GPT-6 Sol is the name at the centre of OpenAI’s latest rumour cycle, and it arrived quickly. Less than three days after OpenAI launched GPT-6 Astra on 3 September 2026, posts on X began describing a second GPT-6 model in internal testing. It was said to be faster than Astra and possibly easier on usage limits. By 15 September, chief executive Sam Altman was promising a big launch “this week” and a string of launches for DevDay, while an OpenAI developer forum thread asked outright whether Sol, Terra and Luna versions of GPT-6 were coming.
None of that is an announcement. OpenAI has published no GPT-6 Sol page, no model ID and no price. What it has published is a naming system that makes a GPT-6 Sol very likely. When it introduced GPT-5.6 in June, OpenAI said Sol, Terra and Luna were “durable capability tiers that can advance on their own cadence”. Our complete guide to GPT-5.6 Sol covers that generation in depth.
This article lines up every GPT-6 Sol claim against the evidence. It dates each leak to the minute, checks the “6x faster” figure against its own arithmetic, and maps where smaller GPT-6 models would sit on OpenAI’s price and usage ladders. It closes with what teams building on OpenAI should do before DevDay on 29 September.
Table of contents
- What the GPT-6 Sol Reports Actually Say
- Why OpenAI’s Own Naming Makes GPT-6 Sol Likely
- What Altman and Sottiaux Actually Teased
- GPT-6 Sol Speed Claims Put to the Arithmetic
- The Contradictions Inside the GPT-6 Sol Leaks
- Where GPT-6 Sol Would Sit on OpenAI’s Price Ladder
- GPT-6 Sol and the Usage-Limit Problem
- How Fast OpenAI Ships Smaller Siblings
- What GPT-6 Sol Would Mean for Teams Building on OpenAI
- The Market Pressure Behind a Cheaper GPT-6 Tier
- Frequently Asked Questions About GPT-6 Sol
- References
What the GPT-6 Sol Reports Actually Say
The GPT-6 Sol story did not start with a news report. It started with a handful of posts on X from accounts that test models and share outputs, then spread through Chinese tech media, rumour blogs and the OpenAI developer forum. Every claim below traces back to those posts. OpenAI itself has said nothing about GPT-6 Sol on the record.
The first GPT-6 Sol post: one SVG and a stopwatch
The earliest post we found came from the account @lyraxana at 23:44 UTC on Saturday 5 September. It was labelled “[GPT-6 Sol] BMW M4 Competition SVG” and described the run as “~28k tokens; ~3 minutes; effort: max; zero-shot”. By 16 September it had 104,170 views. The post named a model and showed an output, but it did not explain how the account had reached an unreleased model.
Lentils: “noticeably worse” but “a monster model”
About nine hours later, at 08:37 UTC on 6 September, the account @Lentils80 quoted that post and made the claim everyone else repeated: “Now that GPT-6 Astra is finally out, OpenAI is now internally testing GPT-6 Sol.” The same post was candid about quality. Outputs were “noticeably worse than Astra in general, from testing”, it said. “However it’s fast and seems to also be a monster model.”
It showed a pixel-art sandbox world made “in zero-shot, 15 minutes and 60k total tokens on Max effort”. The post drew 273,247 views, making it the single most influential GPT-6 Sol claim of the first week. Chinese site QbitAI later named the sandbox “The Realm of Aurellune” and listed a town, farmland, a river, a castle and a minimap among its parts.
Pankaj Kumar and the 29 September date
An hour later, at 09:35 UTC, developer Pankaj Kumar posted a three-point summary headed “GPT-6 Sol Leaks”. He called the model “fast and quite good, though not at the level of Astra”. He said “It could launch at OpenAI DevDay on September 29”. He framed GPT-6 Sol as “the middle ground between performance and limit especially for Plus users, with much better usage limits than Astra”. The DevDay date is real. The link between GPT-6 Sol and that date is his speculation.
The second wave: “beats GPT-6 Astra”
A week later the tone changed. On 13 September at 14:09 UTC the account @Mr_Salio posted “GPT-6 Sol Second Output Got LEAKED”. It claimed the model “beats GPT-6 Astra in coding and frontend generation and it will be a lot cheaper”. That post reached 390,878 views, the most of any GPT-6 Sol leak. A “Third Output” post followed at 14:18 UTC on 14 September, predicting that GPT-6 Sol “Will be a massive upgrade over Astra”.
| Posted (UTC) | Account | Claim | Views (16 Sep) |
|---|---|---|---|
| 5 Sep, 23:44 | @lyraxana | BMW M4 SVG, ~28k tokens, ~3 minutes, max effort | 104,170 |
| 6 Sep, 08:37 | @Lentils80 | Internal testing; “noticeably worse than Astra”; fast | 273,247 |
| 6 Sep, 09:35 | @pankajkumar_dev | Could launch at DevDay on 29 September; better limits | 153,931 |
| 13 Sep, 14:09 | @Mr_Salio | “beats GPT-6 Astra in coding”; “a lot cheaper” | 390,878 |
| 14 Sep, 14:18 | @Mr_Salio | “massive upgrade over Astra” | 179,919 |
| 14 Sep, 19:42 | @0x0SojalSec | “a cinematic 3D treehouse sequence” | 2,526 |
That is six posts from five accounts, with 1,104,671 combined views. None of them links to an OpenAI source, a model card or an API response.
Why OpenAI's Own Naming Makes GPT-6 Sol Likely
The leaks would be easy to dismiss if the name were invented. It is not. OpenAI created the structure that GPT-6 Sol would slot into, and it did so in writing, twice.
“Durable capability tiers”
When OpenAI previewed GPT-5.6 on 26 June 2026, it explained a new naming system: “the number identifies a model’s generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence.” The general-availability post on 9 July repeated the idea almost word for word. Under that rule GPT-6 is the generation. Astra is a new name above the old tiers, which leaves Sol, Terra and Luna free to arrive under the GPT-6 number whenever each is ready.
OpenAI already sells smaller models on the same logic
The 9 July post also made the business case for smaller tiers. “That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable,” OpenAI wrote. It added that GPT-5.6 Terra and GPT-5.6 Luna “outperform Fable 5 at around one-sixteenth the cost”. A GPT-6 generation with only one very expensive model would sit awkwardly against that pitch. That is why forum users and leak accounts alike expect a GPT-6 Sol, and possibly smaller siblings below it.
The gap on OpenAI’s model menu
OpenAI’s API documentation shows the gap plainly. Its “Choosing a model” advice reads: “use GPT-6 Astra, our flagship model for complex reasoning and coding. Choose GPT-5.6 Terra to balance intelligence and cost, or GPT-5.6 Luna for cost-sensitive, high-volume workloads.” Every tier below Astra still carries the 5.6 number. The same page still describes GPT-5.6 Sol as a “Flagship model for complex professional work”. A GPT-6 Sol would close that mismatch.
| Tier | GPT-5.6 generation | GPT-6 generation | Status of the GPT-6 slot |
|---|---|---|---|
| Above Sol | None | GPT-6 Astra (gpt-6-astra) | Launched 3 September 2026 |
| Sol | GPT-5.6 Sol (gpt-5.6-sol) | GPT-6 Sol | Reported in internal testing; unconfirmed |
| Terra | GPT-5.6 Terra (gpt-5.6-terra) | None announced | Asked about on the forum; no leak |
| Luna | GPT-5.6 Luna (gpt-5.6-luna) | None announced | Asked about on the forum; no leak |
What the developer forum asked
On 11 September a user opened a thread titled “GPT-6 Sol, Terra, and Luna?” on the OpenAI Developer Community. They asked whether those versions of GPT-6 would follow Astra. One reply claimed there was “some speculation and proof that GPT-6-Sol will be coming up as it appeared in Arena and API”, without links. Another pointed to precedent: “it was 12 days after the launch of GPT 5.4 we got mini & nano. We are only at 9 days”. The thread had 758 views when we checked, and no reply from OpenAI staff.
What Altman and Sottiaux Actually Teased
The “teases” in this story come from two OpenAI executives, and neither of them named GPT-6 Sol. Their posts matter because they set expectations for the next fortnight, and because at least one outlet read more into them than they say.
Thibault Sottiaux: “a level of ships”
At 06:19 UTC on 15 September, OpenAI’s Thibault Sottiaux posted: “This week will also be a level of ships that you could have expected for DevDay 2025. Crazy”. It had 2.99 million views by 16 September. Sottiaux is also the executive who announced the pause on new $200 Pro subscriptions on 10 September, so Astra demand is the backdrop to anything that ships this week.
Sam Altman: “big 🚢 this week”
Eight and a half hours later, at 14:46 UTC, Altman quoted that post with his own: “big 🚢 this week and then for devday 🚢🚢🚢🚢🚢🚢”. It drew 1,145,994 views and 843 replies. One big ship this week and six ship emojis for DevDay make up the whole tease. There is no model name, no product category and no date beyond “this week” and “devday”.
What the posts do not say
Wccftech read Altman’s post as teasing “a major model release this week, with next week expected to see multiple launches”. Two parts of that go beyond the text. Altman did not say the big ship was a model. DevDay is also not next week: OpenAI’s event page lists it for Tuesday 29 September at Fort Mason in San Francisco, 14 days after the post. The honest reading is that something big lands this week, and several launches follow at DevDay.
Two official posts drew far more attention than the leaks: Altman’s alone out-viewed all six GPT-6 Sol leak posts combined, by 41,323 views.
GPT-6 Sol Speed Claims Put to the Arithmetic
The most repeated number in the GPT-6 Sol story is “6x faster than Astra”. It comes from QbitAI, which published the comparison on 7 September under a headline saying Sol’s internal test results showed a sixfold speed-up. The figure is real arithmetic, but it rests on a single task.
Where “6x” comes from
QbitAI described one user giving the same prompt, an SVG drawing of a BMW M4 Competition, to four models. By its account, GPT-6 Sol at Max effort produced about 28,000 tokens in about three minutes. GPT-6 Astra at Max produced about 25,000 tokens in about 19 minutes. Gemini 3.1 DeepThink took about 29 minutes and Gemini 3.8 Flash about 42 seconds. Nineteen minutes divided by three gives 6.3x, which QbitAI rounded to six.
On wall-clock time alone, GPT-6 Sol sat between Google’s fastest and slowest entries in the same test.
Tokens per second tells a sharper story
Wall-clock time hides output size, so divide tokens by seconds. GPT-6 Sol’s 28,000 tokens in 180 seconds is about 156 tokens per second. GPT-6 Astra’s 25,000 tokens in 1,140 seconds is about 22. On that basis GPT-6 Sol is roughly 7.1 times faster per token, not six. Gemini 3.8 Flash, at 19,000 tokens in 42 seconds, works out at about 452 tokens per second, nearly three times the GPT-6 Sol rate.
| Run | Tokens | Time | Tokens per second | Source |
|---|---|---|---|---|
| GPT-6 Sol, SVG | ~28,000 | ~3 min | ~156 | @lyraxana; QbitAI |
| GPT-6 Astra, SVG | ~25,000 | ~19 min | ~22 | QbitAI |
| Gemini 3.8 Flash, SVG | ~19,000 | ~42 s | ~452 | QbitAI |
| GPT-6 Sol, sandbox | ~60,000 total | ~15 min | ~67 | @Lentils80 |
The same model, 2.3x apart
The Lentils sandbox run gives a second GPT-6 Sol data point: 60,000 total tokens in 15 minutes, or about 67 tokens per second. That is 2.3 times slower than the SVG run. “Total tokens” may include input and reasoning that the SVG post did not count, and the tasks differ. Either way, two GPT-6 Sol posts disagree on throughput by more than a factor of two. That is exactly why one zero-shot run cannot settle a speed ranking.
What a fair speed test would need
A credible comparison would fix the prompt, the effort level, the harness and the hardware, then repeat each run and report the spread. None of the GPT-6 Sol posts meets that bar. OpenAI’s own Astra launch post shows how much the setup matters. In latency simulations on OSWorld 2.0, it said Astra reached “higher computer-use performance in about 47% less time per task than GPT‑5.6 Sol”: roughly 40 minutes a task against 75.
The Contradictions Inside the GPT-6 Sol Leaks
Read side by side, the GPT-6 Sol posts do not describe one model. They describe a model that is worse than Astra, then better, depending on who is posting and when.
“Noticeably worse” against “beats Astra”
The earliest detailed posts, both from 6 September, agree that GPT-6 Sol trails Astra on quality. Lentils called it “noticeably worse”, and Pankaj Kumar said it was “not at the level of Astra”. The 13 September post claims the reverse, that GPT-6 Sol “beats GPT-6 Astra in coding and frontend generation”. Both cannot describe the same checkpoint on the same tasks. A later checkpoint could explain the change, but no post says so.
One video, two “leaks”
The clearest problem is a duplicated clip. The “Third Output LEAKED” post from @Mr_Salio at 14:18 UTC on 14 September and a “just leaked” post from @0x0SojalSec at 19:42 UTC the same day carry the identical video file. The second account described it as “a cinematic 3D treehouse sequence” and “a meaningful step up from Astra”. It did so five hours and 24 minutes after the first account had posted the same clip. A reshared clip is not a second GPT-6 Sol leak.
Claims set against evidence
| Claim | Who says it | What would confirm it | Status on 16 Sep |
|---|---|---|---|
| GPT-6 Sol is in internal testing | @Lentils80, @pankajkumar_dev | An OpenAI post, model card or API listing | Unconfirmed |
| About 6x faster than Astra | QbitAI, from one SVG run | Repeated runs on a fixed harness | Single data point |
| Weaker than Astra | @Lentils80, @pankajkumar_dev | Published benchmarks | Contradicted by later posts |
| Beats Astra at coding | @Mr_Salio | Published benchmarks | Contradicted by earlier posts |
| Launches at DevDay on 29 September | @pankajkumar_dev | An OpenAI announcement | Speculation; DevDay date confirmed |
| Seen in Arena and the API | OpenAI forum reply | A dated screenshot or model ID | No evidence shared |
| Better usage limits than Astra | @pankajkumar_dev, TheWinCentral | A Codex or ChatGPT limits table | Unconfirmed |
How fast rumour runs ahead of release
The same QbitAI article guessed that GPT-6 Terra, GPT-6 Luna and “GPT-Image 2.5” might appear together at DevDay. ChatGPT Images 2.5 actually launched on 8 September, the day after QbitAI published and three weeks before DevDay. TheWinCentral’s 7 September write-up of the GPT-6 Sol leak was more careful than most: “None of these details have been officially confirmed by OpenAI.”
Where GPT-6 Sol Would Sit on OpenAI's Price Ladder
If GPT-6 Sol is real, the most useful question for buyers is price. Here the evidence is solid, because OpenAI’s current API price sheet defines the gap a new tier would fill.
Today’s price ladder
| Model | Input per 1M | Cached input | Output per 1M | Note |
|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | Standard rate, short context |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 | Promotional; launched at $5 and $30 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 | Cut 20% on 30 July 2026 |
| GPT-5.6 Luna | $0.20 | $0.02 | $1.20 | Cut 80% on 30 July 2026 |
| GPT-6 Sol | Not published | Not published | Not published | Unconfirmed model |
The multiples that matter
GPT-6 Astra’s input price is 2.5 times the GPT-5.6 Sol promotional rate and twice its launch rate.
Astra is also five times Terra and 50 times Luna on input. On output, the gap to Sol is 2.5x at today’s promotional price. That spread is the space a GPT-6 Sol would occupy. A buyer who wants GPT-6-generation behaviour today has exactly one option, and it costs $50 per million output tokens.
Batch, Flex and Fast mode
The same sheet halves every rate for Batch and Flex processing, so Astra drops to $5 input and $25 output. Fast mode doubles the rate, taking Astra to $20 and $100. Long prompts cost more again: Astra’s long-context rate is $20 input and $75 output. Every model on the sheet follows the same multipliers, so a GPT-6 Sol would most likely inherit them too. That is a pattern on the page, not a forecast.
Why the promotional Sol price matters
On 21 August, OpenAI cut GPT-5.6 Sol’s API and credit pricing by more than 20% for three months. Its pricing page says the promotional rate is available at least through 21 November 2026. That date matters for anyone comparing a future GPT-6 Sol with the current one, because a late-September launch would be judged against a discounted incumbent. If the promotion lapses, GPT-5.6 Sol returns to $5 input and $30 output, and Astra’s premium narrows to 2x on input and 1.67x on output.
GPT-6 Sol and the Usage-Limit Problem
Pankaj Kumar’s most practical claim was that GPT-6 Sol would come “with much better usage limits than Astra”. Whether or not that proves true, OpenAI’s own Codex documentation shows why subscribers want it.
Messages per five hours by plan
| Model in Codex | Plus | Pro 5x | Pro 20x |
|---|---|---|---|
| GPT-6 Astra | 5–45 | 25–225 | 100–900 |
| GPT-5.6 Sol | 10–100 | 50–500 | 200–2,000 |
| GPT-5.6 Terra | 25–200 | 125–1,000 | 500–4,000 |
| GPT-5.6 Luna | 250–2,000 | 1,250–10,000 | 5,000–40,000 |
| GPT-5.4 mini | 60–350 | 300–1,750 | 1,200–7,000 |
A Plus subscriber’s upper bound more than doubles when moving from Astra to GPT-5.6 Sol, and even the older GPT-5.4 mini allows nearly eight times Astra’s ceiling.
GPT-5.6 Luna is left off the chart because its ceiling of 2,000 would flatten every other bar.
Credits tell the same story
For pay-as-you-go usage, the same page prices GPT-6 Astra at 250 credits per million input tokens and 1,250 per million output tokens. GPT-5.6 Sol costs 100 and 500. That is the same 2.5x ratio as the API’s promotional prices, so the arithmetic is consistent across products. An Astra-class answer costs about two and a half Sol-class answers.
Why Plus users care most
A Plus subscriber gets 5 to 45 Astra messages per five hours in Codex, against 10 to 100 for GPT-5.6 Sol. Demand for Astra was high enough that OpenAI paused new $200 Pro subscriptions on 10 September. A GPT-6 Sol with limits close to today’s Sol would let Plus users stay on the new generation for everyday work, instead of rationing Astra for the hardest jobs.
How Fast OpenAI Ships Smaller Siblings
The forum reply about GPT-5.4 raised a fair question: how long does OpenAI usually take to follow a flagship with smaller models? Its own launch posts give a clear, if uneven, record.
| Flagship | Launched | Smaller models | Launched | Gap |
|---|---|---|---|---|
| GPT-4o | 13 May 2024 | GPT-4o mini | 18 July 2024 | 66 days |
| GPT-5 | 7 August 2025 | gpt-5-mini and gpt-5-nano | 7 August 2025 | Same day |
| GPT-5.4 | 5 March 2026 | GPT-5.4 mini and nano | 17 March 2026 | 12 days |
| GPT-5.6 Sol | 26 June 2026 (preview) | GPT-5.6 Terra and Luna | 26 June 2026 (preview) | Same day |
| GPT-6 Astra | 3 September 2026 | None announced | Not yet | 13 days and counting |
A DevDay launch would put the GPT-6 gap between the GPT-5.4 and GPT-4o precedents.
The 12-day precedent
GPT-5.4 launched on 5 March 2026, and GPT-5.4 mini and nano followed on 17 March: exactly the 12 days the forum reply cited. OpenAI called them “our most capable small models yet”. It also explained the thinking behind them: “developers can compose systems where larger models decide what to do and smaller models execute quickly at scale.” By that clock, a small GPT-6 model would have been due on 15 September. It did not arrive.
The DevDay clock
If GPT-6 Sol or smaller GPT-6 models launch at DevDay, the gap from Astra would be 26 days. That is longer than GPT-5.4’s 12 days and far shorter than GPT-4o mini’s 66. The event page says the opening keynote starts at 10:00 a.m. Pacific Time on 29 September, “featuring Sam Altman”. It also lists follow-up DevDay Exchanges in cities including Bengaluru, Tokyo, London and São Paulo.
Why Astra broke the pattern
The GPT-5 and GPT-5.6 families launched their tiers together. Astra did not, and its launch post never mentions Sol, Terra or Luna. One likely reason is caution: Astra is the first OpenAI model rated at the Critical level for cybersecurity under the company’s Preparedness Framework, and enterprise access was switched off by default at launch. That link is our inference, not an OpenAI statement, but it fits the staged rollout OpenAI described.
What GPT-6 Sol Would Mean for Teams Building on OpenAI
A rumoured model is not something to build on, but it is something to plan around. Three practical steps follow from the evidence above, and none of them depends on GPT-6 Sol being real.
Route by tier, not by model name
OpenAI’s naming scheme lets tiers “advance on their own cadence”. A router that maps tasks to a tier, such as flagship, balanced or low-cost, will survive a GPT-6 Sol launch far better than hard-coded model names. Keep Astra for the hardest end-to-end work, where OpenAI’s documentation positions it. Send volume work to the cheapest tier that passes your own evaluations.
Pin model IDs and watch aliases
The models page lists GPT-5.6 Sol under the ID gpt-5.6-sol with the alias gpt-5.6. Aliases can move when a vendor updates a tier, so pin the full ID in production and test any new model under its own ID. If a GPT-6 Sol identifier appears in the API, run it against your evaluation set before switching any traffic.
Budget for three prices, not one
Plan cost scenarios at today’s promotional GPT-5.6 Sol price, at its $5 and $30 launch price, and at Astra’s $10 and $50. A GPT-6 Sol that lands anywhere in that band will then be easy to evaluate. A promotion that ends on or after 21 November will not surprise your finance team either.
| Action | Why | When |
|---|---|---|
| Map workloads to tiers | OpenAI says tier names persist across generations | Before DevDay |
| Pin full model IDs | Aliases such as gpt-5.6 can move | Now |
| Keep an evaluation set ready | Leaks disagree on GPT-6 Sol quality | Before any switch |
| Model three price points | The Sol promotion runs at least to 21 November | This quarter |
| Track Codex limits | Astra allows 5 to 45 Plus messages per 5 hours | Weekly |
For teams already running agents, OpenAI’s GPT-Live-1 API and Agents API public beta is the other launch to watch, since cheaper tiers change the cost of long agent loops.
The Market Pressure Behind a Cheaper GPT-6 Tier
OpenAI has commercial reasons to fill out the GPT-6 family quickly. Two data points posted on 15 September suggest Astra is already moving spend, and that much of it comes from OpenAI’s own older tier.
Ramp: Astra takes 13% of enterprise spend
Ramp’s Ara Kharazian wrote that “As of this week, Astra takes 13% of enterprise AI spend vs. Fable (8%) per Ramp data.” He added that “OpenAI’s growth is primarily coming from shifts from Sol and some Anthropic models, net-new usage too.” If GPT-5.6 Sol customers are trading up to a model that costs 2.5 times as much per token, a GPT-6 Sol would give them a same-generation step back down.
OpenRouter: a 2.5-year pattern flips
Writer Tae Kim relayed a claim that OpenRouter users spent more on OpenAI’s models than on Anthropic’s in the week before 15 September. According to the post, that had not happened for more than two and a half years. It is a single, second-hand data point from one model router rather than a market survey, but it points the same way as the Ramp figures.
Shipping while arguing for pacing
The teasers land in an awkward week. OpenAI chief scientist Jakub Pachocki published an essay titled “An Alien Mind” on 6 September, and OpenAI has since confirmed AI safety coordination with Anthropic and Google. Promising six ships for DevDay while the industry debates how fast frontier AI models should advance is not a contradiction on its own. It will, however, shape how any GPT-6 Sol launch is received.
Frequently Asked Questions About GPT-6 Sol
Has OpenAI confirmed GPT-6 Sol?
No. As of 16 September 2026, OpenAI has published no GPT-6 Sol announcement, model card, API ID or price. The name comes from posts on X starting on 5 September, and from OpenAI’s own rule that Sol, Terra and Luna are tiers that persist across generations.
Will GPT-6 Sol launch at DevDay on 29 September?
Possibly, but that is speculation. OpenAI has confirmed DevDay for 29 September in San Francisco, and Sam Altman has promised several launches for it. No OpenAI source links GPT-6 Sol to that date.
Is GPT-6 Sol faster than GPT-6 Astra?
One shared run suggests so. An SVG task took about 3 minutes on GPT-6 Sol against about 19 on Astra, roughly 6.3x faster on time and 7.1x on tokens per second. A single zero-shot run is not a benchmark, and a second GPT-6 Sol post implied a much slower rate.
Will there be GPT-6 Terra, GPT-6 Luna or mini models?
Nobody has leaked them. OpenAI’s naming scheme allows GPT-6 versions of Terra and Luna. Its history includes smaller models shipped on launch day, 12 days later and 66 days later. Today the API still offers GPT-5.6 Terra and GPT-5.6 Luna below Astra.
How much will GPT-6 Sol cost?
OpenAI has not said. The current ladder runs from GPT-6 Astra at $10 input and $50 output per million tokens, through GPT-5.6 Sol at a promotional $4 and $20, down to GPT-5.6 Luna at $0.20 and $1.20. A GPT-6 Sol would most plausibly be priced between Astra and today’s Sol.
References
GPT-6 Astra: A new generation of intelligence (OpenAI)
GPT-5.6: Frontier intelligence that scales with your ambition (OpenAI)
Previewing GPT-5.6 Sol: a next-generation model (OpenAI)
Models (OpenAI API documentation)
Pricing (OpenAI API documentation)
Codex pricing (OpenAI Developers)
Introducing GPT-5.4 mini and nano (OpenAI)
Introducing GPT-5 for developers (OpenAI)
GPT-4o mini: advancing cost-efficient intelligence (OpenAI)
GPT-6 Sol, Terra, and Luna? (OpenAI Developer Community)
Sam Altman on X, 15 September 2026
Thibault Sottiaux on X, 15 September 2026
Lentils on X, 6 September 2026
Pankaj Kumar on X: GPT-6 Sol Leaks
GPT-6不只Astra！Sol内测结果曝光，速度快6倍 (QbitAI)
GPT-6 Sol Leak: OpenAI May Be Testing a Faster Model With Better Usage Limits (TheWinCentral)
GPT-6 Astra Helped OpenAI Attract More Enterprise Dollars Than Anthropic Last Week (Wccftech)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.