Receipts: the mounting evidence
- Sam Altman’s weekend selfie-leak: The CEO shared a screenshot with “ChatGPT 5” in the menubar, because apparently NDAs don’t apply to founders.
- Applied-research boss can’t sit still: On Monday, OpenAI’s head of applied research posted that he’s “excited to see how the public receives GPT-5!”, not exactly poker-face messaging.
- Microsoft’s Azure upgrade: Redmond has reportedly been stockpiling GPU racks for months in anticipation, suggesting GPT-5’s appetite will dwarf even GPT-4o’s.
- Altman’s own hype timeline: Back in July he told podcaster Theo Von he was already test-driving the model and felt “useless relative to the AI.”
- The “big-but-small today, big later this week” tweet: Altman teased Tuesday’s announcement of gpt-oss, two open-weight models, and promised a heavier drop later in the week.
OpenAI announced tomorrow’s LIVE5TREAM on X
Polymarket odds are closing in all but confirming launch tomorrow, Source: Polymarket
Why GPT-5 matters beyond the version number
- Performance > Scale
GPT-4 dropped in March 2023, which feels like pre-Cambrian in LLM time. GPT-4.5 landed this February with longer context windows and better tool use. The industry expects GPT-5 to marry that with sharper reasoning, a multimodal “eyes and ears” upgrade, and, if the gossip is right, a context window creeping toward 256 k tokens, square in Gemini 2.5’s crosshairs. Google upped the ante on reasoning and Anthropic’s Claude 4 series is already cannibalizing dev mindshare for code. OpenAI needs a leap, not a shuffle, to keep ChatGPT the default tab.
- Capacity crunch
Altman keeps warning that launch week could be “slightly choppy,” which is corporate-speak for “pray the GPUs don’t melt.” Microsoft’s new racks aside, even Azure feels the strain when a hundred million prompt junkies all hammer “/everlong lyrics but in the voice of Yoda” at once. Expect staged access tiers or a wait-list.
- Open vs. Closed
Tuesday’s gpt-oss-120B and -20B releases mark the first open-weight drop since GPT-2. OpenAI’s pivot toward transparency looks strategic: toss hackers a bone, divert scrutiny from the proprietary flagship, and position GPT-5 as the premium upsell. Apache 2.0 licensing means anyone can fine-tune, great for startups, hair-raising for regulators.
Reasons to keep your skeptic hat handy
- Still no parameter count: The company hasn’t confirmed whether GPT-5 is a “modest” 300 B or something more Giga-scale. Sam’s coyness suggests they’re bracing for carbon-footprint backlash.
- Safety trade-offs: Remember GPT-4o’s too-nice fiasco that let users coax self-harm advice? Faster iterations make safety nets harder to test. If GPT-5 ships half-baked, Altman’s “slightly choppy” could turn into regulatory tsunami.
- Over-promising déjà vu: GPT-5 was “coming early 2025,” then “spring,” then “Q2.” Missed ship dates are practically tradition now.
What to watch Thursday at 10 AM PT
- Multimodal demos: If the livestream shows real-time video reasoning, that’s a differentiator. Otherwise, expect incremental text-only gains.
- Pricing shake-up: GPT-4’s paywall throttled adoption; GPT-5 might introduce consumption-based tiers or bring GPT-4 down to freemium to widen the funnel.
- Hardware partnerships: Rumor mill says NVIDIA’s Blackwell chips and a couple of AWS regions are already whitelisted for GPT-5 fine-tuning jobs.
- Trust & Safety slide deck: If the team dedicates more than 30 seconds to red-teaming, regulators will breathe easier; if not, cue Senate hearing invites.
The contrarian take
The AI press, yours truly included, gets high on version numbers. But step back: gpt-oss-120B is arguably the bigger disruptor. An Apache-licensed, 117 B-parameter model that runs on a single 80 GB GPU guts the moat OpenAI spent five years digging. If GPT-5 is merely “GPT-4-plus-epsilon,” the open-weight release could prove the real story by commoditizing good-enough reasoning for the masses.
Either way, Thursday’s “LIVE5TREAM” will tell us if OpenAI can still set the agenda, or if it just set itself up for the loudest reality-check in Silicon Valley.
