OpenAI cut prices on its newest AI models by half on September 22, 2026, rolling out GPT-6 Sol and GPT-6 Luna as direct replacements for the GPT-5.6 lineup. The move lands GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, while GPT-6 Luna comes in at just $0.10 input and $0.50 output. Both figures are half of what OpenAI charged for the equivalent GPT-5.6 tier just weeks earlier, according to the company’s own announcement, “Introducing GPT-6 Sol and Luna.”
The pricing update matters beyond the spreadsheet. Wccftech’s September 22 report framed it as OpenAI unleashing a new price war, arguing that GPT-6 Sol now lands below Claude Opus 5.5 and that GPT-6 Luna undercuts DeepSeek’s V4.1 Flash, a widely used open-weight model. If that framing holds up, it changes the calculus for any team that chose an open-weight model purely to save money on inference.
What OpenAI Announced on September 22
OpenAI’s blog post introduced GPT-6 Sol and GPT-6 Luna as the mid-tier and lightweight members of the GPT-6 family, sitting below the flagship GPT-6 Astra, which the company had already priced at $10 per million input tokens and $50 per million output tokens for its top-end release. Sol and Luna are built for the bulk of production traffic: chat assistants, coding copilots, and high-volume API calls where price per token decides whether a product is profitable.
VentureBeat’s coverage added a detail that matters for anyone budgeting compute spend: an OpenAI spokesperson told the outlet the new rates are permanent list prices, not a limited-time promotion. That distinction separates this cut from earlier short-term discounts OpenAI has run to win developer mindshare. TheNewStack and Unite.AI both confirmed the same headline numbers in their own write-ups the same day, and MarkTechPost published benchmark context alongside the pricing.
GPT-6 Sol and GPT-6 Luna Pricing, Broken Down
The headline numbers are straightforward, but the cached-token rate is where the real savings show up for high-traffic applications. OpenAI cut cached-input pricing to $0.20 per million tokens for Sol and $0.01 per million tokens for Luna, according to the September 22 announcement. For a chatbot that reuses the same system prompt thousands of times a day, that cache discount can matter more than the headline input price.
| Model | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| GPT-5.6 Sol (prior tier) | $4.00 | not disclosed | $20.00 |
| GPT-5.6 Luna (prior tier) | $0.20 | not disclosed | $1.20 |
| GPT-6 Astra (flagship) | $10.00 | not disclosed | $50.00 |
Every row in that table tells the same story: OpenAI is pushing its mainstream tiers down while keeping the flagship Astra model priced at a premium for teams that need the extra reasoning headroom. That’s a familiar playbook. Cloud providers have run the identical move for years, subsidizing the entry tier to keep developers inside the ecosystem while charging a premium for the top-end product.
How the New Prices Compare to GPT-5.6
OpenAI’s own comparison, cited in its September 22 post, puts both new models at 50% below their GPT-5.6 equivalents on a straight input-and-output basis. Sol drops from $4/$20 to $2/$10. Luna drops from $0.20/$1.20 to $0.10/$0.50. That’s not a rounding-error trim, it’s a full halving of the sticker price for models OpenAI says were trained using techniques similar to those behind GPT-6 Astra, carrying over gains in factuality, coding, and computer-use tasks into cheaper, faster models.
For a startup running a few hundred million tokens a month, that cut isn’t cosmetic. A team spending $40,000 a month on GPT-5.6 Sol calls would pay roughly $20,000 for the same volume on GPT-6 Sol, assuming similar token counts per request. Multiply that across every company running production LLM workloads and the aggregate savings, or lost revenue depending on which side of the transaction you sit, run into real money fast.
The Real Story: Undercutting Claude and DeepSeek
The pricing cut alone would be a routine story. What pushed it onto Wccftech’s front page, and what multiple other outlets picked up the same day, is the comparison to rival models. Wccftech reported that GPT-6 Sol at $2/$10 now sits below Claude Opus 5.5, which outlets including Cryptopolitan and TheNextWeb cite at $4 input and $20 output per million tokens, the same rate OpenAI itself was charging for GPT-5.6 Sol before this cut.
The Luna comparison is arguably sharper. Reports say DeepSeek’s V4.1 Flash, an open-weight model that developers can run on their own infrastructure, lists off-peak API pricing around $0.15 input and $0.60 output per million tokens. GPT-6 Luna’s $0.10/$0.50 rate undercuts that on a closed, hosted model. That detail is the crux of the “negates the rationale for open-weight models” framing: if a closed API from a well-funded lab beats an open-weight model on price and doesn’t require anyone to manage their own GPU fleet, the traditional argument for going open-weight, namely cost control, gets a lot weaker.
It’s worth being precise about what’s confirmed here. OpenAI’s own announcement discloses its pricing and the GPT-5.6 comparison directly. The Claude Opus 5.5 and DeepSeek V4.1 Flash comparisons come from third-party reporting, not from an official OpenAI statement naming those competitors. Treat the “undercuts Claude and DeepSeek” framing as journalism built on public pricing pages, not as an OpenAI marketing claim.
Why Cached Tokens Matter More Than the Headline Price
Most coverage of the launch led with the input and output numbers, but the cache discount is where OpenAI is making its bigger bet. Cutting cached-input pricing to $0.20 per million tokens on Sol and $0.01 on Luna rewards applications that reuse context: long system prompts, retrieval-augmented pipelines that resend the same document chunks, and multi-turn conversations that keep replaying earlier messages.
Here’s a simplified way to estimate the effect on a typical high-cache-hit workload:
# Rough monthly cost estimate for a chat app on GPT-6 Sol
# Assumes 80% of input tokens hit the prompt cache
total_input_tokens = 500_000_000 # 500M tokens/month
cache_hit_rate = 0.80
fresh_input = total_input_tokens * (1 - cache_hit_rate)
cached_input = total_input_tokens * cache_hit_rate
cost_fresh = (fresh_input / 1_000_000) * 2.00
cost_cached = (cached_input / 1_000_000) * 0.20
print(f"Fresh input cost: ${cost_fresh:,.2f}")
print(f"Cached input cost: ${cost_cached:,.2f}")
print(f"Total input cost: ${cost_fresh + cost_cached:,.2f}")
Run that math and a workload with an 80% cache hit rate pays roughly $260 for 500 million input tokens instead of $1,000 at the flat rate. That gap is why engineering teams building on GPT-6 should spend more time optimizing prompt structure for cache hits than debating whether Sol or Luna is the better base model.
The Hardware Economics Behind Cheaper Tokens
None of this pricing exists in a vacuum. Every token OpenAI serves runs on physical GPU capacity, and inference cost is ultimately a function of how many tokens a given cluster can push per second, how much of that capacity sits idle between requests, and how aggressively a provider can batch and cache repeated work. A 50% price cut on a model built with the same underlying techniques as GPT-6 Astra implies either meaningfully better GPU utilization per query, cheaper access to compute, or a deliberate decision to run these tiers at thinner margins to hold market share.
That last point lines up with a broader pattern this site has tracked all year: compute costs and memory pricing have been anything but stable. AI chip prices in China have climbed as an HBM shortage bites, which makes it notable that a major US lab is cutting API prices at the same time hardware inputs are getting more expensive elsewhere in the supply chain. The gap is bridged by scale: OpenAI can spread fixed infrastructure costs across a much larger request volume than a smaller open-weight deployment running on rented or owned GPUs, and that scale advantage is exactly what lets a closed provider undercut a model that’s technically free to download.
Competitive Landscape: GPT-6 Sol vs Claude Opus 5.5 vs DeepSeek V4.1 Flash
Laying the three models side by side, using OpenAI’s own disclosed pricing for GPT-6 and the figures cited by outlets including Wccftech, Cryptopolitan, and TheNextWeb for the competitors, shows how tight the gap has become at the top of the market.
| Model | Type | Input (per 1M tokens) | Output (per 1M tokens) | Source |
|---|---|---|---|---|
| GPT-6 Sol | Closed API | $2.00 | $10.00 | OpenAI, Sept. 22, 2026 |
| Claude Opus 5.5 | Closed API | ~$4.00 | ~$20.00 | Reported by Cryptopolitan, TheNextWeb |
| GPT-6 Luna | Closed API | $0.10 | $0.50 | OpenAI, Sept. 22, 2026 |
| DeepSeek V4.1 Flash | Open-weight | ~$0.15 | ~$0.60 | Reported by Wccftech and others |
The pattern holds at both ends of the market. At the premium tier, GPT-6 Sol is reported to be roughly half the price of Claude Opus 5.5. At the budget tier, GPT-6 Luna is reported to slightly undercut an open-weight model that developers would otherwise need to host themselves. For a team already running on Grok 4.7, which launched at its own aggressive $2/$6 pricing to chase OpenAI and Anthropic, the field just got more crowded at the low end.
A Brief History of AI Inference Price Wars
Price cuts on frontier models aren’t new, but the cadence has sped up. OpenAI has trimmed API pricing repeatedly since 2023, each time framing the move as passing efficiency gains to developers. The bigger shock came in early 2025, when a low-cost reasoning model out of China rattled the market and forced every major lab to reconsider what “premium” pricing could actually hold. Since then, the pattern has repeated with each new model generation: ship a flagship at a premium, then follow within weeks or months with a cheaper, faster sibling aimed at the volume market.
What’s different this round is the explicit competitive framing. Earlier price cuts were usually pitched as internal efficiency wins. This one, per Wccftech’s reporting, is being read by the press as a direct shot at both a closed rival (Anthropic) and the open-weight ecosystem (DeepSeek, and by extension Alibaba’s Qwen line and Moonshot’s models). That’s a shift from “we got more efficient” to “we can out-price anyone still trying to compete on cost.”
What This Means for Open-Weight Models
Open-weight models like DeepSeek’s line, or Alibaba’s Qwen3.8-Max, have built their appeal on three pillars: lower cost, the ability to self-host for data control, and freedom from vendor lock-in. This pricing move only really touches the first pillar. If GPT-6 Luna genuinely lands below hosted DeepSeek V4.1 Flash pricing, cost stops being the deciding factor for teams that don’t have a hard requirement to run inference on their own hardware.
That doesn’t kill the open-weight argument outright. Regulated industries, defense contractors, and companies in regions with strict data-residency rules still need models they can run entirely inside their own infrastructure, no matter what a hosted API charges. But for the much larger pool of developers who chose an open-weight model purely to save money on a hosted alternative, the math OpenAI just published removes a big part of the incentive.
- Cost advantage narrows or disappears for teams comparing hosted-vs-hosted pricing
- Self-hosting still wins on data control and compliance requirements
- Fine-tuning and customization remain easier with open weights
- Vendor lock-in risk stays unchanged regardless of price
Market Reaction and What Analysts Are Watching
The pricing news landed the same week AI infrastructure spending has been under scrutiny from investors watching for signs of a slowdown in the broader build-out. A price war among model providers cuts against that narrative in one specific way: if labs are racing each other down on price, that’s a sign of intense competition for developer traffic, not shrinking demand for AI compute overall. Analysts covering the space will likely watch whether Anthropic responds with its own Opus pricing adjustment, and whether DeepSeek or other open-weight providers push their next release to reclaim the cost advantage.
For the coding-tools market specifically, cheaper GPT-6 Sol and Luna pricing adds pressure on every AI coding assistant built on top of OpenAI’s API, since their own margins move in lockstep with whatever OpenAI charges upstream. That mirrors what happened when Grok 4.7’s steep discount squeezed coding-agent rivals earlier this year: a model-layer price cut doesn’t just help end users, it reshapes the margins of every product built on top of the API.
Risks and Unanswered Questions
A few things remain genuinely unclear. OpenAI’s post confirms its own pricing and the GPT-5.6 comparison, but the company hasn’t issued an official statement naming Claude Opus 5.5 or DeepSeek V4.1 Flash as pricing targets, that framing comes entirely from press analysis of public rate cards. Rate cards also change. Anthropic or DeepSeek could adjust their own pricing in response within days, which would undercut the “undercuts rivals” headline almost immediately.
There’s also the durability question. VentureBeat reported OpenAI called these permanent prices rather than a promotion, but “permanent” in API pricing has historically meant “until the next model generation ships.” Given how quickly OpenAI, Anthropic, and the open-weight labs have been iterating this year, permanent may realistically mean a matter of months rather than years.
What Comes Next: Predictions
Based on the pattern of moves across the industry this year, a few outcomes look likely in the weeks ahead.
- Anthropic responds within weeks. A direct undercut on Claude Opus 5.5 pricing is the kind of move that typically draws a countermove, whether through a price adjustment or a cheaper Claude tier aimed at the same volume segment.
- Open-weight labs lean harder on customization, not price. Expect DeepSeek, Alibaba, and Moonshot to shift marketing away from raw cost savings and toward fine-tuning flexibility and self-hosting control, since the price gap is narrowing.
- Cache-hit optimization becomes a bigger engineering focus. With cached tokens priced 90% below fresh input on Sol, expect more tooling and best-practice guides aimed specifically at maximizing prompt cache reuse.
- More price announcements land before year-end. Given the speed of moves from OpenAI, Anthropic, and Grok’s own aggressive pricing this year, another round of cuts before December looks likely rather than unusual.
- Coding-assistant margins keep compressing. Products built on top of these APIs will keep passing savings to users to stay competitive, squeezing their own margins further in the process.
Why This Matters for Developers Right Now
For any team actively choosing a model provider this quarter, the practical takeaway is simple: run the numbers on your actual workload rather than trusting headline pricing. A high-cache-hit chat application will see dramatically different economics than a batch-processing pipeline that rarely reuses context. Teams that assumed open-weight was the automatic budget choice should re-run their cost comparisons now that GPT-6 Luna is reportedly priced below at least one major open-weight alternative on a hosted basis.
It’s also worth remembering that pricing is only one input. Benchmark performance, latency, context window size, and tool-use reliability all factor into a real production decision, and none of those are captured by a per-token rate card alone.
Frequently Asked Questions
What are GPT-6 Sol and GPT-6 Luna?
They are two new models OpenAI introduced on September 22, 2026, sitting below the flagship GPT-6 Astra in OpenAI’s model lineup, aimed at production-scale chat, coding, and API workloads.
How much do GPT-6 Sol and GPT-6 Luna cost?
GPT-6 Sol costs $2 per million input tokens, $0.20 per million cached-input tokens, and $10 per million output tokens. GPT-6 Luna costs $0.10 input, $0.01 cached input, and $0.50 output per million tokens.
Are these prices cheaper than the previous GPT-5.6 models?
Yes. OpenAI says both new models are priced 50% below the equivalent GPT-5.6 Sol and GPT-5.6 Luna promotional rates.
Is GPT-6 Sol really cheaper than Claude Opus 5.5?
According to reporting from outlets including Wccftech, Cryptopolitan, and TheNextWeb, GPT-6 Sol’s $2/$10 pricing is roughly half of Claude Opus 5.5’s reported $4/$20 rate. OpenAI itself has not issued an official statement naming Claude Opus 5.5 as a direct comparison.
Does GPT-6 Luna undercut DeepSeek V4.1 Flash?
Reports cite DeepSeek V4.1 Flash’s off-peak API pricing around $0.15 input and $0.60 output per million tokens, which GPT-6 Luna’s $0.10/$0.50 rate would undercut. This comparison comes from press analysis of public rate cards, not an official statement from either company.
Is this pricing temporary or permanent?
An OpenAI spokesperson told VentureBeat the rates are permanent list prices rather than a limited-time promotion, though pricing across the industry has historically shifted again once a new model generation ships.
Does this pricing cut affect open-weight models like DeepSeek or Qwen?
It narrows the cost advantage that open-weight models have relied on when compared to a hosted API. It doesn’t remove other advantages of open weights, such as self-hosting for data control and easier fine-tuning.
What’s the difference between GPT-6 Sol, Luna, and Astra?
Astra is OpenAI’s flagship GPT-6 model, priced at $10 input and $50 output per million tokens for maximum reasoning capability. Sol and Luna are lower-cost, faster models built with similar training techniques, aimed at high-volume production traffic rather than top-end reasoning tasks.
