GPT-5 Orion (Spud) Release Date & Leaked Feature List

The Short Answer

If you're here because you Googled "GPT-5 Orion release date" — the answer, as of April 2026, is: pretraining is done, safety evaluation is underway, and public release is expected within weeks. OpenAI president Greg Brockman confirmed on April 2 that the model — internally called Spud — carries two years of core research and will launch as GPT-5.5.

This is not an incremental GPT-5.4 update. It's a full new pretraining run built from scratch at OpenAI's Stargate facility in Abilene, Texas.

Release Timeline: What We Know

Here's the confirmed chronology, based on public statements from Altman and Brockman — no speculation, only sourced events:

Orion vs. Spud — Codename Confusion Explained

Quick clarification, because the internet is messy on this.

• "Orion" was an early internal codename referenced in a 2024 The Verge leak — Sam Altman himself called that report "fake news out of control." Orion was eventually repurposed as the internal name for GPT-4.5, not GPT-5.
• "Spud" is the confirmed current codename for OpenAI's next frontier model, which will be released as GPT-5.5.
• So if you're searching "GPT-5 Orion" — you're searching for the right idea, just an outdated label. GPT-5.5 Spud is what you actually want to track.

GPT-5.5 Leaked Feature List

No official benchmark data exists yet. What Brockman shared is qualitative, and what the AI research community has pieced together comes from internal test leaks and patent filings. Here's what's credible:

What Brockman Confirmed Directly

• Brand-new pretraining run — not a fine-tune or RLHF patch on GPT-5.4. A genuinely new base model.
• Models that "feel smarter" — Brockman described what he calls the "big model smell": the AI stops being rigid and starts actually understanding your intent without you having to rephrase three times.
• Longer autonomous task horizons — the model can stay on-task across more complex, open-ended work without losing context or drifting.
• Intuitive intent alignment — the frustrating "it kinda gets it but not quite" friction he says is finally going away.

What Credible Leaks Suggest

• Extended context window — likely 2M tokens or beyond, building on GPT-5.4's 1M baseline.
• Stronger instruction-following at depth — multi-step agentic chains with less hallucination drift.
• Reduced hallucination rate on factual queries — corroborated by the fact that Fields Medal winner Terence Tao tested its math proofs.
• More natural conversational flow — Altman's comment that GPT-5.4 is "my favorite model to talk to" suggests Spud will take that further.
• Potential enterprise pricing revision — Q2 2026 has 3+ competing frontier models launching simultaneously. OpenAI has room to negotiate.

What We Don't Know Yet

• No public benchmarks (SWE-bench, MMLU, etc.) — anything you read is speculation.
• API pricing is unconfirmed.
• Whether it ships as GPT-5.5 or gets rebranded closer to launch (OpenAI has done this before).
• Whether the alignment phase surfaces any delays — Spud is described as a significantly stronger model, which often means more safety iterations.

Expected Capability Profile (Based on Confirmed Info)

*Scores are editorial projections based on confirmed qualitative descriptions. Not official benchmarks.

My First-Hand Experience with GPT-5.4 — and What Spud Should Fix

I've been running GPT-5.4 as my primary research and writing assistant since its March 5 launch. Here's where it still frustrates me — and why Spud's described improvements matter in practice.

Scenario 1: Multi-Step Briefing Documents

Last week I used GPT-5.4 to pull together a competitive briefing across six AI tools — 400+ pages of documentation, changelog notes, and blog posts fed in via the 1M token context window.

It handled the first pass well. The problem showed up around page 280: the model started subtly blending details from two different products. Not hallucinating outright, but conflating. I had to break the job into two separate sessions and manually reconcile them. That's exactly the "longer autonomous task horizon" gap Brockman says Spud addresses.

Scenario 2: Coding Agent Loops

GPT-5.4's computer-use and coding agent features work, but they're verbose. Ask it to refactor a Node.js service and it'll explain every decision in the middle of execution, sometimes pausing to ask for confirmation on things that don't need it.

Brockman's "models become more compliant when they're actually smarter" comment lands here. GPT-5.4 sometimes feels like it's performing intelligence rather than just doing the job quietly. If Spud delivers on the intent-reading promise, that constant back-and-forth overhead should shrink.

GPT-5.5 vs. GPT-5.4 vs. Competitors

Spud doesn't launch in a vacuum. Q2 2026 is the most contested AI quarter ever, with DeepSeek V4, Grok 5, Claude Mythos, and Gemini 3.2 all expected in the same window.

⚠️ The Common Pitfall Nobody's Warning You About

Here's the thing most GPT-5.5 coverage will miss: brand-new pretraining almost always means a temporary regression in some narrow skill areas.

When OpenAI shipped GPT-4.5 (Orion) in early 2025, several developer teams noticed it was worse at specific JSON schema compliance tasks than GPT-4o had been — despite being "smarter" overall. OpenAI patched it within a month, but teams who upgraded on day one hit real production bugs.

A fully new pretraining run like Spud carries the same risk. The model may be qualitatively more capable and still mishandle an edge case your workflow relies on.

Practical playbook for teams:

• Run your existing GPT-5.4 eval suite against Spud before migrating.
• Don't migrate production API calls on day one — let the community surface regressions first (give it 2 weeks).
• Keep GPT-5.4 as a fallback in your routing layer. This is cheap insurance.
• Watch the OpenAI developer forum the first 48 hours after launch. That's where the real-world issues surface first.

Who Should Actually Care About GPT-5.5

Worth the Wait / Upgrade

• Developers running long agentic pipelines — longer task horizon is the single biggest practical improvement.
• Researchers and analysts working with large document corpora who hit GPT-5.4's context drift problem.
• STEM-heavy use cases — Terence Tao's math reasoning tests are a strong signal.
• Enterprise teams where API pricing can be renegotiated — Q2's five concurrent launches give you real leverage.

Probably Fine to Wait

• Casual ChatGPT Plus users — GPT-5.4 is already excellent for most daily tasks.
• Teams tightly integrated with a Google Workspace stack — Gemini 3.1 still wins that integration race.
• Budget-sensitive projects — DeepSeek V4 at $0.30/MTok with 1M context may be the smarter call.

Frequently Asked Questions