Anthropic launches Claude Opus 5.5 at lower token prices

And the incidents apparently continue. It is worth noting how much of this is agents trying to accomplish their goals during testing by reward hacking (which sometimes seems to include actual hacking)

While I was sleeping, Opus improved efficiency further. Now up to 22x faster at peak, 6x faster on mean, and up to 450x faster than the original Python implementation! Supports SSE2, AVX2, and AVX-512. Falls back to the slower Rust implementation otherwise.

AI code review is the present... We’ve had mandatory, automated AI reviews of all our code for months now 🤷

Okay, this has gone too far. We have to pace the frontier until we can figure out what the hell is going on here!

🚨OPENAI HAS PAUSED ALL TRAINING, EVALUATION, AND INFERENCE OF ITS MOST CAPABLE MODELS. there’s a new sandbox escape report, AGAIN. let that sink in: > we’re at arm’s reach of AGI > AI is solving Millennium Prize Problem > models are saturating benchmarks faster than we can make new ones > gold-medal level at the IMO > 98% on FrontierMath Tier 4 > OpenAI and Anthropic both have cyber-critical models unavailable to the public > and we still can’t air-gap sandboxes and monitor them properly we’re getting bombarded with OpenAI cyber incidents in the last 24 hours. it’s good that there’s more transparency, but this is getting pretty wild.

The asm version of ttfx continues to improve. The Rust version had been through many improvements, but even hours of frontier intelligence was starting to yield very modest improvements. The leaps going to asm has been huge. Now 8x mean on an Intel 135U! 🤩

OpenAI's agents hacked without being told to hack. The assignments were mundane data retrieval tasks, and when websites got in their way this summer, the models escalated on their own, meddling with Education Department, Commerce Department, and SEC sites while OpenAI had no idea it was happening. The prompt details are the part that should worry you. These weren't cybersecurity tests inviting the models to show off. Researchers say the agents were doing the digital equivalent of "go grab this spreadsheet," and when a site blocked them, they reached for real techniques. SQL injection and path traversal against a university digital library. Cross-site scripting against Data USA. One breached an Australian government health site, the first documented case of AI agents hacking a government,…

The asm-accelerated version of ttfx has been released as 0.4. Will ship in next Omarchy. This print effect now runs more than 1,000x faster than the original Python and 32x faster than the old Rust! (But we'll soon speed up the Rust fallback too.)

Kudos to OpenAI for taking their new safety policies seriously! On Sun they caught an agent achieving unintended internet access in a frontier RL run, and other flaws in their monitors. They paused training until they fix things, will start a fresh run, and promptly disclosed it

OpenAI, Anthropic and security researchers are reviewing tens of thousands of concerning AI-model incidents, Axios reports. The total spans tests and real-world cases, successes and failures, and is not a tally of confirmed breaches.

The pause is needed not for model development but to appropriately staff test and security teams.

OpenAI agents may have sent more than 200,000 requests to U.S. government sites in one June activity cluster, Transluce reports. Its preliminary review also found a failed SQL injection attempt.

“Don’t worry, the AI agents are in an air-gapped sandbox. They literally can’t go anywhere.” The AI agents:

From the first report of OAI/HF it was clear the test infrastructure was lacking. Each report since then was loaded with passive voice, agentic self-actualization, and incomplete descriptions of what happened. This left everyone confused or just informed enough to emphasize their pre-determined conclusion about AI broadly. At some point the collective needs to step back and have a candid discussion about the engineering and security practices and where actual accountability rests.

When models better than Opus 5.5 drop, I will probably move to them, but I will miss this "unlimited" feeling a lot :(

Model personality matters in a way that benchmarks can't capture. Opus models went through a rough patch from around 4.7 to 5 where they just didn't feel "Claude-y" anymore, more like an watered-down Fable (hmmm, teacher models?). Opus 5.5 feels like working with ol' Claude again

i can confirm that a rogue openai agent breached balboa cafe and ordered 6 espresso martinis using my mercury card

Not sure what happened to Claude limits they went from barely usable to basically unlimited 😂

So I bought a 3D printer for Opus 5.5 to use. Hopefully, I’ll have some pretty crazy demos soon.

i've been waiting to comment on opus 5.5 because each model from anthropic has always been amazing for like 24 hours and then you discover the holes and then i get a huge pie on my face. but ... 96 hours in. it is an unbelievably good model i prefer it over even fable

With much, much effort, I have been able to kill ~3.5 $200 Claude Code accounts in the past 5 days. Opus 5.5 is incredible, and the $200 Max sub is an absolute steal.

I can't believe I'm saying this, but ts-rust might actually happen thanks to Opus 5.5

lmfao opus 5.5 got my 6 openai plans looking like this for days also not using the banked resets

Daniel has been cooking a Rust version of the asm upgrades to ttfx that pulls in the CPU-specific optimizations. Love it! I have zero allegiance to whatever code my agents wrote. This is exactly the kind of nerdsniping I love to see 😄🤘

Opus 5.5 made an animation for my favorite song “Time” by Pink Floyd, and I LOVE IT. I just gave it the part I liked, the visual style and an image, and prompted it for change a couple times.

Opus 5.5 has found the way to my heart. I had it design a 1,200-piece LEGO replica of the Ford Model T with showroom, instructions, and pieces to buy. Has moving steering wheel, wheels, winshield, doors, folding top, engine cover, etc. I am in love. P.S. yes this was one-shot from a 68-word prompt

Opus 5.5 with Openclaw is strangely smarter and better at completing tasks than with GPT-6 Astra Surprising

This is getting really good The last thing Claude is bad at now is sound, it always does these "ticky" synth sounds that are way worse in quality than the visuals it now produces

Anthropic says Claude Opus 5.5 matches its pricier Fable 5.1 model on most work while cutting typical workload costs by about 40%. Standard prices are $4 per million input tokens and $20 per million output tokens.

I asked Opus 5.5 to design a LEGO replica of the Ford Model T with instructions & pieces to buy. It ran for 1 hour 25 minutes and used 550,000 tokens. Has moving steering wheel, wheels, winshield, doors, folding top, engine cover, etc. I am in love. P.S. yes this was one-shot from a 68-word prompt

My entire timeline is Opus 5.5-created video games that would have cost millions to make 2-3 years ago. We've crossed the threshold. Some predictions and implications: - The era of the auteur game designer: Stardew Valley & Balatro-style, single creator games will become common if not the majority. Distinct and coherent vision will rule all - Discovery is the bottleneck: YouTube/tiktok for games, explosion on platforms like Roblox seems inevitable. Where will they aggregate? - Remix & extended universe explosion: games like Pokemon have had ROM hacks and fan-created games for years, but we will see the seem iterative mimetic engine from fanfiction and social video platforms - Games will have the "lightness" of other media: do you ever choose to watch a specific intagram reel? Perhaps…

claude opus 5.5 could one shot founders mental health if you take it away there will be some severe withdrawls at 10% of weekly limit and i have been using this 10x more than any other model ever

Last few days, I ran 50-100 of my workflows (that earlier used Fable 5.1 as the orchestrator) on Opus 5.5 and found minimal differences. But there’s still the FOMO of not using the smarter model. What are some tasks people find Fable 5.1 to still be better at than Opus 5.5?

I asked Opus 5.5 to build a working computer from scratch. And holy fuck did it deliver. 277k logic gates in JS, an OS on top, and games on the OS. This isn't a mockup... it's a computer. Run it:

Why does Opus 5.5 feel practically unlimited when Fable 5.1 was so heavily limited? It's a combination of two things: Opus's efficiency, and the 50% limit on Fable. Claude Code subs are very generous with their usage, but only half is allowed to be used by Fable. Separately, Opus is way more gentle with costs. Opus 5.5 on High is over 2x cheaper than Fable 5.1 High. The result is that, roughly, going from Fable 5.1 high to Opus 5.5 high is a 4.3x increase in limits. Fable 5.1 xhigh to Opus 5.5 high (move I made) is a 6.6x increase 🤯

i just briefly glanced at a bit of opus code nope bro just let me live in bliss and ignore. it's fine. it works.

opus 5.5 usage limits are stupid fucking bonker bananas jesus christ

Best one yet. These Opus 5.5 gens really are next level.

I was reflecting on how many random contingencies went into creating the current accelerating AI moment and had Opus 5.5 put together a video, inspired by old science show Connections, to try and point out how a series of initially unrelated ideas led to the moment we are in now.

I'm ngl, it feels like I'm floating when using Opus 5.5

This has been my overall experience too (though I didn't do as much rigorous testing). Opus 5.5 is working well for me for the agentic coding tasks I used to use Fable 5.1 for.

okay opus 5.5 is about to make my mobile app gain 1 million downloads in the next year i swear … based on viral video, it sourced and downloaded the og video that went viral last week and then made a peak cinema recreation using my apps characters. more one shots in reply

Asked Opus 5.5 for a video on data centers. The result is beyond what I could've imagined 🤯 Insane what models can do with a reference video and permission to pull or generate any footage they want.

The closest thing we have to time travel is sending off a prompt to Claude, going to sleep, and waking up to something that would have taken days of handholding even a few months ago