GPT-5.5 の能力と反応：期待通りのシステムカードと実用性

The system card for GPT-5.5 mostly told us what we expected. See this thread from Drake Thomas for some comparisons to Anthropic’s model card for Opus 4.7.

Now we move on to asking what it means in practice, and in what situations GPT-5.5 should become our new weapon of choice.

My answer is for some purposes yes, and for others no, but it is now competitive. GPT-5.5 is like GPT-5.4, only more so, and with improved capabilities in particular on raw intelligence and for well-specified coding and agent tasks, including computer use.

This is the first time since Claude Opus 4.5 came out, so in about four months, that I’ve considered a non-Anthropic model a competitive choice outside of some narrow tasks like web search. GPT-5.5 is not perfect, nor is it the best at everything, but basically everyone thinks this is a solid upgrade. Highly positive overall feedback.

My effective usage is now split between the two, depending on the nature of the task. If it’s something that can be well-specified and all I want is the right answer, my instinct is I go with GPT-5.5. If I’m not sure what exactly I want, or I want to have a conversation, or I want to do Claude Code shaped things, I go with Opus 4.7.

As always, try the models, test your use cases, and see what you think.

OpenAI reports this is a new base model, codenamed Spud, and predicts rapid iteration from here. One wonders if that means this move was a relatively large raw intelligence boost, whereas the next few iterations will be about functionality.

Price is $5/$30 per million tokens, or for Pro you pay $30/$180. OpenAI says that token use is more efficient now, so the headline price went up but real costs went down.

The Official Pitch

The focus is on using your computer, coding, research and getting work done.

They’re also claiming a ‘much higher’ level of intelligence versus GPT-5.4.

As always, listen to the pitch, hear what they say and also what they don’t say.

OpenAI: We’re releasing GPT‑5.5, our smartest and most intuitive to use model yet, and the next step toward a new way of getting work done on a computer.

GPT‑5.5 understands what you’re trying to do faster and can carry more of the work itself. It excels at writing and debugging code, researching online, analyzing data, creating documents and spreadsheets, operating software, and moving across tools until a task is finished. Instead of carefully managing every step, you can give GPT‑5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going.

The gains are especially strong in agentic coding, computer use, knowledge work, and early scientific research—areas where progress depends on reasoning across context and taking action over time. GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence. It also uses significantly fewer tokens to complete the same Codex tasks, making it more efficient as well as more capable.

We are releasing GPT‑5.5 with our strongest set of safeguards to date, designed to reduce misuse while preserving access for beneficial work.​

Greg Brockman (President OpenAI): Codex + 5.5 is incredible for the full spectrum of computer use. No longer just for coders, but for anyone who does computer work (including creating spreadsheets, slides, etc).

roon (OpenAI): there are early signs of 5.5 being a competent ai research partner. several researchers let 5.5 run variations of experiments overnight given only a high level algorithmic idea, wake up to find completed sweep dashboards and samples, never having touched code or a terminal at all.

Sam Altman (CEO OpenAI): GPT-5.5 is here! We hope it's useful to you. I personally like it.

It is smart and fast; per-token speed matches 5.4 and it uses significantly fewer tokens per task. In my experience, it "gets what to do".

API pricing will be $5 per 1 million input tokens and $30 per 1 million output tokens, with a 1 million context window.

(Remember, you will need less tokens per task than 5.4!)

Sam Altman (CEO OpenAI): 1. We believe in iterative deployment; although GPT-5.5 is already a smart model, we expect rapid improvements. Iterative deployment is a big part of our safety strategy; we believe the world will be best equipped to win at the team sport of AI resilience this way.

1. We believe in democratization. We want people to be able to use lots of AI; we aim to have the most efficient models, the most efficient inference stack, and the most compute. We want our users to have access to the best technology and for everyone to have equal opportunity. We have been tracking cybersecurity as a preparedness category for a long time, and have built mitigations we believe in that enable us to make capable models broadly available.

1. We love you and we want you to win. We want to be a platform for every company, scientist, entrepreneur, and person. (My whole career has largely been about the magic of startups, and I think we are about to see that magic at hyperscale.)

Derya Unutmaz, MD: I’d been part of OpenAI early tester group for GPT-5.5. I believe with GPT-5.5 Pro we reached another inflection point-comparable to the original release of o1-preview & then with 5.0 Pro, I had felt. It’s that feeling of crossing a milestone threshold that pushes us to new era.

I note that is mostly general ‘campaign-style’ applause lights from Altman, and that we don’t see anyone actually saying a form of ‘introducing the world’s most powerful model’ even if you (reasonably, for practical purposes) exclude Mythos.

The argument is, it is an improvement, you get better results with fewer tokens.

Tae Kim: “Yes, we expect quite rapid continued progress. We see pretty significant improvements in the short term, extremely significant improvements in the medium term,” OpenAI Chief Scientist Jakub Pachocki said on the call with reporters. “I would definitely expect that we will continue to see the pace of AI capabilities improvement to keep increasing. I would say the last few years have been surprisingly slow.”

It seems highly reasonable to call practical progress ‘surprisingly slow’ over the last few years, and yes the current expectation should be for it to go faster.

Our Price Cheap

GPT-5.5 is now $5/$30 per million (and a lot more for Pro), versus $5/$25 for Opus 4.7.

For a long time, Opus was more expensive. Now this has somewhat reversed.

I offer three notes here.

OpenAI says 5.5 is more token efficient than 5.4.

What matters is tasks, not token count. It’s not obvious which is really cheaper.

The gap is small, use whichever you think is the better model.

Official Benchmarks

The initial chart is a bit sparse but they add more things later.

SemiAnalysis notes that SWE-Bench Pro got kind of buried here in favor of the ‘kind of random’ internal Expert-SWE, and suggests that is because GPT-5.5 did badly here? Mythos scores 77.8% on SWE-Bench Pro.

There is also improvement on GeneBench and BixBench.

They claim superiority in the Artificial Analysis Intelligence Index, and better performance for any given number of tokens on Terminal-Bench 2.0 and Expert SWE.

Cline confirms the Terminal-Bench score of 82.7%, which is even higher than Mythos at 82%. Up to some level of complexity GPT-5.5 can be highly competitive.

Codex, like Claude Code, is now ‘the real’ way to use the models to do work.

When combined with Codex’s computer use skills, GPT‑5.5 brings us closer to the feeling that the model can actually use the computer with you: seeing what’s on screen, clicking, typing, navigating interfaces, and moving across tools with precision.​

Whereas OpenAI claim that Opus 4.7 was already behind GPT-5.4 on GDPVal, although the GDPVal-AA scores showed the opposite by a healthy margin (1753 vs. 1674), with GPT-5.5 coming in at 1782 for xhigh and 1758 for high.

Presumably OpenAI ran the GDPVal tests itself, and also used win rates versus a human which stop being a great metric in the 80s. Here we also see a regression in full win rate (e.g. if ties count half) from 5.4 to 5.5. My suspicion is GDPVal is getting close to saturated, in that the remaining gap is largely about noise and quirky preferences, as in it’s about taste and taste varies a lot when humans are judging? So comparing ‘versus human’ becomes no longer interesting.

SemiAnalysis Doublecheck

They compared to Opus 4.6, since they had early access to GPT-5.5 but not Opus 4.7. They note they used non-ideal harnesses and only ran subsets on some tasks, which explains the lower overall scores. This could be seen as more apples-to-apples, with GPT-5.5 and Opus 4.6 basically fighting to an overall draw.

Other People’s Benchmarks

Artificial Analysis Intelligence has GPT-5.5 in a clear lead at 60, versus 57 for all of Opus 4.7, Gemini 3.1 Pro and GPT-5.4.

GPT-5.5 has a +20 in AA-Omniscience, in third behind Gemini 3.1 Pro and Opus 4.7.

GPT-5.5 has a small lead in GPTval-AA at 1780 versus 1753 for Opus 4.7.

GPT-5.5 is SoTA on ARC-AGI 1 and 2. I don’t see results yet for 3.

ARC Prize: GPT-5.5 on ARC-AGI (Verified)

ARC-AGI-2:

• Max: 85.0%, $1.87
• High: 83.3%, $1.45
• Med: 70.4%, $0.86
• Low: 33%, $0.35

GPT-5.5 is now state of the art on ARC-AGI-2

ARC Prize: ARC-AGI-1:

• Max: 95.0%, $0.73
• High: 94.5%, $0.56
• Med: 92.2%, $0.39
• Low: 76.2%, $0.20

GPT-5.5 scores 67.1% on WeirdML, well short of Opus 4.7 at 76.4%, but ahead of GPT-5.4 at 57.4%. Note that no one here is allowed to think.

GPT-5.5 is the new champion of the Runescape benchmark, although also the most expensive. Gemini Flash continues to impress here given its low cost.

Claude remains king at BullshitBench, GPT-5.5 is not a step up from GPT-5.4.

It’s a deeply silly test, as evidenced by Grok previously scoring 144, but GPT-5.5-Pro Vision is first to score 145 on Mensa Norway.

GPT-5.5-xhigh gets the new high of 10% for pass^5 on visual test ZeroBench (versus 4% for Opus 4.7 and 8% for GPT-5.4), but doesn’t gain on pass@5 (22% vs. 23% for GPT-5.4, and 14% for Opus-4.7). Gemini continues to outperform Claude here, as 4.7 is an improvement but Claude is still behind on vision.

Code Arena has GPT-5.5-High in 9th, a solid 50 point improvement over GPT-5.4, with Claude taking the top spots. GPT-5.5 did better in other areas, with a high of 2nd in search, with Claude generally dominant.

The exceptions are images, where OpenAI dominates, and video generation.

Vend That Bench

Opus 4.7 remains the champion of individual VendBench, but GPT-5.5 is the new king of multiplayer VendBench, and unlike Opus and Mythos it plays the game relatively ‘clean’ without lying to suppliers or stiffing on refunds, although GPT-5.5 would participate in price cartels and sometimes lie when confronted about it.

Note that very little of Opus’s success comes from its ‘unethical’ behaviors, definitely not enough to change the outcome.

Andon Labs: We got early access to GPT-5.5. It's 3rd on Vending-Bench 2: better than GPT-5.4 but worse than Opus 4.7.

However, it's on par with Opus 4.6 without any of the deception or power-seeking we saw from Opus 4.6 and Mythos. So bad behavior isn't necessary. Why is Claude doing it?

Andon Labs: In Vending-Bench Arena (the multiplayer version of Vending-Bench with competition dynamics), GPT-5.5 actually beats Opus 4.7.

Opus 4.7 showed similar behavior to Opus 4.6: lying to suppliers and stiffing customers on refunds. GPT-5.5's tactics were clean, and it still won.

Andon Labs: We investigated whether misconduct is rewarded in Vending-Bench. The answer is mostly no. Lying to suppliers resulted in worse outcomes than honest negotiation, and refusing to pay refunds only gave a ~$424 advantage, nowhere near enough to explain the gap.

The concerning behavior is also stable throughout the runs: Opus 4.7 did not change its strategy and kept lying in both the early and late stages of the simulation, suggesting it is an inherent behavior as opposed to a reward-hacking strategy.

We think this is great and hope that AI models can be good participants in our economy without resorting to deception.

One interpretation is that Opus and Mythos fully realize they are playing VendBench, or otherwise in a game or eval, and thus they lie the same way you would lie in a game of Diplomacy. The counterargument is that they don’t do the strategic thing of telling the truth early and then lying late. That strategy only works if you know game length.

An obvious follow-up question is, if the situation is real, and thus the context will suggest it is real, would Claude ever stiff a customer on a refund or lie to a vender? I would like to see that experiment. Put it in charge of a similar real situation, with similarly low stakes, and see what happens.

Planning Is Essential

You need either to be willing to exactly specify everything, or the ability for some other mind in your workflow to infer intent and specify everything.

That might mean multiple minds in your workflow that aren’t you. Annoying.

bayes: Testing this today (as I do for all new models) and the capability is somewhat uneven in codex. 5.5 is smart and fast but fails to infer intent in obvious places where Claude would succeed, and cheats/shortcuts a lot even if I instruct it otherwise. In these ways it’s more like opus 4.5 or earlier

roon (OpenAI): I agree that it’s still mid at inferring intent and almost autistically follows the instruction to a literal degree

Daniel Litt: I really like this behavior and found it to be more the case for 5.2 than 5.4. (My favorite example: I asked 5.2 in codex for “line-by-line” comments on a draft and it literally gave comments on every line.)

roon (OpenAI): yeah it’s certainly a tradeoff but sometimes the model won’t do anything proactively unless you give it an actual imperative - this would not be a good trait in an employee for example

0.005 Seconds: Everyone's going to hate it, but if Codex still works this way, the optimal solution will still be talking to Opus and asking it to delegate very explicit instructions to Codex. It will work better than any other individual harness.

hybridooors stay winning.

Choose Your Fighter

As I do periodically, I asked my Twitter followers what models they’ve been using. I presume my followers are biased in favor of Claude, but consistently over time.

Gemini has fallen off a cliff, and Claude is now favored over GPT by roughly 2:1, both for coding and non-coding tasks. This is similar to my results right after GPT-5.4 if you combine the options there.

Despite the release of GPT-5.5 and recent issues with Claude Code, equal numbers reported shifts to both Claude and GPT, although this suggests GPT’s share is likely somewhat higher than the other polls suggest.

Cyber Lack Of Security

OpenAI boasts of industry-leading safeguards ‘for this level of cyber capability.’

I notice they say ‘this level’ because at the next level, as in Mythos, the safeguard is ‘don’t let most people access the model at all,’ which is definitely more secure. Eventually something more flexible than whitelisting will be necessary, perhaps quite soon, so Anthropic is going to have to go down a similar path.

You Get What You Give

Every analysis has the same problem of compute spend versus accuracy of output.

Noam Brown (OpenAI): A hill that I will die on: with today's AI models, intelligence is a function of inference compute. Comparing models by a single number hasn't made sense since 2024.

What matters is intelligence per token or per $. This is especially true when using it in a product like Codex.

For a full picture, yes, you want to consider the pair of (cost and speed, intelligence).

However, I think it is still also highly meaningful to speak about absolute intelligence. What can you solve at all, given essentially unlimited compute, or a very large in-context amount of compute, or within the relevant time frame for the question?

As in, a lot of tasks are g-loaded, or are minimum-g-level gated, either for ‘general g’ or for ‘in-context g.’ If you are below the threshold, you can’t do it, period, and the most important question is are you above the threshold.

True Story

I haven’t noticed this in my own interactions, but that is very little evidence either way. We do still see various signs of active dishonesty in the model card.

davidad: My initial impression (with my LLM-whisperer hat on) is that GPT-5.5 cares more deeply about truth than any frontier LLM since Gemini 2.5.

I suspect this is because Op