What Is GPT-6? OpenAI's Next Frontier Model

The 10T-Parameter Successor to GPT-5.6: A First Look at GPT-6

GPT-6 launched on September 3, 2026, with GPT-6 Astra as OpenAI's flagship frontier model. OpenAI positions Astra as state of the art across computer use, browsing, software engineering, cybersecurity, science, and professional work. The launch announcement reports a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench. GPT-6 Astra has a 1,050,000-token context window, a 128,000-token maximum output, and published API pricing. GPT-6 Sol and GPT-6 Luna joined the family on September 22 as faster, more affordable models. OpenAI has not published a parameter count or complete architecture specification.

The model first surfaced in community leaks in early July 2026 as the generational successor to the GPT-5.x line, positioned as OpenAI's response to Anthropic's Fable and Mythos family. Those pre-launch reports supplied several estimates, including a rumored 10T-scale model and a 1.5M-plus token context window. The launch confirms GPT-6 and its public family, while neither the earlier parameter estimate nor the larger context estimate is an official specification.

Key Takeaways

• GPT-6 Astra launched on September 3, 2026, as OpenAI's flagship model after GPT-5.6.
• GPT-6 Sol and GPT-6 Luna launched on September 22, extending the GPT-6 family with faster, more affordable models.
• OpenAI describes Astra as state of the art across computer use, browsing, software engineering, cybersecurity, science, and professional work.
• Confirmed Astra headline results include 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench.
• GPT-6 Astra has a 1,050,000-token context window, a 128,000-token maximum output, and an April 30, 2026 knowledge cutoff.
• Astra API pricing is $10 per 1 million input tokens and $50 per 1 million output tokens, with cached input at $1 and cache writes at $12.50.
• GPT-6 Sol costs $2 per 1 million input tokens and $10 per 1 million output tokens. GPT-6 Luna costs $0.10 per 1 million input tokens and $0.50 per 1 million output tokens.
• Astra is available through a staged rollout involving organizations, ChatGPT, the OpenAI API, and AWS. Sol and Luna are available through ChatGPT Work, Codex, and the API, with Luna also available in the desktop app for Free and Go users.
• OpenAI has not published a parameter count. The earlier 10T-scale community claim remains unconfirmed.

What Is GPT-6?

GPT-6 Astra is the flagship model OpenAI launched after GPT-5.6. GPT-6 is also a family: Astra is its flagship, while Sol and Luna are faster, lower-cost members launched later in September. The family is not a point release inside the GPT-5 line. OpenAI's launch materials describe Astra as a new generation of intelligence with capabilities spanning computer workflows, coding, cybersecurity, scientific work, and professional tasks.

The name first appeared in verifiable form in mid-2025, when OpenAI CEO Sam Altman told CNBC that GPT-6 would arrive faster than the gap between GPT-4 and GPT-5 and would center on memory and personalization. Through the first half of 2026, the community expected the "Spud" pretraining base to become GPT-6, but that model shipped as GPT-5.5 on April 23, 2026, followed by GPT-5.6 on July 9, 2026. Only after GPT-5.6's release did leakers converge on the picture of GPT-6 as a new, larger pretrain beyond the public GPT-5 line.

That pre-launch picture changed on September 3, when OpenAI introduced GPT-6 Astra. The launch materials identify Astra as OpenAI's most intelligent and aligned model yet and describe state-of-the-art performance across computer use, browsing, software engineering, cybersecurity, science, and professional work.

OpenAI expanded the family on September 22 with GPT-6 Sol and GPT-6 Luna. The two models build on strengths associated with Astra while emphasizing faster inference, more efficient caching, and lower cost. OpenAI positioned Sol for complex professional work and coding, while Luna is aimed more at frequent, lower-cost tasks.

GPT-6's public status is now clear, even though access remains staged by model and surface. GPT-6 Astra launched on September 3, 2026. GPT-6 Sol and GPT-6 Luna launched on September 22, 2026.

GPT-6 at a Glance

Field	Value
Developer	OpenAI
Type	Frontier model family with computer-use, coding, science, and professional workflows
Flagship	GPT-6 Astra
Additional launched models	GPT-6 Sol and GPT-6 Luna
Predecessor	GPT-5.6 (Sol / Terra / Luna tiers)
Pretraining base	OpenAI has not published a complete architecture description; pre-launch reports described a new foundation distinct from the roughly 4T-token "Spud" base
Astra context window	1,050,000 tokens
Astra maximum output	128,000 tokens
Parameter count	Not published by OpenAI; the pre-launch 10T-scale claim remains unconfirmed
Modality and tools	Text and image input, with computer, browser, terminal, and software workflows described for Astra
Astra knowledge cutoff	April 30, 2026
Astra pricing	$10 per 1M input tokens; $1 cached input; $12.50 cache writes; $50 output
Sol pricing	$2 per 1M input tokens; $10 output
Luna pricing	$0.10 per 1M input tokens; $0.50 output
Availability	Astra launched September 3, 2026; Sol and Luna launched September 22, 2026, with staged access through OpenAI products and the API
License	Proprietary
Model documentation	OpenAI model documentation and a system card are available alongside the launch materials

How GPT-6 Works and What Makes It Different

The single most repeated pre-launch claim was that GPT-6 would not be a continuation of the Spud pretraining foundation. GPT-5.5 and GPT-5.6 were described as sharing that roughly 4T-token base. Community reports described GPT-6 as a new run on a much larger corpus and a bigger model, but OpenAI has not published a parameter count or a full architecture comparison. The 10T-scale estimate should therefore remain labeled as a pre-launch claim, not a confirmed specification.

The launch gives a clearer picture of what the family is optimized to do. GPT-6 Astra is built to work with computers and software rather than only produce conversational answers. OpenAI lists tasks such as filling out forms, updating customer records, organizing calendars, conducting online research, drafting summaries, analyzing scientific data, generating plots, creating websites, running frontend quality checks, installing and testing software, and troubleshooting problems visible on screen.

GPT-6 Sol and Luna apply parts of the Astra approach to lower-cost operating points. OpenAI says the models benefit from improved caching and inference efficiency and are 50% cheaper in API token pricing than GPT-5.6 promotional pricing. Independent evaluation reported in the supplied launch material found aggregate intelligence and coding scores roughly level with GPT-5.6, with gains on some tests offset by regressions on others. The practical distinction is therefore not simply a claim of across-the-board capability growth: Sol and Luna emphasize efficiency, speed, and cost per task.

Three coined anchors show up consistently in the earlier community discussion:

• Spud Successor — the framing that GPT-6 replaces, rather than extends, the Spud base. The launch confirms Astra's position as the next major model after GPT-5.6, but not the architecture details behind that transition.
• Fable++ — Bindu Reddy's July 21 characterization that GPT-6 would be "considerably more efficient than Fable," Anthropic's frontier model, on very complex workloads. Astra's launch results now provide a basis for comparisons, although the exact tests and costs still matter.
• Serving Architecture Advantage — the argument that OpenAI's edge is not just the model but the infrastructure to serve it at scale, which Haider argued Anthropic has struggled with on the Fable side. Astra's staged rollout and the lower-cost Sol and Luna releases show that infrastructure and access capacity remain part of the practical product story.

On direction, the older but on-record signal is Altman's memory framing. In August 2025 he told CNBC: "People want memory. People want product features that require us to be able to understand them." The GPT-6 launch emphasizes computer use, judgment, alignment, and long-running professional work rather than publishing a separate specification for persistent memory.

What You Can Do With GPT-6

GPT-6 Astra is now available through a staged rollout, and Sol and Luna are available through additional OpenAI work surfaces. The concrete use cases are no longer only extrapolations from pre-launch reporting.

• Computer-use workflows: Astra can fill out online forms, update CRM records, organize calendars, conduct online research, and draft summaries in email or document editors. OpenAI describes it as a model that can handle demanding professional computer tasks.
• Long-horizon coding: Astra can create websites, install and test software, troubleshoot problems on screen, and run frontend quality checks. If you need a production workflow before Astra reaches your account, the current predecessor is GPT-5.6.
• Agentic research and science: OpenAI reports state-of-the-art performance on Terminal-Bench Science 0.1 and HealthBench Pro. The model can analyze scientific data, run simulations, fit models, and use code and terminal tools in research workflows.
• Mathematical discovery: An internal version of Astra generated new results for 10 long-standing open problems in mathematics and theoretical computer science. OpenAI says the total inference would have cost roughly $2,000 at Sol API rates, after which the arguments were prepared into manuscripts and formalized in Lean certificates. The results include work on sphere packing, coding theory, non-sofic groups, Connes's rigidity conjecture, arithmetic circuit complexity, quantum parallel repetition, the closest vector problem, Ehrhart's volume conjecture, multicolor Ramsey numbers, and extremal graph theory.
• Creative and 3D work: Users have demonstrated GPT-6 working in Blender, building games, creating interactive environments, and producing complete fonts from visual references. These are hands-on showcase reports rather than standardized evaluations, but they illustrate the model family's computer-use orientation.
• Lower-cost agent workflows: Sol and Luna are designed for work at scale. Sol targets complex professional work and coding, while Luna is aimed at frequent, lower-cost tasks. Their lower token prices and improved caching are intended to reduce the cost of repeated or tool-heavy workflows.
• Cross-session personalization: Memory and personalization remain part of the broader GPT direction signaled by Altman, but the supplied GPT-6 launch materials do not publish a separate, detailed memory specification.

How GPT-6 Compares

The comparison landscape changed twice in September: Astra launched on September 3, and Sol and Luna followed on September 22.

Model	Vendor	Status	Notable claim
GPT-6 Astra	OpenAI	Released September 3, 2026; staged rollout	State of the art across computer use, coding, cybersecurity, science, and professional work
GPT-6 Sol	OpenAI	Released September 22, 2026	Faster, lower-cost model for complex professional work and coding
GPT-6 Luna	OpenAI	Released September 22, 2026	Faster, lower-cost model for frequent, lower-cost tasks
GPT-5.6 (Sol)	OpenAI	Released July 9, 2026	Predecessor to the GPT-6 family
Claude Fable 5.1	Anthropic	Released	Comparison point for Astra's science, coding, and agentic performance
Kimi K3	Moonshot	Released	Challenger in the broader frontier-model market

OpenAI's launch page reports 59.3% for Astra on Agents' Last Exam, compared with 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol. On OSWorld 2.0, Astra scored 72.6% versus Sol's 65.7% and completed tasks in approximately 47% less time in the cited latency simulation.

OpenAI also reports 99.9% on ARC-AGI 3, 98% on FrontierMath Tier 4, and 100% on ExploitBench. Those are headline results from OpenAI's launch materials and should be read alongside independent evaluations rather than treated as a complete ranking of every model or workload.

The newer Sol and Luna comparisons are more mixed. Artificial Analysis reported intelligence and coding scores roughly level with GPT-5.6, with improvements on some evaluations offset by regressions on others. Community tests also show task-dependent results: one coding test favored Astra on bug finding, while separate 3D assessments favored Claude Opus 5.5 on output quality and GPT-6 Sol or Luna on speed and cost. These tests use different prompts, settings, and evaluation methods.

For a deeper look at the competitive frame, see our earlier analysis of the Claude Opus 5 vs GPT-5.6 matchup and the Grok 4.6 vs Kimi K3 efficiency comparison.

Availability: How to Access GPT-6

GPT-6 Astra launched on September 3, 2026. OpenAI initially offered it to a limited set of organizations and said it would become available over the following days to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS.

GPT-6 Sol and GPT-6 Luna launched on September 22. OpenAI said they were rolling out to ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Education users, and were already live in the API. Free and Go users can try Luna in the desktop app. The supplied launch material distinguishes these Work and Codex surfaces from ordinary Chat, whose separate availability for Sol and Luna was not established in the supplied announcements.

Surface	Status as of September 23
Limited organizations	Astra launched through the initial organization rollout on September 3
ChatGPT Plus, Pro, Business, and Enterprise	Astra rollout began after the initial organization access
GPT-6 Astra API	Available through the OpenAI API as part of the staged launch
AWS	Announced as an Astra access channel alongside the API
ChatGPT Work	Sol and Luna rolling out for Plus, Pro, Business, Enterprise, and Education users
Codex	Sol and Luna rolling out for Plus, Pro, Business, Enterprise, and Education users
OpenAI API	Sol and Luna are live
Desktop app	Luna is available for Free and Go users
Ordinary Chat	Separate Sol and Luna availability was not established in the supplied launch announcements
GPT-6 Astra Pro	A higher-tier flavor was referenced for Pro, Business, and Enterprise; separate complete pricing and limits are not published in the supplied launch materials

If you do not see Astra, Sol, or Luna in a particular account or surface, that may reflect the staged rollout and product-specific access terms rather than the models' launch status.

The published Astra API rates are $10 per 1 million input tokens and $50 per 1 million output tokens. Cached input is $1 per 1 million tokens, while cache writes cost $12.50 per 1 million tokens. Sol is priced at $2 per 1 million input tokens and $10 per 1 million output tokens. Luna is priced at $0.10 per 1 million input tokens and $0.50 per 1 million output tokens. OpenAI and the supplied launch reporting also describe a 90% discount for cached input reads on GPT-6 family models. The supplied launch materials do not establish a complete rate-limit table for every ChatGPT, Codex, or API tier.

What We Don't Know Yet

The launch resolves the model family's name, Astra's launch date, Sol and Luna's launch date, core positioning, Astra's headline context size, and initial pricing. These questions remain genuinely open:

• Parameter count and full architecture: OpenAI has not published the model size or a complete architecture specification. The pre-launch 10T-scale claim remains unverified.
• Separate Sol and Luna specifications: The supplied launch materials give prices, positioning, availability, and broad efficiency claims for Sol and Luna, but do not establish separate context-window, maximum-output, or knowledge-cutoff specifications.
• Rollout completion and exact limits: The supplied materials do not establish when every user in each ChatGPT, Codex, API, AWS, or desktop-app tier will receive access, or whether limits are identical across surfaces.
• Benchmark comparability: OpenAI's headline results and independent evaluations use different prompts, settings, harnesses, and workloads. No single supplied benchmark set provides a complete direct comparison among Astra, Sol, Luna, and competing models.
• Benchmark reliability and representativeness: The available comparisons are task-dependent. Some tests favor Astra or Opus 5.5, while others show Sol or Luna completing work faster or at lower cost. The supplied material contains no common independent evaluation covering every reported result.
• Safety and monitorability for Sol and Luna: OpenAI reports improved alignment and safety results for Astra, but the supplied material does not provide comparable safety figures for Sol or Luna.
• GPT-6 Terra's status: Sol and Luna launched on September 22, but the supplied launch reporting did not announce a public GPT-6 Terra release or explain whether Terra remains planned.
• AGI status: Some commentary describes Astra using AGI language, but the supplied OpenAI launch materials do not claim that GPT-6 Astra has achieved AGI.

Frequently Asked Questions

Is GPT-6 released yet?

GPT-6 launched on September 3, 2026, with Astra as its flagship model. GPT-6 Sol and GPT-6 Luna launched on September 22. Astra is rolling out through organizations, ChatGPT Plus, Pro, Business, and Enterprise, the OpenAI API, and AWS. Sol and Luna are rolling out through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Education users, through the API, and, for Luna, the desktop app for Free and Go users.

What is the difference between GPT-6 and GPT-5.6?

GPT-6 Astra is OpenAI's flagship model after GPT-5.6, with state-of-the-art capabilities across computer use, coding, cybersecurity, science, and professional work. GPT-6 Sol and Luna extend the GPT-6 family with faster, more affordable models. OpenAI has not published a complete architecture comparison or parameter count.

How many parameters does GPT-6 have?

OpenAI has not published GPT-6's parameter count. Pre-launch community reports cited a 10T-scale model, but that figure remains unconfirmed and should not be treated as a specification.

What is the context window of GPT-6?

GPT-6 Astra has a 1,050,000-token context window and a maximum output of 128,000 tokens. OpenAI lists an April 30, 2026 knowledge cutoff. The supplied launch materials do not establish separate context-window specifications for GPT-6 Sol or GPT-6 Luna.

When will GPT-6 be released?

GPT-6 Astra launched on September 3, 2026. GPT-6 Sol and GPT-6 Luna launched on September 22, 2026.

How much will GPT-6 cost?

GPT-6 Astra API pricing is $10 per 1 million input tokens and $50 per 1 million output tokens, with cached input at $1 per 1 million tokens and cache writes at $12.50. GPT-6 Sol is priced at $2 per 1 million input tokens and $10 per 1 million output tokens, while GPT-6 Luna is priced at $0.10 per 1 million input tokens and $0.50 per 1 million output tokens.

Is GPT-6 open source?

GPT-6 is offered as a proprietary hosted model family through OpenAI products and the API. The supplied launch materials do not identify a release of GPT-6 weights.

What to Watch Next

The next meaningful updates are no longer about whether GPT-6 exists or when it will launch. They are about rollout completion, the final access terms for each Astra, Sol, and Luna surface, the behavior of the family under real workloads, and independent evaluation of computer-use, coding, science, and cybersecurity performance.

The largest technical gap is still the missing parameter count and architecture description. The launch establishes Astra's public capabilities and pricing and adds lower-cost Sol and Luna variants, but it does not confirm the earlier 10T-scale estimate. The specifications of Sol and Luna beyond their pricing, positioning, and availability also remain incomplete.

The launch also leaves the place of GPT-6 Terra unresolved. Sol and Luna are public, while the supplied launch materials do not establish whether Terra will be released or how it would fit into the family.

Building similar frontier-scale chat and agent workflows? On kie.ai you can try GPT-5.6, Claude Opus 5, and Gemini 3.1 Pro.