GPT-5.6 Lands With Three Models, but Early Benchmarks and Hands-On Testing Tell a Messier Story

Today, the most explosive news in the tech world is undoubtedly the triple launch of GPT-5.6 — Sol, Terra, and Luna going live simultaneously.

After the launch, all subscriptions received a quota reset.

This is the most important version of the second half of this year, and the last product before the GPT 6.x era. I estimate 6.x will be online before long.

Regarding GPT-5.6's capabilities, I posted about them the day before yesterday; you can check my previous article.

Just update the Codex version directly.

Sol is incredibly strong, outright beating Fable 5.

But GPT-5.6 - Sol is only available on Pro 5x and 20x plans; Plus accounts only get Terra and Luna.

However, I'm not sure why — it was previously stated that Plus could use Sol. This time, I don't know if it hasn't been rolled out yet or what, but it's different from what was said at launch. If anyone can use it, please let us know.

In addition, Sol has added Max and Ultra thinking quotas. Both Max and Ultra have very good reasoning capabilities. Max makes the model spend more time thinking deeply, while Ultra calls multiple sub-agents to process tasks in parallel. Ultra also consumes significantly more tokens, so use it with caution.

OpenAI also released a benchmark. Looking at it this way, GPT-5.6 Sol not only does good work but also saves tokens. Fable 5... forget it, I won't bother.

And OpenAI specifically emphasized that the Luna version was used for Sol's post-training.

Although GPT 5.6 has been officially released, it's time to say goodbye to Codex.

OpenAI has officially launched ChatGPT Codex and ChatGPT Work.

The current Codex is now called ChatGPT Codex.

These two modes can be seamlessly switched directly within Codex.

The launch of ChatGPT Work feels like it could be OpenAI's entry point into the office productivity space.

Unlike Codex for developers, which shows the thinking process, ChatGPT Work hides the technical details and is primarily aimed at non-developers.

ChatGPT Work is the most noteworthy new feature this time. It no longer just answers questions; it can also execute tasks across applications: connecting to your Google Drive, Slack, email, and other tools, breaking down complex projects on its own, completing them step by step independently, and only asking you when necessary.

For example, if you give it a sentence like "organize the customer research into a marketing plan," it can read the materials, write the plan, and generate material versions for different markets, all while maintaining context throughout.

This is also what was previously mentioned: ChatGPT and Codex might be integrated together.

Work will also be available on web and mobile, meaning you can initiate and track tasks without being in front of a computer.

Many products like Qoder, Trae, and WorkBuddy (domestic products) actually have two versions: one for programmers and one for office workers.

It's just that this time OpenAI has integrated them together.

The browser has also received a major update this time.

Using the Computer Use feature within Codex is smoother, and it now supports authenticated websites, multiple tabs, and file downloads. Additionally, your tabs can be preserved, making it even smoother.

This Computer Use feature allows your ChatGPT to use any app, increasingly resembling human usage. Most crucially, the usage speed is faster and it saves more tokens.

I used the previous Computer Use feature frequently in my daily routine, and I did feel the efficiency was somewhat slow.

Beyond that, GPT 5.6 has made substantial improvements in front-end, visualization, and other areas. OpenAI also released a front-end rendering comparison chart for different generation types this time.

There are too many small improvements.

This update has brought everyone quite a few new things. I think these small details and features will keep me excited for a while.

Oh right, after GPT 5.6 was officially released, OpenAI chose to reset the quotas for all users. Let that sink in.

The triple launch of GPT 5.6 naturally puts everyone in a euphoric mood, and perhaps they can't stand to hear any dissenting opinions.

But I might have to say something different. I might have to pour some cold water on this.

First, let me post a picture from my favorite evaluation organization.

This picture essentially says: the contest between Fable 5 and GPT-5.6 needs to be viewed in two ways: comprehensive capability and coding agent capability.

• The upper chart is the comprehensive Intelligence Index (a mix of 9 different tests): Fable 5 scored 60, GPT-5.6 Sol scored 59, Terra scored 55. This means Fable slightly leads Sol in the composite score, and Terra is a tier lower.
• The lower chart is the Coding Agent Index (the sum of DeepSWE, Terminal-Bench, SWE-Atlas-QnA): Sol scored 80, Terra scored 77, Fable 5 scored 77. This means Sol is first when acting as a coding agent; Terra and Fable are basically tied.

So it cannot be concluded that GPT-5.6 comprehensively beats Fable. Instead, it shows:

Sol is the top performer in the coding agent track; Fable still holds a slight lead in comprehensive capability; Terra's positioning is very much like a coding agent, with scores close to Fable, but its overall capability is a tier lower, making it a mid-range model.

Additionally, max is the highest reasoning tier, and "with fallback" indicates Fable's safe fallback configuration is also running; so this is not a pure, naked model showdown. Different toolchains, reasoning tiers, and evaluation sets will all change the rankings.

And Simon Willison mentioned in his blog that GPT 5.6... might not be as effective as Fable 5.

Simon Willison was fortunate to be among the group that got early access, so he had enough testing time. He explained the ins and outs clearly.

He said they designed an "Agent Ultimate Exam," aimed at making every token count more.

This test covered 55 long-running professional workflow evaluations. GPT-5.6 Sol set a new high score of 53.6, beating Claude Fable 5 (using adaptive reasoning mode) by 13.1 points.

Even in medium mode, it led Fable 5 by 11.4 points at about a quarter of the estimated cost.

This is crucial for achieving more widespread and economical intelligence: GPT-5.6 Terra and GPT-5.6 Luna both outperformed Fable 5, at only about one-sixteenth of the cost.

However, in the highly watched SWE-Bench domain, Fable 5 significantly beat GPT-5.6. Fable 5 achieved a score of 80%, while GPT-5.6 Sol scored 64.6%.

Of course, OpenAI published a dedicated article the next day, stating that about 30% of the questions in SWE-Bench Pro are inherently problematic, possibly having incomplete descriptions, overly strict tests, or non-unique answers, and that this scorecard shouldn't be treated as the final judge.

He said that based on his usage so far, for complex coding tasks executed by OpenAI's models, its performance has not been better than Fable's.

He showcased 18 different "pelican" configurations for the different versions of GPT 5.6 — covering the performance of three different models at various reasoning intensities. He also listed their token consumption and computational costs: the cheapest was gpt-5.6-luna (reasoning intensity: None), costing 0.71 cents; the most expensive was gpt-5.6-sol (reasoning intensity: Max), costing 48.55 cents.

Image source: Simon Willison. The three columns are Luna, Terra, and Sol; the six rows are none to max.

I ran a few tests myself. I didn't get to use Sol; I tested the Terra ultra mode. My impression is that it's not as good as GPT 5.5. The output speed is very slow, and I encountered two problems right from the start.

The first round, it sent an empty final message at the end of the task, causing the interface to appear blank.

Then the second round, it mixed English and Chinese in the output.

So, who is stronger? I don't think there's a need to jump to a conclusion so quickly.

Perhaps it's because I didn't get to use Sol mode, so maybe testing with Terra doesn't reveal anything at all.

GPT-5.6 Sol might have the upper hand in token efficiency, multi-agent collaboration, and tool invocation; Fable 5, at least in the complex engineering tasks Simon actually performed, remains in an irreplaceable position.

I tested Grok 4.5 yesterday and felt it was pretty good. If my later tests with GPT-5.6 - Sol also turn out poorly, I might just switch to Grok 4.5.

So, what's the verdict? It depends on the actual test results from all you smart people here.

Finally, a digression. Many reviewers probably stayed up all night waiting for the launch, wrote their articles, and went to sleep. Good for them.

As for me, I got up at 4 AM to watch the France vs. Morocco match, and then casually wrote this article.

So, am I going to sleep now? No, I'm going to work. :)