GPT-6 is OpenAI's unreleased next-generation flagship — codenamed Spud. Pre-training finished March 24, 2026 at Stargate Abilene. This page is a live tracker of every credible GPT-6 rumor, leak, and benchmark prediction — every claim cited inline.
published · updated · by ZeroTwo Editorial
GPT-6 is OpenAI's unreleased next-generation frontier large language model, internally codenamed Spud. As of April 24, 2026 it has not been officially announced. Pre-training finished March 24, 2026 at the Stargate Abilene data centre, per reporting in Fello AI and Dr Alan Thompson's lifearchitect.ai. On the same day, Sam Altman publicly described the launch as "a few weeks" away.
The naming is not yet committed. Multiple analysts including TokenMix report that the public name will depend on the size of the performance delta over GPT-5.4 — small delta means a GPT-5.5 launch (which shipped April 23, 2026), large delta means GPT-6. With GPT-5.5 already on the market, a standalone GPT-6 release window has likely shifted to May–Q3 2026.
The headline expected feature is persistent cross-session memory. Sam Altman has called current AI memory "GPT-2 era" relative to what's coming. Per analyst reporting at Voiceflow and Greg Brockman's public statements, GPT-6 is also expected to ship native agentic execution — multi-step browse, code, file, and API actions without per-step prompting.
Each card carries a source URL and a 1–5 star credibility score. 5 stars = on-record from OpenAI or a verified primary source. 1 star = anonymous claim with no corroboration. We don't publish anything we can't link.
OpenAI internally confirmed that pre-training of the model codenamed 'Spud' finished March 24, 2026 at the Stargate Abilene data centre. Sam Altman publicly described the launch as 'a few weeks' away the same day. This is the most concrete GPT-6 milestone to date.
Polymarket's 'GPT-6 released by' contract has been re-priced multiple times in April. The April 30 contract has fallen from ~93% to ~45–72% depending on the day, with Q3 2026 priced as a meaningful tail. The most defensible window is now late April through late May 2026.
Multiple reporters note OpenAI has not committed to the GPT-6 name internally. If the performance delta over GPT-5.4 is large enough, the launch is GPT-6; if not, the model ships as GPT-5.5. (GPT-5.5 was confirmed and released April 23, 2026 — see OpenAI blog — leaving 'GPT-6' open as a future label.)
Sam Altman has repeatedly described current AI memory as being in the 'GPT-2 era' relative to what's coming, signalling persistent cross-session memory as the headline GPT-6 capability. Greg Brockman has emphasised that the next model better infers context without explicit re-prompting.
An unverified insider claim circulated April 10–14 of a 2,000,000-token context window and a ~40% benchmark improvement over GPT-5.4, alongside a 'super app' bundling Codex and the Atlas browser. The April 14 launch date passed with no announcement; numbers are suspiciously precise. Treat as low-credibility until corroborated.
Epoch AI's training-compute analysis argues GPT-5 used less compute than GPT-4.5, but GPT-6 'probably won't' — Stargate's GB200 capacity finally lets OpenAI scale a base model again. Spud was trained at the Abilene Stargate site where Oracle has deployed >450,000 GB200 GPUs under a 15-year lease.
TokenMix's GPT-6 pricing forecast extrapolates from GPT-5.4's $1.75 in / $14 out and GPT-5.5's $5 in / $30 out. Expected GPT-6 launch range: $10–$30 per million input tokens, $40–$80 output, with a 2× multiplier above 200K-token context — explicitly speculative.
GPT-5.5 appeared in OpenAI's internal Codex model picker on April 22, hours before Altman heavily hinted at a Thursday release. GPT-5.5 then shipped April 23, 2026 — meaning the 'next model' shipped under the 5.5 banner, not 6. The GPT-6 rumour cycle continues.
Multiple analysts including Voiceflow and lifearchitect expect GPT-6 to ship native agentic execution — autonomous multi-step browse, code, file, and API actions without per-step prompting. This aligns with OpenAI's Operator and Codex roadmap and would push token usage per task into the 50K–500K range.
The exact channels we refresh whenever the rumor cycle heats up. If something hits any of these, this page updates within a day.
openai.com/blog — official 'Introducing GPT-6' is the only definitive signal
@sama posts have preceded every model launch by 24–72 hours
Live odds on the public release date; reprices daily
developers.openai.com/api/docs/models — first place a 'gpt-6' ID will appear
Past launches have leaked via Codex Pro accounts hours before announcement
Oracle GB200 deployment reports from The Information / Bloomberg
None of the named "GPT-6 competitors" — Claude Opus 5, Grok 5, Gemini 4 — have shipped under those exact names. Here is the actual April 2026 frontier.
| Model | Released | Known for | Source |
|---|---|---|---|
| GPT-5.5 (current OpenAI flagship) | Apr 23 2026 | 82.7% Terminal-Bench 2.0, 88.7% SWE-bench, 1M context, $5/$30 | openai.com |
| Claude Opus 4.6 / Mythos preview (Anthropic) | Q1 2026 | Reasoning + writing leader; 1M context; $5 / $25 | aimagicx.com |
| Gemini 3.1 Pro (Google) | Q1 2026 | Multimodal + 2M context; best price-performance | aimagicx.com |
| Grok 4 (xAI) | 2025–26 | 75% SWE-bench Verified — narrowly leads coding bench | aimagicx.com |
See deeper write-ups on GPT-5, Gemini 2.5 Pro, and Grok.
"GPT-5.5 is a sign of the future. What's notable about GPT-5.5 is it demonstrates that rapid AI improvement is continuing."
Mollick's framing matters: GPT-5.5 just shipped April 23, 2026, but it landed without resetting the GPT-6 watch. Multiple Stargate sources cited by Epoch AI describe a separate, larger run still in progress — the actual GPT-6 scale-up enabled by Oracle's GB200 deployment in Abilene.
The cheapest hedge while waiting for GPT-6 is also the most honest test of "do I actually need it." Run the same prompt against the four current frontier models — GPT-5.5, Claude Opus 4.6, Gemini 3.1 Pro, Grok 4 — and see which one wins for your tasks. ZeroTwo's multi-model chat is the only consumer surface that puts all four side-by-side in one window.
When OpenAI ships GPT-6, ZeroTwo adds it to the model selector the same day — every previous frontier launch (GPT-5, GPT-5.2, 5.4, 5.5; Claude Opus 4.5, 4.6; Gemini 3 Pro; Grok 4) appeared within hours. That means no separate ChatGPT Plus subscription, no API keys, no waiting list. See our ChatGPT alternative comparison for the side-by-side breakdown.
For developers, the highest-leverage prep is building agentic tooling now against GPT-5.5's existing multi-step tool calling. Per Voiceflow's analysis, GPT-6's expected agentic capability extends — not replaces — the current Codex / Operator surface. Build for tool use today and the upgrade path is mostly free.
Ten direct answers, sourced from Polymarket, OpenAI, Epoch AI, Fello AI, and TokenMix.
As of April 24, 2026, GPT-6 has not been officially released. OpenAI completed pre-training of the model codenamed 'Spud' on March 24, 2026 at the Stargate Abilene facility, and Sam Altman publicly described the timeline as 'a few weeks' the same day. Polymarket's 'GPT-6 released by' contract currently prices late April through late May 2026 as the most likely window, with a meaningful tail extending into Q3 2026. Note that GPT-5.5 launched on April 23, 2026, suggesting the next major numbered release may slip further into 2026.
OpenAI has not officially confirmed the GPT-6 name. Reporting from TokenMix, Fello AI, and lifearchitect indicates the public name will depend on how large the performance delta is over GPT-5.4 — if the gap is small, the model ships as GPT-5.5 (and indeed GPT-5.5 launched April 23, 2026); if the gap is large, GPT-6. As of the most recent OpenAI blog post, no GPT-6 announcement exists.
Expected capabilities, per Sam Altman public comments and analyst speculation: persistent cross-session memory (Altman called current AI memory 'GPT-2 era'), native agentic execution (multi-step browse / code / API actions), substantially longer context (rumours of 2M tokens, unverified), better multimodal grounding, and lower hallucination rates. None of these are confirmed in writing yet.
GPT-6 pricing is purely speculative. TokenMix's forecast extrapolates from GPT-5.4 ($1.75 in / $14 out) and GPT-5.5 ($5 in / $30 out per million tokens) to suggest a $10–$30 input / $40–$80 output range at launch, with a likely 2x multiplier for context above 200K tokens. Expect a Pro tier in the $200–$300/month ChatGPT range, mirroring GPT-5 Pro.
Unverified. The 2M-token context claim originates from an April 10 anonymous leak that also predicted an April 14 launch date that did not occur. Suspiciously precise numbers from unverified sources should be treated as rumour. GPT-5.4 already extended to 1.05M tokens and GPT-5.5 ships at 1M, so 2M for GPT-6 is plausible but not confirmed.
OpenAI's Stargate facility in Abilene, Texas. Oracle has deployed over 450,000 NVIDIA GB200 GPUs at the Abilene campus under a 15-year lease, and Stargate's first two buildings went operational in September 2025. Epoch AI's compute analysis confirms GPT-6 will use materially more training compute than GPT-5, which itself used less than GPT-4.5.
GPT-5 (Aug 7, 2025) and the GPT-5.x family already cover most of OpenAI's near-term roadmap — 400K to 1M context, 88.7% SWE-bench Verified (5.5), unified fast/thinking modes. The expected step-change for GPT-6 is in three places: persistent memory across sessions, native multi-step agentic execution, and a meaningful step in compute (>2x GPT-5 training FLOPs, per Epoch AI). Hallucinations are expected to drop further versus the 60% reduction GPT-5.5 already delivered.
As of April 24, 2026, none of the named GPT-6 competitors have shipped under those exact names. Claude Opus 4.6 leads on reasoning and writing benchmarks; Grok 4 narrowly leads SWE-bench at 75%; Gemini 3.1 Pro leads multimodal and price-performance; GPT-5.5 currently tops Terminal-Bench 2.0 at 82.7%. Expect Anthropic and Google to respond within weeks of any GPT-6 announcement.
Probably not. ChatGPT Plus locks you into one provider at $20/month for GPT-5.x access. ZeroTwo bundles GPT-5.5, Claude Opus 4.6, Gemini 3.1 Pro, and Grok 4 in one workspace for $29.99/month, with day-zero access to GPT-6 once it ships. Side-by-side comparison across providers is the only way to know which model handles your specific tasks best.
Sign in to ZeroTwo today and use GPT-5.5 alongside Claude, Gemini, and Grok. ZeroTwo adds new frontier models to the model selector within hours of release — including the entire GPT-5 family within a day of OpenAI's August 2025 launch. Your existing subscription unlocks GPT-6 the moment it appears in the OpenAI API, with no upgrade required.
The ZeroTwo editorial team tracks every frontier model release in real time, refreshes our 60+ model catalog, and updates pages like this one with primary-source citations within a day of any new credible signal.
published · updated
Use GPT-5.5 — plus Claude, Gemini, and Grok — on ZeroTwo today. The moment OpenAI ships GPT-6, it lands in your model selector. No upgrade fee. No ChatGPT Plus required.
Try GPT-5.5 on ZeroTwo