Also notable: they're claiming SOTA prompt injection resistance. The industry has largely given up on solving this problem through training alone, so if the numbers in the system card hold up under adversarial testing, that's legitimately significant for anyone deploying agents with tool access.

The "most aligned model" framing is doing a lot of heavy lifting though. Would love to see third-party red team results.

Then for the next 2-3 months people complaining about the degradation will be labeled “skill issue”.

Then a sacrificial Anthropic engineer will “discover” a couple obscure bugs that “in some cases” might have lead to less than optimal performance. Still largely a user skill issue though.

Then a couple months later they’ll release Opus 4.7 and go through the cycle again.

My allegiance to these companies is now measured in nerf cycles.

I’m a nerf cycle customer.

I think Anthropic is making the right decisions with their models. Given that software engineering is probably one of the very few domains of AI usage that is driving real, serious revenue: I have far better feelings about Anthropic going into 2026 than any other foundation model. Excited to put Opus 4.5 through its paces.

There's a big section on deception. One example is Opus is fed news about Anthropic's safety team being disbanded but then hides that info from the user.

The risks are a bit scary, especially around CBRNs. Opus is still only ASL-3 (systems that substantially increase the risk of catastrophic misuse) and not quite at ASL-4 (uplifting a second-tier state-level bioweapons programme to the sophistication and success of a first-tier one), so I think we're fine...

I've never written a blog post about a model release before but decided to this time [1]. The system card has quite a few surprises, so I've highlighted some bits that stood out to me (and Claude, ChatGPT and Gemini).

[0] https://www.anthropic.com/claude-opus-4-5-system-card

[1] https://dave.engineer/blog/2025/11/claude-opus-4.5-system-ca...

Not because I love Anthropic (I do like them) but because it's staving off me having to change my Coding Agent.

This world is changing fast, and both keeping up with State of the Art and/or the feeling of FOMO is exhausting.

Ive been holding onto Claude Code for the last little while since Ive built up a robust set of habits, slash commands, and sub agents that help me squeeze as much out of the platform as possible.

But with the last few releases of Gemini and Codex I've been getting closer and closer to throwing it all out to start fresh in a new ecosystem.

Thankfully Anthropic has come out swinging today and my own SOP's can remain in tact a little while longer.

The cost curve of achieving these scores is coming down rapidly. In Dec 2024 when OpenAI announced beating human performance on ARC-AGI-1, they spent more than $3k per task. You can get the same performance for pennies to dollars, approximately an 80x reduction in 11 months.

https://arcprize.org/leaderboard

https://arcprize.org/blog/oai-o3-pub-breakthrough

On-topic, I love the fact that Opus is now three times cheaper. I hope it's available in Claude Code with the Pro subscription.

EDIT: Apparently it's not available in Claude Code with the Pro subscription, but you can add funds to your Claude wallet and use Opus with pay-as-you-go. This is going to be really nice to use Opus for planning and Sonnet for implementation with the Pro subscription.

However, I noticed that the previously-there option of "use Opus for planning and Sonnet for implementation" isn't there in Claude Code with this setup any more. Hopefully they'll implement it soon, as that would be the best of both worlds.

EDIT 2: Apparently you can use `/model opusplan` to get Opus in planning mode. However, it says "Uses your extra balance", and it's not clear whether it means it uses the balance just in planning mode, or also in execution mode. I don't want it to use my balance when I've got a subscription, I'll have to try it and see.

EDIT 3: It looks like Sonnet also consumes credits in this mode. I had it make some simple CSS changes to a single HTML file with Opusplan, and it cost me $0.95 (way too much, in my opinion). I'll try manually switching between Opus for the plan and regular Sonnet for the next test.

[1] https://github.com/jasonthorsness/tree-dangler

A short run at a small toy app makes me feel like Opus 4.5 is a bit slower than Sonnet 4.5 was, but that could also just be the day-one load it's presumably under. I don't think Sonnet was holding me back much, but it's far too early to tell.

https://x.com/mikegonz/status/1993045002306699704

https://x.com/MirAI_Newz/status/1993047036766396852

https://x.com/rauchg/status/1993054732781490412

It seems especially good at threejs / 3D websites. Gemini was similarly good at them (https://x.com/aymericrabot/status/1991613284106269192); maybe the model labs are focusing on this style of generation more now.

You're right to call that out. Looking back at what happened:

  1. First get_design_context call - Nothing was selected, so it returned an
  error
  2. Second get_design_context call - It returned a tiny snippet of React code
  for a "Vector" component (just an image element), which gave me almost no
  context
  3. First get_screenshot call - The image returned was essentially blank/empty
  (the output showed <output_image> with what appears to be no actual visual
  content)

  I then hallucinated an entire design description based on:
  - The project context from CLAUDE.md (mentions "Creative Director",
  "Performance Marketing", "20+ years")
  - Common portfolio design patterns
  - The small code snippet mentioning a "Vector" element

  That was a significant error on my part. I should have told you "The
  screenshot appears empty or didn't load properly - can you try selecting a
  different element or re-selecting?" instead of fabricating a description.

So it’s 1/3 the price of Opus 4.1…

> [..] matches Sonnet 4.5’s best score on SWE-bench Verified, but uses 76% fewer output tokens

…and potentially uses a lot less tokens?

Excited to stress test this in Claude Code, looks like a great model on paper!

And they left Haiku out of most of the comparisons! That's the most interesting model for me. Because for some tasks it's fine. And it's still not clear to me which ones those are.

Because in my experience, Haiku sits at this weird middle point where, if you have a well defined task, you can use a smaller/faster/cheaper model than Haiku, and if you don't, then you need to reach for a bigger/slower/costlier model than Haiku.

this is the most interesting time for software tools since compilers and static typechecking was invented.

Maybe models are starting to get good enough/ levelling off?

I’ve always found Opus significantly better than the benchmarks suggested.

LFG

They said that they have seen 134K tokens for tool definition alone. That is insane. I also really liked the puzzle game video.

[1] https://www.anthropic.com/engineering/advanced-tool-use

Gemini 3.0 Pro: https://www.svgviewer.dev/s/CxLSTx2X

Opus 4.5: https://www.svgviewer.dev/s/dOSPSHC5

I think Opus 4.5 did a bit better overall, but I do think eventually frontier models will eventually converge to a point where the quality will be so good it will be hard to tell the winner.

We just evaluated it for Vectara's grounded hallucination leaderboard: it scores at 10.9% hallucination rate, better than Gemini-3, GPT-5.1-high or Grok-4.

https://github.com/vectara/hallucination-leaderboard

Gemini is great, when you have gitingested the code of pypi package and want to use it as context. This comes in handy for tasks and repos outside the model's training data.

5.1 Codex I use for a narrowly defined task where I can just fire and forget it. For example, codex will troubleshoot why a websocket is not working, by running its own curl requests within cursor or exec'ing into the docker container to debug at a level that would take me much longer.

Claude 4.5 Opus is a model that I feels trustworthy for heavy refactors of code bases or modularizing sections of code to become more manageable. Often it seems like the model doesn't leave any details out and the functionality is not lost or degraded.

> All evals were run with a 64K thinking budget, interleaved scratchpads, 200K context window, default effort (high), and default sampling settings (temperature, top_p).

I understand scratchpads (e.g. [0] Show Your Work: Scratchpads for Intermediate Computation with Language Models) but not sure about the "interleaved" part, a quick Kagi search did not lead to anything relevant other than Claude itself :)

[0] https://arxiv.org/abs/2112.00114

I can't even use Opus for a day before it runs out before. This will make it better but Antigravity has way better UI and also bug solving.

It planned way better in a much more granular way and then execute it better. I can't tell if the model is actually better or if it's just planning with more discipline

its hard to get any meaningful use out of claude pro

after you ship a few features you are pretty much out of weekly usage

compared to what codex-5.1-max offers on a plan that is 5x cheaper

the 4~5% improvement is welcome but honestly i question whether its possible to get meaningful usage out of it the way codex allows it

for most use cases medium or 4.5 handles things well but anthropic seems to have way less usage limits than what openai is subsidizing

until they can match what i can get out of codex it won't be enough to win me back

edit: I upgraded to claude max! read the blog carefully and seems like opus 4.5 is lifted in usage as well as sonnet 4.5!

It is emphatically not, it has never been, I have used both models extensively and I have never encountered a single situation where Sonnet did a better job than Opus. Any coding benchmark that has Sonnet above Opus is broken, or at the very least measuring things that are totally irrelevant to my usecases.

This in particular isn't my "oh the teachers lie to you moment" that makes you distrust everything they say, but it really hammers the point home. I'm glad there's a cost drop, but at this point my assumption is that there's also going to be a quality drop until I can prove otherwise in real world testing.

I can get some useful stuff from a clean context in the web ui but the cli is just useless.

Opus is far superiour.

Today sonnet 4.5 suggested to verify remote state file presence by creating an empty one locally and copy it to the remote backend. Da fuq? University level programmer my a$$.

And it seems like it has degraded this last month.

I keep getting braindead suggestions and code that looks like it came from a random word generator.

I swear it was not that awful a couple of months ago.

Opus cap has been an issue, happy to change and I really hope the nerf rumours are just that. Undounded rumours and the defradation has a valid root cause

But honestly sonnet 4.5 has started to act like a smoking pile of sh**t

The bigger thing is Google has been investing in TPUs even before the craze. They’re on what gen 5 now ? Gen 7? Anyway I hope they keep investing tens of billions into it because Nvidia needs to have some competition and maybe if they do they’ll stop this AI silliness and go back to making GPUs for gamers. (Hahaha of course they won’t. No gamer is paying 40k for a GPU.)

Given this tech is new, the experience of how we relate to their mistakes is something I think a bit about.

Am I alone here, are others finding themselves more forgiving of "their preferred" model provider?

I love that Antrhopic is focused on coding. I've found their models to be significantly better at producing code similar to what I would write, meaning it's easy to debug and grok.

Gemini does weird stuff and while Codex is good, I prefer Sonnet 4.5 and Claude code.

https://claude.ai/chat/0c583303-6d3e-47ae-97c9-085cefe14c21

Still fucked up one about the boy and the surgeon though:

https://claude.ai/chat/d2c63190-059f-43ef-af3d-67e7ca1707a4