GPT-5.6, to be released next week?
Just yesterday, netizens excitedly discovered that the underlying code of the Codex application surprisingly contained identifiers for GPT-5.6's three sub-models: Sol, Terra, and Luna.
Even more exciting, a brand-new "Speed Dial" feature also appeared in the code.
This hints that users will be able to freely adjust between speed and quality based on their needs, undoubtedly bringing an unprecedented level of control.
According to leaks, OpenAI has internally set a hard deadline: the target release window for GPT-5.6 points directly to next Tuesday (July 7th) through July 9th.
Why July 7th? Because this day coincides precisely with the vacuum period after the expiration of Claude Fable 5's specific quota plan.
This is a meticulously timed commercial hunt, calculated down to the hour.
Recently, Anthropic has driven countless developers crazy with a series of missteps, and Google's Gemini 3.5 Pro was forced into an emergency "rework." OpenAI is seizing this moment to strike and scoop up users!
Dissecting Codex Code
Sol, Terra, Luna Are All Coming
"To be honest, OpenAI is acting like it's no big deal, quietly stuffing the model names into dead code as if we wouldn't notice," one netizen joked.
Since GPT-5.6's limited release, geeks have been closely watching every front-end update from OpenAI.
Finally, in a recent merge of the Codex application, someone spotted traces of GPT-5.6.
Another user shared a short video. Although backend API limitations currently prevent successful model calls, the front-end popup clearly shows the styles of the three models and the new "Speed Selector."
Furthermore, the code faintly revealed the phrase "Sol Ultra." Industry speculation suggests Sol Ultra will be the ace card directly competing with top-tier flagship rivals, matching Fable 5's performance but at a much more affordable price.
Besides these three models, the code also revealed a key piece of information: the highly anticipated "Real-time Voice Support" is still under development and likely won't launch directly next week.
Leaked Hands-on Tests: GPT-5.6 vs. Fable 5
Although most haven't gotten access yet, a few players with internal testing permissions have already shared comparative evaluations of GPT-5.6 in real engineering environments.
The result can be summed up in four words—a dimensional strike.
Round One: The Extreme Tug-of-War Between Efficiency and Understanding
Overseas tech blogger Shivam shared his experience using GPT-5.6-terra and Fable 5 to solve the same complex technical prompt.
Fable-5 started with a 100% 5-hour session limit. This model frantically "Thought" in the background, burning through a whopping 21% of the quota limit, only to finally respond by asking a bunch of cross-questions, telling him to reconfirm the technical details to solve.
For the same task, GPT-5.6-terra consumed only 13% of the quota, with astonishingly fast response speed.
It didn't waste words but efficiently listed several different methods and architectural paths to solve the problem and quickly began execution.
Shivam stated bluntly: While using Fable, my mind was preoccupied wondering if it would suddenly downgrade to Opus 4.8; the decisiveness of GPT-5.6-terra was extremely satisfying.
Round Two: "Blind Test" of a Hardcore WebGL Frontend Mini-Game
Oracle Director Gilson Melo presented a tough challenge to GPT-5.6 High and Fable 5 High:
Build a fully functional browser-based game from scratch in a single HTML file (using WebGL or HTML5 Canvas). The game must have real-time rigid body physics, gravity, and user-controlled paddle/ship mechanics. Write the complete CSS, JS, and HTML, omit no logic, and must support mouse drag-and-drop with real-time physical feedback.
This question severely tests the model's ability to handle extreme detail, long code without reduction, and calculations of underlying physics formulas.
The two models exhibited different strategies in their workflows.
Fable 5 High's performance was stunningly impressive; it went all-in with extreme confidence, generating the entire game's code in one go.
GPT-5.6 High, however, very humanely paused twice during generation, actively asking the developer to clarify two final key decisions.
Even more impressive, without being asked, it took the initiative to add sound effects to the game.
The final results showed that GPT-5.6 High scored more solidly in overall game experience, smoothness of physical collisions, and robustness of details.
In summary, both testers believed GPT-5.6 had the upper hand in efficiency and response style, especially in clarity and speed when handling complex tasks.
Based on these results, it's very much worth looking forward to GPT-5.6's launch next week.
Precise Timing
OpenAI Poaches Users While They're Vulnerable
If the model leak was an accident, the release timing is definitely a深思熟虑的布局 (deliberate layout).
OpenAI plans to重磅发布 (heavily release) GPT-5.6 on July 7th, precisely卡在 (timed to) the day Claude users lose access to Fable 5.
Claude has been losing many users recently. OpenAI sees its chance and is ready to scoop them all up.
An insider revealed: "The usage quota limits for GPT-5.6 will be significantly relaxed, more generous than Fable 5's. Tighter safety guardrails are also being rolled out gradually, but they won't be as aggressive as Fable's to the point of affecting normal use."
User Dissatisfaction Peaks, Perfect Timing for OpenAI's Poaching
In comparison, Anthropic has been facing a lot of public anger lately.
Claude Fable 5, despite just returning, has already triggered strong user dissatisfaction.
Ask just a few questions, and Fable 5 will downgrade to Opus 4.8.
Biomedical engineer Derya Unutmaz tried to get Fable 5 to explain the word "human."
After typing just "Explain human," the model thought for a few seconds before popping up a card saying "Switched to Opus 4.8," because Fable 5's safety mechanism judged this message contained content needing interception.
Even more ridiculous, semiconductor analyst Dylan Patel asked an extremely simple question: "How many letter r's are in the word 'raspberry'?"
This question was also intercepted, with the interface showing "Chat paused," indicating Fable 5's safety mechanism blocks most cybersecurity or biology-related topics.
Additionally, Opus 4.8's recent hallucination problems are also very severe, with conversations even containing others' information.
This cliff-like drop in user experience creates the perfect window for OpenAI to poach users.
Moreover, GPT-5.6 is also likely to be more cost-effective.
Leaks show GPT-5.6 Sol will be more than twice cheaper than Fable 5, due to its higher token efficiency. But the key is, is its performance sufficient to rival Fable 5?
Some predict Sol Ultra should be comparable to Fable 5 while being cheaper. If this prediction holds true, OpenAI will completely outperform its competitor in price-performance ratio.
Developer Reminder
4 Codex Reset Credits, Don't Let Them Go to Waste
Finally, a "tips for maximizing benefits/avoiding pitfalls" guide for all hardcore developers planning to return to Codex.
According to in-depth digging by Reflection's CTO, if you previously accumulated 4 rate-limit reset credits in Codex, immediately check your account backend.
OpenAI's official underlying rules show these reset credits have a validity period of only 30 days. If your first credit arrived around June 11th or 12th, then around July 12th, they will start expiring in batches!
If you want to know your precise expiration time, you can have Codex call your ChatGPT token to request this backend API: GET https://chatgpt.com/backend-api/wham/rate-limit-reset-credits.
You will receive a JSON response similar to the following:
If GPT-5.6 is truly unlocked on schedule next Tuesday, you'll have only a brief 4 to 5 days to use up your first reset credit.
Next Tuesday, OpenAI will most likely gift everyone a brand-new Reset. So, hurry up and use your old credits wisely these next few days.
See you next week, GPT-5.6!
References:
https://x.com/testingcatalog/status/2073049917266821338https://x.com/synthwavedd/status/2073084352251232435
https://x.com/ShivamS1123/status/2072664629445275897
https://x.com/gmelo33/status/2072822933194437035
Editor: Aeneas
This article is from the WeChat public account "New Zhiyuan" by ASI启示录
