Two new OpenAI models landed on a Tuesday, and one rumored model didn’t. On September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna, expanding the GPT-6 family that previously consisted only of flagship GPT-6 Astra. Pricing for the new tiers starts at $0.10 per million input tokens. Then there’s the model that wasn’t there. As The New Stack put it plainly: “As of now, there is no GPT-6 Terra.”
TL;DR: OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026, bringing Astra-class improvements to cheaper tiers. API prices dropped 50% versus GPT-5.6, starting at $0.10 per million input tokens. Sol reportedly doubles accuracy at half the cost. The models roll out to ChatGPT Plus, Pro, Business, Enterprise, Edu, and Copilot — and GPT-6 Terra does not exist.
What Did OpenAI Actually Release on September 22, 2026?
OpenAI released two new models: GPT-6 Sol and GPT-6 Luna. Both join the previously released flagship GPT-6 Astra, and according to TechCrunch, OpenAI says the new models are “cut from the same cloth as Astra” — meaning the company positions them as carrying the flagship’s improvements into lower-cost tiers.
The release date was crowded. Anthropic released Claude Opus 5.5 the same day, and Simon Willison noted on his blog that OpenAI’s launch came roughly an hour after Anthropic’s announcement. A day earlier, xAI shipped Grok 4.7 and MiMo v2.6 arrived in Flash and Pro variants. Willison described the moment as “a new price war.”
The launch covered more than just the API. Coverage from 9to5Mac indicates OpenAI upgraded both ChatGPT and Codex with the two new models. That matters for developers who use Codex in their daily workflows — the cheaper tiers are not limited to raw API calls.
And the Terra-shaped hole in the lineup is real, not a rumor debunked. Sources report simply that no GPT-6 Terra exists in the current lineup. The Myclaw comparison of “GPT-6 Luna vs GPT-5.6 Terra” refers to the older GPT-5.6 Terra model, not a new GPT-6 release under that name.
How Much Cheaper Are GPT-6 Sol and Luna?
The headline change is price: API pricing for the new models is 50% lower than GPT-5.6. Per gHacks, the 50% cut is measured against GPT-5.6 promotional pricing, and per MarkTechPost, the new models start at $0.10 per million input tokens. For teams running high-volume workloads, that math compounds quickly.
The pricing drop comes alongside what OpenAI describes as improved intelligence, not a downgrade. The New Stack’s framing was blunt: OpenAI “cuts token prices in half” while releasing the models. ZDNET’s headline went further, reporting that GPT-6 Sol “doubles its accuracy rate — for half the cost.”
A few practical implications of this pricing structure:
- High-volume tasks shift toward Sol and Luna instead of Astra by default
- The $0.10 per million input tokens entry point lowers the barrier for experimentation
- Existing GPT-5.6 pipelines have a direct, cheaper migration target
- MarkTechPost reports the launch includes caching upgrades, which can further reduce effective costs
Whether the discount applies to output tokens at the same 50% rate is not stated in the available sources. The sources specify input token pricing and the overall 50% reduction versus GPT-5.6.
How Do Sol and Luna Differ From GPT-6 Astra?
GPT-6 Astra remains the flagship. Sol and Luna complement it, occupying faster and cheaper tiers in the lineup, per The New Stack and gHacks. The family now has three members — Astra, Sol, and Luna — and Microsoft’s Azure blog groups all three as options for “production AI agents, complex workflows, and high-volume tasks.”
The clearest stated performance claim concerns Sol: ZDNET reports that Sol doubles its accuracy rate at half the cost. Luna’s positioning leans harder on price. The Myclaw comparison states that “GPT-6 Luna’s strongest case is its much lower token price for focused work,” while GPT-5.6 Terra “earns a place when its results save enough correction time or failed attempts to justify the premium.”
So the tiering logic looks like this:
| Model | Position in lineup |
|---|---|
| GPT-6 Astra | Flagship, previously the only GPT-6 model |
| GPT-6 Sol | Faster/cheaper tier; reportedly 2x accuracy at half the cost |
| GPT-6 Luna | Lowest-price tier, from $0.10 per million input tokens |
| GPT-6 Terra | Does not exist |
What the sources don’t provide: benchmark tables comparing Sol and Luna against Astra, context window sizes, or latency figures. Those details aren’t stated in the available coverage, so treat any specific numbers you see elsewhere with caution.
Where Can You Use the New Models?
Availability is broad. Per Neowin, GPT-6 Sol and Luna are rolling out to ChatGPT Plus, Pro, Business, Enterprise, and Edu customers. Free and Go users get access too, but only to GPT-6 Luna, and only through the ChatGPT desktop app.
Developers have more surface area than just ChatGPT. The GitHub Changelog confirms both models are available in GitHub Copilot, joining GPT-6 Astra as selectable options. Microsoft’s Azure blog covers the full trio — Astra, Sol, and Luna — in Microsoft Foundry, positioned for production AI agents and high-volume work.
Here’s the rollout in one list:
- ChatGPT Plus, Pro, Business, Enterprise, and Edu — both models
- ChatGPT Free and Go — GPT-6 Luna via the desktop app
- Codex — both models, per 9to5Mac
- GitHub Copilot — both models as selectable options
- Microsoft Foundry — all three GPT-6 models for production agents
The enterprise angle deserves attention. Enterprise and Edu customers get both models, and Foundry’s positioning around “production AI agents, complex workflows, and high-volume tasks” suggests OpenAI and Microsoft expect the cheaper tiers to absorb serious workloads — not just chat experimentation.
Is There a GPT-6 Terra?
No — according to The New Stack, “as of now, there is no GPT-6 Terra.” Despite widespread use of the name in the rumor mill before the September 22, 2026 launch, OpenAI’s GPT-6 lineup currently consists of three models: Astra, Sol, and Luna. The Terra name does not appear in any official announcement.
So where did the idea come from? The topic of a “GPT-6 Terra” model circulated widely ahead of the release, and some publications referenced an existing “GPT-5.6 Terra” — which appears to be a real, earlier model in OpenAI’s catalog. That’s likely the root of the confusion. A GPT-5.6 Terra exists in the sources’ discussion of pricing and comparisons; a GPT-6 Terra does not.
Could Terra arrive later? The sources don’t state any roadmap for a future model under that name, so any speculation about a delayed GPT-6 Terra is just that — speculation. What is confirmed: the cheaper, faster tier of the GPT-6 family is covered by Luna, and Sol sits alongside it as the higher-accuracy option of the two new releases.
For now, buyers evaluating “Terra” in a pricing table should double-check whether they’re looking at GPT-5.6 Terra, the previous generation, not a GPT-6 product.
What Do the Benchmarks Say About Sol’s Accuracy?
The headline claim comes from ZDNET’s coverage: GPT-6 Sol “doubles its accuracy rate — for half the cost.” In other words, the new model delivers a 2x improvement in accuracy compared to its predecessor while the API price drops by 50%, a combination OpenAI is clearly positioning as the core selling point of the launch.
Accuracy is the metric that matters most for professional use. A model that is cheap but frequently wrong costs more in corrections than it saves in tokens. TechCrunch reports that OpenAI itself is boasting about “fewer mistakes” as a defining trait of the Sol and Luna releases, not just the lower price. Both models are described as being “cut from the same cloth” as GPT-6 Astra, the flagship of the family.
The sources don’t publish the exact benchmark names, scores, or the baseline model for the “doubled accuracy” comparison in the available excerpts, so treat the 2x figure as OpenAI’s claim rather than an independently verified result. MarkTechPost notes that benchmarks accompany the launch alongside the pricing details, and the reduced error rate is a recurring theme across every source covering the release.
Pairing that accuracy gain with a price from $0.10 per million input tokens (MarkTechPost) is the actual story: same-or-better quality, half the bill.
What Does This Mean for Developers and AI Agents?
Cheap, accurate models are the raw material of the agent economy, and this launch is explicitly aimed at that market. Microsoft’s Azure blog frames the trio — GPT-6 Astra, Sol, and Luna — as “scalable model options for production AI agents, complex workflows, and high-volume tasks” inside Microsoft Foundry. That’s a direct signal: these models are meant to run autonomous workflows at scale, where token costs multiply fast.
The economics changed significantly. With API pricing cut 50% versus GPT-5.6 promotional pricing (gHacks) and entry-level cost starting at $0.10 per million input tokens, high-volume agent loops that were expensive on the previous generation become dramatically cheaper. For an agent making thousands of calls per day, halving the per-token rate is the difference between a viable product and a money pit.
Distribution is equally important. GitHub’s changelog confirms both models are now available in GitHub Copilot, joining GPT-6 Astra, so developers get the new options inside their existing tooling without any integration work. MarkTechPost also mentions “caching upgrades” bundled with the pricing change, which further reduces effective costs for repetitive agent workloads.
The net effect: production agent deployments on OpenAI models get cheaper on day one, across GitHub, Azure, and the OpenAI API.
How Does the Release Fit OpenAI’s Broader GPT-6 Rollout?
Sol and Luna complete the middle and lower tiers of a three-model GPT-6 family. The New Stack describes them as models that “complement the flagship GPT-6 Astra,” meaning the rollout follows a familiar pattern: flagship first, then cheaper, faster variants extended across the product line. 9to5Mac frames it the same way — OpenAI is “upgrading ChatGPT and Codex with two more GPT-6 models.”
The timing was aggressive. Simon Willison notes that Anthropic released Claude Opus 5.5, and roughly an hour later OpenAI released GPT-6 Sol and Luna — on the same day that Grok 4.7 and MiMo v2.6 Flash/Pro also shipped. Willison’s framing of the moment as “a new price war” captures the competitive context: multiple frontier labs cutting prices within a single 24-hour window.
Availability is broad from launch day. Neowin reports the models are rolling out to ChatGPT Plus, Pro, Business, Enterprise, and Edu customers, with free and Go users getting access to GPT-6 Luna through the ChatGPT desktop app. Enterprise and education plans receive the models alongside standard business tiers.
In short, this wasn’t a quiet addendum to Astra — it was a coordinated, multi-platform expansion of the GPT-6 line across ChatGPT, Codex, GitHub Copilot, and Microsoft Foundry, timed deliberately against rival launches.
Should You Switch From GPT-5.6 Terra to GPT-6 Sol or Luna?
For most workloads, the math favors switching. The new models cost 50% less than GPT-5.6 promotional pricing (gHacks) while Sol reportedly doubles its accuracy rate. Paying half as much for twice the accuracy is a rare combination, and it’s hard to construct a cost-based argument for staying on the older generation.
That said, one source offers a useful counterpoint. A Myclaw comparison of GPT-6 Luna and GPT-5.6 Terra argues that GPT-6 Luna’s “strongest case is its much lower token price for focused work,” while GPT-5.6 Terra “earns a place when its results save enough correction time or failed attempts to justify the premium.” In other words, if Terra’s specific behavior on your tasks genuinely reduces retries and manual fixes, its higher price can still pay for itself.
A practical approach, then:
- Run your representative workloads on both GPT-6 Sol and GPT-6 Luna before migrating anything
- Compare not just output quality, but the number of failed attempts and corrections each model requires
- Check whether your use case involves focused, bounded tasks (Luna’s sweet spot per the Myclaw analysis)
- Factor in caching upgrades mentioned by MarkTechPost if your workload has repetitive prompts
- Note that Enterprise and education customers get the new models per Neowin’s rollout details
- Verify any task where GPT-5.6 Terra currently saves correction time — that’s the only real argument for the premium
For high-volume agent work, the 50% price cut alone likely settles it. For specialized tasks, measure before you migrate.
Frequently Asked Questions
What are GPT-6 Sol and GPT-6 Luna?
They are two new OpenAI models launched on September 22, 2026, that complement the flagship GPT-6 Astra in the GPT-6 family. Both are described as cut from the same cloth as Astra, with ZDNET reporting that GPT-6 Sol doubles its accuracy rate while costing half as much as the previous generation.
How much do GPT-6 Sol and Luna cost?
API pricing is 50% lower than GPT-5.6 promotional pricing, according to gHacks, with MarkTechPost reporting a starting point of $0.10 per million input tokens alongside caching upgrades. Simon Willison described the simultaneous launches by OpenAI, Anthropic, and others as a “new price war.”
Is GPT-6 Terra a real model?
No. The New Stack states explicitly: “As of now, there is no GPT-6 Terra.” A GPT-5.6 Terra from the previous generation does appear in sources — Myclaw published a GPT-6 Luna versus GPT-5.6 Terra comparison — which is likely the source of the GPT-6 Terra confusion.
Can free ChatGPT users access GPT-6 Sol and Luna?
Partially. Per Neowin, paid tiers — Plus, Pro, Business, Enterprise, and Edu — get both models, while free and Go users will get access to GPT-6 Luna specifically through the ChatGPT desktop app. Free users do not get Sol according to the available sources.
Summary
OpenAI’s September 22, 2026 launch of GPT-6 Sol and Luna reshapes the price-performance curve for its model lineup. The key takeaways:
- There is no GPT-6 Terra. The GPT-6 family is Astra, Sol, and Luna; The New Stack confirms Terra doesn’t exist in the current generation, though GPT-5.6 Terra from the prior generation is real.
- Prices dropped 50%. API pricing is half of GPT-5.6 promotional pricing, starting from $0.10 per million input tokens, with caching upgrades on top.
- Accuracy doubled. ZDNET reports GPT-6 Sol doubles its accuracy rate at half the cost — the core claim of the launch.
- Availability is broad. ChatGPT Plus, Pro, Business, Enterprise, and Edu get both models; free and Go users get Luna via the desktop app; both models are live in GitHub Copilot and Microsoft Foundry.
- The price war is on. Sol and Luna landed within an hour of Claude Opus 5.5, the same day as Grok 4.7 and MiMo v2.6.
If you’re building on OpenAI models, the switch decision is mostly arithmetic: benchmark your workloads on Sol and Luna, compare correction costs against GPT-5.6 Terra, and let the 50% price cut do the arguing. Have you tested the new tiers yet? The comments are open.