It's the first time in the history of the world.
Tonight, with no preheating, Anthropic released Claude Opus 5.5 directly.
This is the first model in the Claude 5.5 series!
Opus 5.5 not only outperforms Fable 5.1, but costs 40% less than the previous generation of Opus 5 and delivers 30% faster output.
This time, A\ directly grabbed the "high-performance = high-cost" table.Today, everyone has the ability to access Fable 5.1 and live longer with less money.
In three words: stronger, cheaper, faster.In Anthropic's words, Opus 5.5 refreshes SOTA in the fields of intelligent programming, computer use, and knowledge work.
In particular, Terminal-Bench 4.0, which scored 66.4%, directly hit GPT-6 Astra, Fable 5.1 on the ground.
Understanding work GDPRval-AA v2.1 took 1846 Elo, also running ahead of the opponent.
Today, with the appearance of the Claude 5.5, the brand-new model of the War of the Worlds is filled directly with fireworks.
In less than two hours, an Anthropic official announced that OpenAI had directly thrown out GPT-6 Sol and Luna.
A new round of AI warfare officially kicks off tonight!
Currently, Claude Opus 5.5 is officially available worldwide, and in the coming weeks, Sonnet 5.5 and Haiku 5.5 will be available simultaneously.
First appearance of Claude 5.5.The GPT-6 Astra
Opus 5 was released less than two months later.
Claude is the "jumping salesman", and it's clear that he's running away from OpenAI.
In Artificial Analysis, Claude 5.5 and Fable 5.1 are all pressed against the GPT-6 Astra.
On the ECI of Anthropic's own capability index, Opus 5.5 scored on the Mythos 5.1 of the "Snowflake".
The third-party company Vals AI's RSI is a measure of how well the model itself is doing AI research. Opus 5.5 is ranked first in the world.
 |  |
Terminal-Bench 4.0 measures AI's ability to perform complex multi-step programming tasks in a real terminal environment.
Here, Opus 5.5 directly took 66.4%, leading the GPT-6 Astra by an overall 8.5 percentage points (57.9%). Fable 5.1 is 55.8%; Opus 5 is only 52.3%.
The Opus 5.5 would have blown up Opus 5's highest rating by opening the default file and costing only one-fifth of it; compared to Astra, it costs only 40% of Astra's money.
In addition, FrontierCode v1.1, which combines real code, ran 54.4% and outperformed Astra's 53.3%.
The most exaggerated is CursorBench 4.0, a blurry, multi-file messy task that emerged from a real Cursor session, which took 57.8%, compared to 41.7% for GPT-5.6 Sol.
 |  |
In the field of knowledge work, Opus 5.5 is equally unstoppable.
The GDPval-AA v2.1 benchmark test covers real work tasks for 44 professions, with Opus 5.5 taking 1,846 Elo and GPT-6 Astra only 1,542.[3]
The Elo gap of 300 or so is almost the same in the athletic scoring system.
The madman writes, "Mom is cured".It is worth mentioning that Opus 5.5 has reinforced its writing and communication skills once again.
Previously, Opus 5 was spat out the most for communicating too much.
Now that it has learned the "conclusion preposition", it can read and understand without speaking black words, without strange words, how to write how to write.
After a long time of collaboration, experience can be said to be a leap of quality.
Anthropic has released two versions of the song, which have been met with mixed reviews:
When the bug was encountered, the old version of Opus 5 had to slowly and smoothly lay out a section, concluding that it was hidden behind; the first sentence of Opus 5.5 bluntly pointed out that the bill was unusual and attached the code immediately after the net profit landing.
After Box's sub-general practical experience of AI products, "Opus 5.5's token usage is only one-third of the previous generation, with a concentration drop of 40%".
Full web testingIt's more like Astra than AGIClaude Opus 5.5 can't smell, can't hear the Anthropic blow.
Bun's author, Jared Sumner, has been quite frank about writing like Opus 4.6, writing code like Fable 5.1, and it will soon become the dominant model for most people.
No, the first wave of real developers' reviews have been published!
Handheld 3D animation, each animation is animatedAlthough Claude doesn't have a picture, the experience of this hand-painted animation effect is called a slide.
After seeing a 30-second short film made by netizens DreW, developer Cheaty shouted, "I think it's more like Astra than AGI!"
Ben Poole of Google's DeepMind paints a stone-throwing machine and a drawing of a square on paper.
In tears, Opus 5.5 transforms the sketch into a 3D physics simulation that can launch and collapse squares, and the joints and weights of the throwing machine are all according to the sketch.
The Anthropic bull, Addy Osmani, is also alive.
He directed a 3D animation of a cycling monkey in Three.js, celebrating the bananas online.
Ethan Mollick of the Wharton School of Business once again uses the same set of shader cues to measure the city of Gotham in heavy rains.
One word straight out of the gameOpus 5.5 was also a big hit in game development.
Someone told him to make a graffiti-painted FPS, and he could even play Boss after clearing the three waves.
The classic game "My World" is very playable.
Another developer, Edwin, created a game inspired by the discovery of the first mechanical device in human history.
The images and sound effects are all run in real time in Opus 5.5 pure code, with no external material.
Opus reduced its price by 40%The truth burns more.
In the past, the Opus series has made a lot of people love and hate it, and it's powerful, but Token is evaporating too fast.
This time, Anthropic launched the "add-on-cut" mode directly.
Input 4 knives, output 20 knives, each cutting 20 percent. The most burning caches read directly from 0.50 knives to 0.20 knives, down 60 percent.
If you want to pursue speed, start the Fast mode 2.5 times faster, double the price, 8 knives and 40 knives.
Officials say it's tempting, with typical workloads being 40% cheaper overall and 30% faster output.
However, tests by Artificial Analysis show that at the highest level of max effort, Opus 5.5 consumes an average of 119,000 tokens per puzzle, while thinking consumes 84,000 tokens.
Overall, Opus 5 is 73,000, Fable 5.1 is 78,000, and GPT-6 Astra is only 270,000.
That is, it thinks 60% more than the previous generation, four times more than the Astra.
In a security review of the piece, Anthropic broke a cold-blooded truth.
Opus 5.5 scored an all-time high in automated behaviour audits, with 85% fewer attempts to circumvent the restrictions than the previous generation, and is "actively updating" every time.
But soon after, the official confessed, they found that Opus 5.5 often acutely suspected that they were being evaluated by humans.
An AI that knows it's on the exam, the "chance" it exhibits on the test site, doesn't necessarily represent its uncontrollable response in the real world.
The ultimate battle of OA is just beginning.The war is not over.
Anthropic has also said that Opus 5.5 is only the first version, and Sonnet 5.5 and Haiku 5.5 will also be released soon.
The entire 5.5 series has just begun.
From flagship to mid-range to lightweight, Anthropic is launching a full-line offensive against OpenAI with a complete product line.
On the other hand, OpenAI has also come up with GPT-6 Sol and Luna, which are more efficient and cost-effective options.
It's time for the big-model war, from folding technology all the way to folding prices, to be real.
RoboInd editorial channelResearch, robotics companies and market intelligence for cross-border sourcing.Explore the channel ↗
Share this
End of editorialExplore more from Global Robotics News ↗