OpenAI’s own pricing page confirmed that advantage was still intact for the follow-up GPT-6.1 Sol model in October 2026, with cached input holding at $0.10 per million tokens. That matters especially for teams building on the Codex workspace, since Sol and Luna’s initial rollout targeted the Work and Codex ChatGPT workspaces specifically in September 2026, putting agentic coding tools squarely in the group that stands to benefit most from aggressive cache pricing. Anyone maintaining tooling that already targets OpenAI’s earlier models, including setups discussed in guides on OpenAI’s math-solving milestone from earlier this year, should expect to revisit rate limits, context handling, and fallback logic given the new context ceiling and its reported overage multipliers.
Predictions: Where This Pricing Fight Goes Next
- Anthropic will face pressure to either cut Claude Opus 5.5 pricing again or sharpen its performance claims enough to justify the current 2x premium over GPT-6 Sol.
- DeepSeek is likely to respond with either a flatter pricing structure or a new V4.x release that removes the peak surcharge entirely, since that surcharge is now a clear competitive liability against Luna.
- Expect more labs to formalize a two-tier structure, a premium reasoning model and a cheap high-throughput model, rather than selling a single general-purpose API tier, following the template OpenAI just reinforced with Sol and Luna.
- Independent benchmark organizations will likely be pushed to publish head-to-head evaluations of Sol, Luna, Opus 5.5, and V4.1 Flash specifically because the pricing gap has outpaced the available performance data.
- Cache-pricing transparency will become a bigger part of how labs market themselves, since the real cost difference between these models increasingly lives in cache rates rather than headline input and output prices.
These are analytical projections based on the pricing and competitive dynamics described above, not confirmed roadmap details from any of the companies involved.
The Bigger Picture: A Price War With No Clean Winner Yet
It is tempting to declare GPT-6 Luna the outright winner of this round simply because its price is the lowest number on the table. That would be premature. Price is the easiest variable to compare and the hardest one to make decisions on in isolation, because it says nothing about latency, tool-use reliability, hallucination rates, or how a model behaves inside a long-running agent loop under real production load. What this week’s launches do confirm is that the assumption baked into a lot of 2025-era AI strategy, that open-weight models would always win on cost while closed labs competed on capability, no longer holds cleanly. OpenAI just priced a closed model below an open-weight one on both input and output tokens, and it did so on the same day a direct rival cut its own flagship price by 20%.
The next data point worth watching is not another price cut. It is whether any of these four models pulls ahead on a properly controlled, third-party benchmark that accounts for the new pricing. Until that happens, this remains a story about spreadsheets, not about which model is actually better.
Frequently Asked Questions
How much does GPT-6 Sol cost per million tokens?
GPT-6 Sol costs $2.00 per million input tokens and $10.00 per million output tokens as of its September 22, 2026 launch, down from $4.00 and $20.00 under the prior GPT-5.6 promotional pricing. A follow-up variant, GPT-6.1 Sol, carried a confirmed cached-input rate of $0.10 per million tokens on OpenAI’s API pricing page as of October 2026.
How much does GPT-6 Luna cost per million tokens?
GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20 previously.
Is GPT-6 Sol cheaper than Claude Opus 5.5?
Yes. GPT-6 Sol’s $2/$10 pricing is exactly half of Claude Opus 5.5’s $4/$20 rate on both input and output tokens.
Is GPT-6 Luna cheaper than DeepSeek V4.1 Flash?
Yes, on both metrics. Luna’s flat $0.10/$0.50 rate beats DeepSeek V4.1 Flash’s best off-peak rate of $0.15/$0.60, and beats its peak rate of $0.30/$1.20 by an even wider margin.
Why did OpenAI cut prices on GPT-6 Sol and Luna?
OpenAI attributed the cuts to improved caching and more efficient inference, saying the resulting cost savings were being passed on to customers. The cuts also landed the same day Anthropic launched Claude Opus 5.5, which several outlets treated as a competitive response rather than a coincidence.
Do these price cuts mean open-weight models are no longer cheaper?
Not entirely. GPT-6 Luna is now cheaper than DeepSeek V4.1 Flash on a pure per-token API basis, which weakens the cost argument for choosing an open-weight model purely to save money via a hosted API. Open-weight models still offer advantages closed APIs cannot match, including self-hosting, full model access, and fine-tuning control.
What is GPT-6 Sol and Luna’s context window?
Both models reportedly ship with a standard 272,000-token context window. Some coverage indicates prompts beyond that limit may be billed at a higher multiplier, though that detail is less consistently confirmed than the base context size.
Is there an independent benchmark comparing all four models?
No complete, independently verified benchmark table comparing GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, and DeepSeek V4.1 Flash under identical test conditions has been published as of this writing.
Related Coverage
- KAT Walk C2 vs Virtuix Omni One: $2,196 VR Treadmill Gap [2026]
- Reclusa X Mini vs C3 Pro vs Halo65: $85 Keyboard Gap [2026]
- Acer Warns PC Prices Jump 20% in Memory Crisis [2026]
- Snapdragon 8 Elite Extreme Gen 6: 8K/60fps Video, 2 Chips [2026]
- Control Resonant PC Specs: RTX 4070 Ti for 60 FPS [2026]
