OpenAI GPT-6: release date, specs, training compute — what we know (Sept 2026)

OpenAI has not officially confirmed GPT-6 as of September 7, 2026. The model name has not been announced, the release date is unknown, and most of what is "known" is industry inference, supply-chain reporting, and Sam Altman's public statements about OpenAI's roadmap. This article consolidates the most credible reporting on what comes after GPT-5 (released August 2025) — whether it's called GPT-6, GPT-5.5, or something else entirely.

GPT-5 was the last "GPT-n" model?

In a June 2026 Sequoia AI Summit interview, Sam Altman hinted that GPT-5 might be the last model in the GPT-n lineage. "We're going to keep naming things, but the actual architectural jumps are getting smaller each generation. The big jumps now come from inference-time compute and tool use, not from retraining a bigger base model."

This suggests two possible paths:

1. GPT-6 ships in 2027 — a successor to GPT-5 with a meaningfully different architecture (e.g., persistent memory, multimodal native, agentic core). Likely 2-3× the training compute of GPT-5.
2. GPT-5.5 ships in late 2026 — a refinement of GPT-5 with new capabilities (e.g., agentic workflows, real-time learning). No GPT-6 in 2026.

OpenAI has used the "5.5" naming pattern only once (the rumored GPT-4.5 in early 2025 was rolled into GPT-5 pre-training). The pattern may not repeat.

What we know about GPT-6 / the next model

Based on supply-chain reporting, leaked OpenAI job postings, and Sam Altman's public statements:

Attribute	GPT-5 (current)	GPT-6 (estimated)
Release window	August 2025	Late 2026 to mid 2027
Training compute	~3 × 10^26 FLOPs	~1 × 10^27 FLOPs (3-5× GPT-5)
Parameters (active)	~500B (MoE, 50B active per token)	~1-2T (MoE, 100-200B active)
Context window	1M tokens	5-10M tokens
Modalities	Text, image, audio, video	+ persistent memory, agentic tool use, real-time world access
Reasoning	Multi-step inference, o3-style thinking	Continuous learning, persistent agent identity
Inference cost	$3 per million input tokens (API)	~$1-2 per million (3× efficiency from inference-time compute)
Multimodal	Native image, audio (voice mode), video (Sora integration)	Native real-time video, embodied (robot control)
Open-source	GPT-OSS-120B / 20B (released August 2025)	Likely an open-weight GPT-6-mini or similar

The Stargate compute deal: why GPT-6 is bottlenecked on hardware

OpenAI's "Stargate" compute deal (announced May 2026 with Oracle, SoftBank, and MGX) commits $500 billion over 5 years to build AI data centers. The first Stargate facility (Abilene, Texas) came online August 2026 with 2 GW of power capacity — enough to train a GPT-5-scale model continuously for 6 months.

This is the infrastructure that would train GPT-6. Without Stargate, GPT-6 training would be impossible by 2027.

Stargate timeline:

Phase	Capacity	Online	Use
Phase 1 (Abilene TX)	2 GW	Aug 2026	GPT-6 training
Phase 2 (5 additional US sites)	8 GW total	2027	GPT-6 training + inference
Phase 3 (international — UAE, Norway)	15 GW total	2028-2029	Global inference + GPT-7

GPT-6's release is gated by Stargate Phase 1 completion (already done) and a 6-9 month training run. Best estimate: GPT-6 announcement late 2026 or Q1 2027, public release mid 2027.

What GPT-6 will be able to do (vs GPT-5)

Six capability jumps expected over GPT-5:

1. Persistent memory across sessions — GPT-5 has chat-level memory (recalled within a session). GPT-6 is expected to add user-level persistent memory (recalled across all sessions, with privacy controls).
2. Real-time world access — GPT-5 uses retrieval-augmented generation (RAG) for up-to-date info. GPT-6 is expected to add native real-time web access, eliminating the RAG latency.
3. Autonomous multi-step agents — GPT-5 can do agentic workflows with human oversight. GPT-6 is expected to handle multi-day autonomous tasks (e.g., "plan my Q4 marketing campaign, execute the email outreach, report back").
4. Native video understanding and generation — GPT-5 has separate Sora for video. GPT-6 is expected to integrate video understanding + generation natively (Sora 2 architecture absorbed).
5. Embodied AI integration — Per OpenAI's robotics partnership with Figure AI (announced March 2026), GPT-6 is expected to power humanoid robot reasoning.
6. Scientific reasoning breakthroughs — GPT-6 is being trained with explicit scientific reasoning datasets (chemistry, biology, physics). Goal: AI co-scientist for drug discovery and materials science.

When will GPT-6 be announced?

The most likely timeline:

Date	Event
Sept 9 2026	Apple's "Surprise and Shine" keynote — Apple Intelligence features powered by GPT-5 (no GPT-6 yet)
Oct 2026	OpenAI DevDay — possible GPT-6 preview or "GPT-5.5" refresh
Late 2026 / Q1 2027	GPT-6 announcement (possibly at OpenAI's Spring event or another DevDay)
Mid 2027	GPT-6 public release (ChatGPT Pro / API)
Late 2027	Open-weight GPT-6-mini / GPT-6-nano release

OpenAI's pattern: announce a model 3-6 months before public release. DevDay (October) is the most likely announcement venue. If GPT-6 ships at DevDay 2026, it will be available to API customers in Q4 2026 and to ChatGPT users in early 2027.

What GPT-6 means for ChatGPT users

For everyday ChatGPT users, the GPT-5 → GPT-6 jump will be less dramatic than GPT-4 → GPT-5. Most user-facing improvements come from inference-time compute (better reasoning, longer planning) rather than raw model size.

The biggest user-facing changes:

• Memory that actually works — GPT-5's chat history recall is limited. GPT-6 will remember user preferences, project context, and past interactions across sessions.
• Agent mode default — GPT-5 has agent mode as opt-in. GPT-6 is expected to make agentic behavior default for paid tiers.
• Real-time info without hallucinations — GPT-5 hallucinates on recent events. GPT-6's native real-time web access should reduce hallucinations on time-sensitive queries.
• Voice that sounds human — GPT-5's voice mode is good but stilted. GPT-6 with end-to-end audio modeling should pass as human in phone calls.

Competitive landscape: GPT-6 vs Gemini 3 vs Claude 4

Model	Company	Status	Differentiator
GPT-6	OpenAI	Announced late 2026 / Q1 2027 (rumored)	Best agentic capabilities, deepest memory
Gemini 3 Ultra	Google DeepMind	Live (I/O 2026)	Native real-time video, Project Astra integration
Claude 4 Opus	Anthropic	Live (Q1 2026)	Best coding + long-context reasoning, safety focus
Grok 4	xAI	Live (Q2 2026)	Real-time X/Twitter integration, edgier personality
Llama 5 405B	Meta	Open-source, Q4 2026 (rumored)	Free for commercial use, customizable

GPT-6's competitive moat is the Stargate compute advantage: 2 GW of dedicated training capacity is more than the combined training compute of every other AI lab except Google. This means GPT-6 can be trained at a scale no one else can match.

What to watch for

Three signals that GPT-6 is imminent:

1. Sam Altman tweets "interesting things cooking" — his usual pre-announcement hint
2. OpenAI DevDay (October) agenda — check if there's a "next-generation model" session
3. GPT-5 deprecation announcement — OpenAI typically announces a successor 6 months before sunsetting the previous model
4. Stargate Phase 1 training run completion — when the Abilene facility completes its first 6-month training cycle

For related reading, see our OpenAI Project Astra deep dive, Sam Altman AGI countdown, and Jensen Huang on AGI and Nvidia GPUs.