GPT-5.6 Lands as an Agent-First Model, Codex Gets Absorbed, and Claude Resets Quotas
So lively! GPT 5.6 is here, Codex is gone, Claude has reset!
As the title says, I woke up to find the foreign AI scene absolutely buzzing. There are so many hot topics, I don't even know which one to chase!
To sum it up in one sentence: GPT 5.6 has arrived, Codex is no more, Claude has reset, and Zuckerberg actually tweeted, to promote their new model Muse!
With this update pace, it's impossible to keep up!
I was originally peacefully testing Fable 5, thinking once my tokens ran out, I could rest. I didn't expect another reset today.
One day in AI is a year in the human world; it changes every day. After today, how should we choose a model? Honestly, model testing is incredibly difficult right now! You can only differentiate them by diving deep into specific scenarios.
Today, let's just watch the excitement first, and at the end, I'll give everyone a reference table for choosing models in different scenarios ⌚️!
1. Client-side Changes
Let's first look at the direct part, the changes happening in the software in your hands.
Because of the Fable 5 situation, GPT 5.6 was also under review for a while. It's finally here today. I wasn't really looking forward to the 5.6 update myself, because it's just a +0.1 update, not very interesting. But this time, OpenAI's adjustments are quite significant!
When I woke up this morning, a 'Work' tab appeared in ChatGPT:
And the 5.6 model can now be selected, but only in this mode!
Under Chat mode, there's only GPT 5.5!
Seemingly a small change, but it's actually very significant for OpenAI. Their ChatGPT previously had 800 million users, but most were using the Chat function. Now they've finally inserted the Work function into the core application.
Not only has the ChatGPT client changed, but Codex has too. Actually, the Codex application no longer exists!
Its name is now ChatGPT (Codex)!
For Codex users, the changes aren't huge; the core functionality is the same!
Just some adjustments to the left menu. The current top-level menu items are: Usage, Pets, Invite Friends, Settings! OpenAI is now the only foreign AI model tool with an 'Invite Friends' feature! OpenAI really wants to dominate the mass market. (Want to give me an email address? I'll invite you, and we can share experiences! Haha)
These foreign "trash" products, I praise them every day, and they don't even give me a commission!
With all this tinkering, OpenAI has created a "Two Suns in the Sky" effect!
Now there are two identical icons on the computer. You can only distinguish them by order, or you can uninstall one. Which one is better to uninstall?
2. Characteristics of GPT-5.6
Based on my experience, GPT 5.6 probably won't significantly improve its programming ability, but rather its agent capabilities. There probably isn't much room to test its front-end either.
I'll just use its Work function to summarize the characteristics of GPT 5.6. Consider it a small agent test!
I chose the GPT 5.6 Sol High mode:
It mentioned:
GPT‑5.6 is a new generation model family officially released by OpenAI on July 9, 2026. It's not just about improving chat ability, but is clearly developing towards a "long-duration autonomous agent model."
I've heard this kind of description way too many times, haha!
Three versions:
| Model | Positioning | API Input/Output Price |
|---|---|---|
| GPT‑5.6 Sol | Flagship, complex programming, research, professional work | $5 / $30 |
| GPT‑5.6 Terra | Performance-cost balance | $2.5 / $15 |
| GPT‑5.6 Luna | Fastest, cheapest, suitable for high concurrency | $1 / $6 |
Looking at this table gives a very familiar feeling! I silently recited: Opus, Sonnet, Haiku!
Core characteristics:
Strong Programming Ability
Sol is currently OpenAI's strongest programming model:
- Terminal-Bench 2.1: 88.8%
- Ultra multi-agent mode: 91.9%
- DeepSWE: 72.7%
- SWE-Bench Pro: 64.6%
Its advantage leans more towards terminal operations, complex engineering, tool calling, and long-duration continuous execution, rather than just generating single code snippets!
Native Multi-Agent Capability
New ultra mode, coordinating four agents to work in parallel by default, and finally unifying the results.
Suitable for:
- Large project analysis
- Multi-module parallel development
- Complex research tasks
- Security audits
- Simultaneous searching, coding, testing, and verification
max is a single model investing more reasoning time; ultra is multiple agents executing in parallel.
Smarter Tool Calling
GPT‑5.6 supports Programmatic Tool Calling, allowing you to write a piece of JavaScript to organize multiple tool calls, execute loops, parallel requests, and filter intermediate results.
This can significantly reduce the round trips of "model calls tool once → returns result → thinks again → calls again," especially suitable for agents and automation systems.
Very Large Context
- Context window: approximately 1.05 million tokens
- Max output: 128K tokens
- Knowledge cutoff: February 16, 2026
- Supports text and image input, text output
In practice, it can read a large code repository, hundreds of documents, or several long novels at once. But official long-context tests show that Sol's accuracy in the 512K to 1M range still drops to about 73.8%, so "being able to fit it in" doesn't equal "completely without omissions."
Significantly Enhanced Frontend and Design Capabilities
GPT‑5.6 doesn't just generate web page code; it also checks rendering results, adjusts layout, hierarchy, fonts, spacing, and interactions. OpenAI particularly emphasizes its design judgment on frontends, PPTs, documents, and spreadsheets.
This is quite valuable for making mini-games, product prototypes, visual materials for public accounts, and complete web projects.
Higher Token Efficiency
Sol's unit price is basically the same as GPT‑5.5, but completing tasks usually requires fewer output tokens, fewer tool calls, and fewer steps. That is, it's not a lower unit price, but the total cost to complete the same thing may be lower.
Enhanced Caching Mechanism
Supports explicitly specifying cache breakpoints, cache retained for at least 30 minutes:
- Cache write: 1.25x the normal input price
- Cache read: 90% reduction in input price
Quite practical for long system prompts, large codebases, and repetitive agent tasks.
This introduction should be fairly clear and concise! It's also in a state of wanting it all, optimizing for different scenarios, then greatly improving stability, efficiency, and optimizing costs.
Overall, GPT is a very comprehensive AI assistant.
3. Claude Resets Quotas!
OpenAI has started unabashedly borrowing and imitating Claude's categorization method and their Cowork!
Claude also didn't hesitate to borrow OpenAI's marketing method: Resetting Quotas!
Gods fight, we watch the show! Comrade Tibe is also watching the show with us. He expressed that he smelled: a whiff of fear!
Tibe is the head of Codex & ChatGPT, very active on X, and he's the one shouting about resets every day. Now seeing ClaudeDev also playing this way, he went to mock them a bit.
Great!
You can only say "competition" is an excellent thing for users!
We can freeload thousands of dollars worth of Fable 5 again! We can use GPT 5.6 for free again (by free, I mean for subscribers, no extra payment needed!)
4. Meta Joins the Fun Too
Meta's previous Llama had a profound impact on open-source models, but it drained itself dry and grew weaker day by day! Later they made closed-source models, but it seems no one even mentions them anymore. But these past few days, they suddenly dropped a new model.
Zuckerberg's X account, which hadn't been updated for over a decade, was updated too. So rare:
Zuckerberg first registered on Twitter in 2009 (we're from the same wave, ah life~~ the gap~~)!
After posting some very abstract content, he stopped posting. Three years later, he posted a "traffic-driving post" and then disappeared for 11 years. Then in 2023, he posted a picture:
The meme of this picture is probably about Threads and X!
After posting this, he disappeared again, and this reappearance is this time, and he posted several in one go!
It's truly very rare. As a social media giant, to promote their own AI model, they posted update notices on a rival's platform. Zuck must have been holding his breath for so long not tweeting, and now that breath has dissipated!
Zuck introduced their new model like this:
Muse Spark 1.1 performs strongest in agent performance, tool use, and computer use. It excels on long-running tasks with a 1 million token context window, can delegate execution tasks to sub-agents running in parallel, and is trained to use computer interfaces on desktop, mobile, or browser.
I told you, my ears have calluses from hearing these descriptions! It's almost the unified expression for all model updates!
Let's look at the specific benchmark data:
It doesn't lead in Coding, but has some highlights in the Agent section!
It mainly surpasses Opus 4.8 and GPT 5.5 on four benchmarks. I made a picture for easier understanding.
Of course, this table of theirs doesn't include the two strongest models, GPT 5.6 Sol and Fable 5!
Meta is probably at this level. They've even started considering renting out GPUs.
5. Line Up and See Who's Tallest!
With several models updated recently, I think it's necessary to sort out which models are good at what. Since I haven't tested them all yet, let's just compare benchmark data first. Foreign models might also game benchmarks, but overall, credibility is still acceptable.
| Scenario | Benchmark | Main Test Content | Sol | Fable 5 | Opus 4.8 | Muse 1.1 | Grok 4.5 |
|---|---|---|---|---|---|---|---|
| 🧠 General | Artificial Analysis Intelligence Index | Reasoning, science, programming, agent general | 58.9 | 🥇 59.9 | 55.7 | — | 54 |
| 📊 Professional Work | GDPval-AA v2 | Documents, spreadsheets, PPTs, etc. professional deliverables | 1747.8 | 🥇 1759.6 | 1600.1 | 1381 | 1543 |
| 💻 Coding Agent | Coding Agent Index | Code modification, terminal, codebase understanding | 🥇 80.0 | 77.2 | 72.5 | — | 76 |
| 🐛 Bug Fixing | SWE-Bench Pro | Locate and fix bugs in large real repositories | 64.6% | 🥇 80.0% | 69.2% | 61.5% | 64.7% |
| 🏗️ Long-Range Dev | DeepSWE v1.1 | Complete real software engineering tasks over long periods | 🥇 72.7% | 69.7% | 59.0% | 53.3% | 53.0% |
| ⌨️ Terminal Ops | Terminal-Bench 2.1 | Complete complex multi-step tasks via command line | 🥇 88.8% | 83.1% | 78.9% | 80.0% | 83.3% |
| 🖥️ Computer Use | OSWorld 2.0 | Operate real desktop software to complete long workflows | 🥇 62.6% | — | 54.8% | 47.3% | — |
| 🛠️ Tool Calling | Toolathlon | Call a large number of real tools to complete complex tasks | 58.0% | 🥇 61.7% | 59.9% | — | — |
| 🔬 Science Reasoning | GPQA Diamond | Graduate-level scientific knowledge and reasoning | 🥇 94.6% | 92.6% | 92.0% | — | — |
| 🧮 Extreme Math | FrontierMath Tier 4 | Frontier-level difficult math problems | 83.0% | 🥇 87.8% | 56.1% | — | — |
| 🩺 Medical Professional | HealthBench Professional | Professional medical analysis and answer quality | 60.5% | 🥇 60.9% | 53.0% | 59.3% | — |
| 📄 PDF Understanding | gdp.pdf | Text, chart, and data understanding in complex PDFs | 🥇 30.7% | 29.8% | 22.5% | — | — |
This TOP list is basically dominated by Fable and Sol.
Of course, in actual use, we might not necessarily choose these top-tier models. Partly because 'good enough' is enough, and partly due to price factors.
Of course, there are also many people like me who just like having the strongest.
I also made a dedicated picture for easier understanding and selection!
This is a summary and picture made by GPT 5.6. The data is all accurate, but the specific descriptions might be biased towards its own model!
In the final summary, I quite agree: Sol is strong in execution, Fable is strong in tackling tough problems, Grok is strong in cost-effectiveness!
That's about it!
OpenAI and Anthropic are the two insurmountable mountains of the AI model industry. No one else can climb over; other companies can only carve out a corner for themselves!