Anthropic Leaks Opus 5.5, Aiming to Challenge GPT-6

This time, is Anthropic holding something big?

Everyone was waiting for Opus 5.2, but the latest update reveals that the next version will jump straight to 5.5!

Early this morning, developer Lyra uncovered a revelation: Claude Opus 5.5, codenamed "claude-wafer-eap," is already undergoing private testing.

It’s said that if you act quickly, it will suddenly drop this Tuesday.

Clearly, Opus 5.5 is targeting GPT-6, with rumors suggesting its performance directly surpasses Astra.

Even more striking is the disclosed pricing: input of one million tokens for just $4, output of one million tokens for $20.

This isn’t over—they also quietly laid the groundwork for Claude Fable 5.2.

Just two days ago, foreign media reported that A must unveil a major model before its IPO to restore market confidence.

Now that these two hole cards are revealed, the real show is just beginning.

Now, before GPT-6 has even cooled down, Anthropic is about to strike.

And all of this happened precisely at the same time that GPT-6 Sol and Gemini 4 Pro were undergoing coordinated gray testing.

Three companies are competing in the same space, and Anthropic grabbed the headlines in the most aggressive way possible: skipping levels, cutting prices, and racing ahead—all in one swift move.

The developer Lyra, who first revealed Claude Opus 5.5, also disclosed another detail: 5.5 can outperform GPT-6-Sol.

But more intriguing are the actions behind this leap:

On July 24, Claude Opus 5 was officially released. Afterward, a brief test checkpoint labeled 5.2 appeared, and now it has moved directly to 5.5.

Anthropic never intended to make a standard update—this time, from the start, they were aiming for a major breakthrough.

Experienced user Harshith, who has actually tested it, said it straight: "Opus 5.5 absolutely crushes it—outperforming both Fable 5.1 and Astra."

He casually generated a 3D model of a Waymo autonomous vehicle, and the level of accuracy was astonishing—

Even small details like the Jaguar logo and the lidar sensors on top are clearly generated.

More importantly, the cost is only half that of Opus 5.

Regarding the consumed tokens, consider the staggering output volume: only 1.4k input tokens generated a full 532k output tokens, with the cache performing 41.2M reads and 1.7M writes.

While performance surged, the price dropped by half.

Developer Veee also joined in, creating a web-based 3D car viewer using Three.js for Opus 5.5.

In terms of rendering quality and interactivity, Opus 5.5 is well-executed.

Let’s look at a more advanced test—it can even pass on the first try, directly outputting a 3D interactive “WALL-E robot” console.

Demonstrated exceptional professionalism in UI layout, mechanical integration, and IP fidelity.

In addition, many developers who participated in the beta testing reported that this new checkpoint's problem-solving intuition, deep code refactoring capabilities, and long-range multi-step planning performance have clearly surpassed the current Opus 5.

Performance soars, prices slashed

Along with the rumors surrounding version 5.5, a set of unverified API prices has emerged.

According to current reports, the price of Opus 5.5 is expected to be—

$4 input and $20 output per million tokens. Compared to the current Opus 5 pricing of $5/$25, this represents a reduction of approximately 20%.

More importantly, the caching mechanism has been significantly optimized. Cache reads are now $0.2 per million tokens, and cache writes are $5.

A cache read price of $0.2 means the cost for enterprises to build autonomous agent systems with long-duration, long-context interactions will drop by an order of magnitude.

This pricing strategy means one thing: Anthropic is turning its top-tier models from a luxury into a commodity.

Beyond Opus, there’s new activity on the Fable line.

Developer Pankaj Kumar has just released the preview of Fable 5.2.

According to him, the official team appears to be quietly switching from 5.1 to 5.2, and he happened to get an early bird ticket for this beta test.

After getting familiar with it, he gave it a very high evaluation: Fable 5.2’s 3D generation capability has not only caught up with Astra but is also showing strong potential to surpass it.

You can accurately analyze what you want without a detailed prompt.

He tried building a 3D scene of a "Cliffside Villa" using Fable 5.2, and every detail—from lighting and material reflections to atmosphere—was flawless.

Even more impressively, it can directly use Three.js to generate a complete mini-game in one go.

Moreover, the frontend design of this model is exceptional—its color scheme, typography, and interface layout are flawless, leaving previous generations far behind in terms of polish and completion.

This demonstration features an extremely detailed exploded rendering, systematically unveiling the layered optical modules, transparent components, precision internal hardware, and lighting effects.

The overall composition of the final image is surprisingly refined.

During another developer Chetaslua’s test, no pre-made assets were provided, no MCP plugin was enabled, and not even a single sketch was needed—Fable 5.2 generated a "miniature world" purely using native JavaScript.

This adorable Claude mascot is entirely rendered pixel by pixel using pure code.

Same test from a netizen.

Fable 5.2 has built a complete, functional small JavaScript system—

Dozens of entirely different head modeling, facial micro-expressions, dynamic effects, control panels, and interaction logics all ran seamlessly together in a single generation.

Finally, let’s look at a real-world head-to-head test: between Fable 5.2 and Astra—who really comes out on top? Just take a look at the images below to see for yourself:

Use Opus 5.5 to solidify enterprise-grade infrastructure while leveraging Fable 5.2 to manage creativity and front-end development.

Anthropic’s two-pronged approach is timed with exceptional precision.

These days, the pace of the large model community has been completely accelerated.

The current AI arena is no longer a solo performance dominated by a single player.

This wave of technological transformation has everyone on edge because Anthropic, OpenAI, and Google have remarkably aligned their efforts, directing the battle toward the same critical frontier.

The technological roadmaps of the three giants have now been fully laid out on the table.

OpenAI: Astra raises the limit, waiting patiently for the daily essential, GPT-6 Sol

The GPT-6 Astra previously released by OpenAI has indeed raised the bar for scientific reasoning and AI intelligence to a new level.

But that high token inference cost ensures it remains nothing more than a "Dragon-Slaying Blade" enshrined on a pedestal—ordinary engineers dare not use it freely in their daily work.

Precisely for this reason, gpt-6-sol, recently uncovered in the OpenAI API, has instantly become the most eagerly anticipated powerhouse in the entire community.

Anthropic: Publicly calling for slowdown, privately releasing a new version

The biggest contrast in this entire round of events has to be Dario Amodei, CEO of Anthropic.

Just last week, Dario loudly called out in a ten-thousand-word article, "AI is advancing too rapidly; the industry needs to slow down."

In secret, they directly moved Opus 5.2 to the recycle bin and skipped two versions to pull out Opus 5.5, preparing for a surprise launch on Tuesday.

Even more aggressive is this "price suicide charge," flooding OpenAI's core territory with millions of tokens priced at just $4.

Google: After four months of silence, now it's all about Gemini 4 Pro.

On Google's side, it can't wait any longer. The evolution of Gemini 3.5 Pro is even less advanced than Flash, and it has been internally canceled for release.

Gemini 4 Pro has become the village's only hope.

In the Arena real-world test, the 4 Pro appeared under the code name 'gemini-3.8-flash' and outperformed all others across every metric.

Logan Kilpatrick explicitly stated that Gemini 4 will be Google's largest pre-training project to date, with the goal of "returning Google to its peak."

Looking back, the real epic showdown among the "Big Three" is happening right under our noses these days.

Anthropic's move this time cuts not only toward GPT-6 but also toward its own IPO anxieties.

Skipping levels, slashing prices, and racing ahead—anyone with eyes can see: it’s about reclaiming developers’ wallet share before the OpenAI Developer Conference.

This Tuesday in Silicon Valley is bound to be another sleepless night.

Who are you rooting for in this round?

Reference materials:

https://x.com/lyraxana/status/2101668658711798083?s=20

This article is from the WeChat public account "New Intelligence Yuan" (ID: AI_era), author: ASI Revelation, editor: Peach