If you've been seeing GPT-5.4 pop up in whispers and screenshots lately, it's not your imagination. The unusual part is where the breadcrumbs came from — not just "someone said something on social." We're talking public code changes, scrubbed references, and little accidents that look a lot like a real model sitting on real servers.
At the same time, there's a lot of emotion swirling around OpenAI right now, especially tied to government and defense work. That's been feeding a "QuitGPT" wave, with plenty of people canceling and trying other tools.
📋 Table of Contents
Quick Updates on the AI Drama (and Why It's Spilling Into Product News)
First, the big headline: the GPT-5.4 chatter is not just rumors. Multiple independent "oops" moments all point to the same thing — an upcoming model that OpenAI didn't mean to name publicly yet.
The messier backdrop: there's a growing backlash tied to OpenAI's relationship with the Department of Defense (the Pentagon). The frustration has turned into a real public movement — people canceling ChatGPT subscriptions on principle, or just out of annoyance, or both. The QuitGPT protest report on Business Insider captures the mood on the ground well.
🔔 Key Developments
- More users appear to be switching from OpenAI to Anthropic, at least by app download rankings.
- The "QuitGPT" angle got louder, showing up in mainstream coverage — Claude surpassed ChatGPT in U.S. app downloads per Axios.
- Anthropic's CEO reportedly called OpenAI's messaging around the military deal "mendacious" (i.e., straight-up lies) — per TechCrunch.
GPT-5.3 Instant Is Already Out — And It's Trying to Be Less Irritating
An everyday "just chatting" setup that matches how most people use fast default models. (AI-generated image)
Before GPT-5.4, there's an important stepping stone: GPT-5.3 Instant. This one matters because it's positioned as the default "fast chat" experience — the model people hit when they just want a decent answer without waiting.
The stated focus sounds almost humble: make ChatGPT less annoying to talk to. Here's what stands out:
- 27% fewer hallucinations — a big claim if it holds up in the messy real world.
- Less "calm down" energy — that phrase almost never lands the way it thinks it does. It doesn't soothe, it escalates.
- Better writing and search integration — answers should read cleaner and connect to useful info more smoothly.
- Fewer unnecessary refusals and weird caveats on harmless questions.
💡 Real-world example: You ask for a cardio music playlist and suddenly it's warning you about headphone decibel ranges, disturbing neighbors, and half a dozen other things that were not the point. If GPT-5.3 Instant actually reduces that "overcautious autopilot" tone, it's a genuine quality-of-life improvement.
For more context on where GPT-5.3 fits in the broader model rivalry, check out our earlier breakdown: GPT-5.3 vs Claude Opus 4.6 coding showdown.
The GPT-5.4 Leaks: Why People Are Taking Them Seriously
A developer spotting an "oops" moment in a code review. (AI-generated image)
What made this GPT-5.4 situation different is the repeat pattern — multiple sightings from different angles, all pointing to the same model name. Not once, not twice, but three times in a single week. And a lot of the signals came from OpenAI's own surfaces.
🔍 Leak #1: A GitHub Pull Request That Triggered a "Fire Drill" Cleanup
A pull request in OpenAI's public Codex GitHub repo referenced a "minimum model" set to 5.4, even though that model didn't officially exist in public product language. Then the cleanup happened fast — code was force-pushed repeatedly over several hours to scrub references from 5.4 back to 5.3.
"Typos get fixed. Fire drills get force-pushed over and over while people try to scrub every trace."
A clean summary of what showed up in the repo and how fast it was removed: Codex repo PRs referencing GPT-5.4 — AwesomeAgents.
🔍 Leak #2: The /slashfast Command That Referenced GPT-5.4 Directly
A second slip landed in similar territory: changes tied to a "fast mode" toggle in Codex. In the original code text, the fast toggle explicitly referenced GPT-5.4. That alone is pretty blunt — it suggests the team was working on a speed tier where you can trade cost for lower latency. Then it got scrubbed too, reportedly within a few hours.
Why this matters: Fast mode only makes sense when you've got enough demand, enough agent-style usage, or enough enterprise pressure that "waiting" becomes a problem. This is a sign of maturity and scale — not a weekend experiment.
🔍 Leak #3: Model Selector Sightings, Error Logs & Reporter Confirmation
The third leak was the classic kind: an employee screenshot showing GPT-5.4 in a model selection dropdown. The post was deleted, but screenshots don't really die once copied around. Then a journalist hit an error message that referenced a model ID including GPT-5.4 — suggesting the model is present on the server side, where IDs and routing live. Finally, a well-connected reporter at The Information confirmed the model exists and is running internally.
"When the same model name shows up in code, UI, and error logs, it stops feeling like fan fiction."
What GPT-5.4 Might Include: Context, Reasoning, Vision & Speed
🧠 A Bigger Context Window (Likely 1 Million Tokens)
Early buzz was a 2-million token context window. The more grounded expectation now looks like 1 million tokens. That's still massive — the difference between "paste a few docs" and "stuff a small library into the conversation."
| Model | Context Window | Status |
|---|---|---|
| GPT-5.2 | 400,000 tokens | Current |
| GPT-5.4 (leaked) | 1,000,000 tokens | Upcoming |
| Top Competing Models | 1,000,000 tokens | Current |
⚠️ Context size isn't magic by itself. Models can struggle to retrieve the right detail even when it's present — sometimes called "context rot." The retrieval behavior inside that context matters just as much as the raw size.
⚡ "Extreme Thinking Mode" and Better Long-Horizon Reliability
Another leaked feature is a ladder of reasoning effort: light → medium → high → extra high → extreme thinking. This points to inference-time compute — basically how long the model spends thinking before responding. In "extreme" mode, it might not be seconds or minutes. It could be closer to hours for genuinely hard tasks.
This matters most for agent workflows. Two common failure examples:
- An agent starts a benchmark with one model, then silently switches mid-task — completely defeating the point of the test.
- A model gets an early instruction like "don't delete emails without confirming," then later deletes everything anyway because it lost the thread.
Those aren't "dumb model" problems. They're reliability problems across time, steps, and tool actions. If GPT-5.4 improves long-running task behavior — staying aligned with the original goal across many steps — that might matter more than a raw IQ bump.
🖼️ Full-Resolution Image Inputs & A "Fast Lane" for Latency-Sensitive Work
One more practical upgrade: full-resolution image support. Images (JPEG, PNG, WEBP) won't be compressed the same way before analysis. This sounds small, but it's not. Compression can erase tiny details that matter:
- Code screenshots where a single character matters
- Diagrams and schematics with fine labels and lines
- Charts, scans, and detailed visual data
Then there's the speed tier concept again. A standard lane and a priority lane. When running real-time agents, waiting even 10 extra seconds can break the product experience.
Why OpenAI's Release Cadence Is Speeding Up
Instead of one giant "wait for GPT-5" moment, releases now seem to arrive closer to monthly. OpenAI has signaled this is deliberate, partly to avoid the hype-and-letdown cycle.
When people wait a long time, they build a model up in their head — then even a strong release feels underwhelming. Regular drops change expectations. You start to look for steady improvements, not a single "everything changes today" day. The other benefit is simple: it keeps pressure on competitors and keeps users engaged — especially useful when a "QuitGPT" wave is trending.
OpenAI's User Numbers & the QuitGPT Pressure Test
|
910M+
Weekly Active Users
|
1B
Target by End of 2025
|
#1
Claude in US App Downloads*
|
*During certain windows, per recent reports.
Despite the backlash, OpenAI is still enormous by usage. But the QuitGPT movement seems to be making a dent — some reports suggest Anthropic's Claude climbed past ChatGPT in estimated first-time downloads in the U.S.
📊 Downloads vs. Weekly Active Users: These are different beasts. Downloads measure momentum. Weekly active users measure gravity. Momentum can flip faster; gravity changes slower.
Also, the Pentagon angle isn't staying neatly contained. See our internal explainer on why so many people are reacting strongly: Pentagon–Anthropic dispute over Claude in military use.
Personal Takeaways
Two thoughts pulling in opposite directions — and both feel true.
On one hand, day-to-day friction with AI tools is real. The "calm down" thing is a perfect example. It's small, but it changes how the tool feels. Same with random caveats on simple questions. Those moments erode trust, even when the model is technically capable.
On the other hand, the defense and government side of the story keeps creeping into the product conversation. It's hard to completely separate "this tool helps me write and code" from "this tool is part of bigger power systems."
🔑 Key Lessons From Watching These Leaks Unfold
- The most useful clues are the boring ones: code references, model IDs, fast-mode toggles, context window numbers.
- That stuff has weight because it's tied to real engineering decisions — less performative than a teaser tweet.
- The one thing to watch in GPT-5.4: long-task failure rate. Can it stay on mission for hours, remember constraints, and not "wander"?
- When it can do that reliably, it stops being a clever assistant and starts being something closer to a dependable worker.
Frequently Asked Questions (FAQ)
❓ Is GPT-5.4 officially confirmed by OpenAI?
No. OpenAI has not officially announced GPT-5.4. The evidence comes from leaked code in GitHub, scrubbed references, an employee screenshot, a journalist's error log, and a reporter at The Information confirming internal use — none of it is an official announcement.
❓ What is the expected context window for GPT-5.4?
Early rumors mentioned 2 million tokens, but more grounded reporting points to 1 million tokens — matching top competitors and roughly 2.5x the current GPT-5.2 window of 400,000 tokens.
❓ What is "extreme thinking mode" in GPT-5.4?
It refers to a ladder of reasoning effort: light, medium, high, extra high, and "extreme." At the extreme end, the model may spend significantly more inference-time compute — potentially hours — on very difficult tasks, rather than seconds or minutes.
❓ What is the QuitGPT movement?
QuitGPT is a user backlash movement tied to OpenAI's partnership with the Department of Defense/Pentagon. People are canceling ChatGPT subscriptions out of ethical objections, frustration, or both — and switching to alternatives like Anthropic's Claude.
❓ When will GPT-5.4 be released?
No official release date has been announced. Given OpenAI's current near-monthly shipping cadence and the fact that the model appears to already be running internally, a release could happen relatively soon — but speculation about exact timing remains just that: speculation.
Conclusion
The GPT-5.4 leaks don't look like random noise — they look like a model that's already being tested and prepared for release. If the 1 million token context window, "extreme thinking," full-resolution images, and speed tiers land as described, the biggest win won't be bragging rights. It'll be reliability over long tasks.
Meanwhile, the QuitGPT backlash keeps pressure on OpenAI to ship improvements faster and explain itself better. If GPT-5.4 drops soon, the response will tell us a lot — not just about the model, but about what users actually care about right now.
🔵 Revolution In AI Verdict
GPT-5.4 is real — the engineering breadcrumbs confirm it. Watch for the long-task reliability story more than any single benchmark number. That's where the next battle in AI is actually being fought.
📚 More from Revolution In AI
0 Comments