Everyone is waiting for Opus 5.2, but Anthropic might jump straight to version 5.5. Developer Lyra revealed a big scoop early: Claude Opus 5.5, codenamed claude-wafer-eap, is already in secret testing, and it could suddenly hit the market as early as Tuesday this week. Along with the leak came a pricing list — only $4 per million tokens input and $20 per million tokens output, clearly targeting GPT-6, and reportedly outperforming Astra. On another front, Claude Fable 5.2 is also quietly preparing its ground. Just two days ago, foreign media reported that Anthropic must release a major model before its IPO to stabilize market confidence. Now, with these two cards unveiled, the real show has just begun.

Looking back at the timeline is even more intriguing. On July 24, Claude Opus 5 was officially released, followed by a brief 5.2 test checkpoint. Now it's jumping directly to 5.5. This company never intended to do regular updates; instead, it was aiming for a big move from the start.

Anthropic is making a big move quietly: Claude Opus 5.5 skips to internal testing, with a price of $4 per million tokens directly hitting the core of GPT-6.

Everyone is waiting for the schedule of Opus 5.2, but this time, Anthropic is not following the usual rules. Latest reports indicate that the next version number jumps directly to 5.5 — the Claude Opus 5.5, codenamed claude-wafer-eap, is already secretly in testing. The news first came out from developer Lyra, and if action is quick, it could suddenly hit the market on Tuesday. Even more eye-catching is the leaked pricing list: only $4 for every million token input and $20 for output, clearly aimed at GPT-6, and supposedly outperforming Astra. At the same time, Claude Fable 5.2 is also being prepared behind the scenes. Anthropic's dual approach is perfectly timed.

The timing of this surprise is especially intriguing. Two days ago, foreign media reported that Anthropic must release a major model before its IPO to stabilize market confidence. Now, with these two cards unveiled, the real show has just begun. Looking back at the timeline feels even more unusual: on July 24, Claude Opus 5 was officially released, then briefly had a 5.2 test checkpoint, and now it skipped directly to 5.5. From the beginning, this company never planned to make conventional iterations; it was aiming for a big move.

Developer Harshith, who has actually tested it, said a truthful statement: Opus 5.5 has won completely, surpassing previous versions Fable 5.1 and Astra. He made a 3D model of a Waymo autonomous car with 5.5, with an incredibly high level of detail, including small details like the Jaguar logo and the lidar sensor on the roof. More importantly, the cost is only half of Opus 5's. You can feel the terrifying output volume from the token consumption: only 1,400 tokens input and 532,000 tokens output, with a cache of 41.2M read and 1.7M write. Developer Veee also joined in, using Opus 5.5 to create a web-based 3D car display using Three.js, with very solid rendering quality and interaction. For a more challenging test, it can even generate a 3D interactive Wall-E robot console in one go. Whether it's UI layout, mechanical binding or IP accuracy, it shows a strong level of professionalism. Many internal testers have reported that the new checkpoint's problem-solving intuition, deep code refactoring capabilities, and long-term multi-step planning have clearly created a generation gap compared to the current Opus 5.

Along with the news about 5.5, there are also a set of unconfirmed API prices. According to current information, the expected price of Opus 5.5 is $4 per million tokens input and $20 per million tokens output, which is a 20% reduction compared to the current Opus 5's $5 and $25. The key point is the significant optimization of the caching mechanism: the cost of cache reading is reduced to $0.2 per million tokens, and writing is $5. The $0.2 cache reading unit price means that enterprises building long-term, long-context autonomous agent systems will see their costs drop by an order of magnitude. This pricing strategy points to one thing: Anthropic is turning top models from luxury items into daily necessities.

Outside of Opus, the Fable line also has new developments. Developer Pankaj Kumar recently released a trial experience of Fable 5.2. According to him, the official seems to be quietly moving from 5.1 to 5.2, and he happened to get an early ticket for this internal test. After trying it, he gave a very high evaluation. Fable 5.2's 3D generation ability not only catches up with Astra, but is even showing signs of surpassing it, and it can accurately grasp user intent without detailed prompts. He tried to build a 3D scene of a cliffside villa, and the lighting, material reflection, and atmosphere were flawless. Even more astonishingly, it can directly generate a complete game using Three.js, with excellent frontend aesthetics, color schemes, layout, fonts, and interface design, all flawlessly completed, far surpassing previous generations. He also released an extremely complete explosion disassembly render, peeling back layer by layer the optical modules, transparent parts, precision internal hardware, and light and shadow details. Another developer, Chetaslua, tested it without any pre-made materials, without enabling the MCP plugin, and even without any inspiration sketches, Fable 5.2 generated a miniature world entirely using native JS. That super-cute Claude mascot was rendered entirely with pure code. In online tests, Fable 5.2 built a complete running small JS system, with dozens of different head models, facial micro-expressions, dynamic effects, control panels, and interactive logic, all running seamlessly in one generation.

Zooming out, the reason why this round of large model upgrades has made everyone so restless is because Anthropic, OpenAI, and Google are moving in perfect unison, directing the battle to the same battlefield. OpenAI's recent GPT-6Astra indeed raised scientific reasoning and AI intelligence to a new peak, but the high cost of token reasoning ensures it will remain a dragon-slaying sword on the altar, something ordinary engineers dare not use freely in their daily work. Therefore, the recent discovery of gpt-6-sol in the API interface has become the most anticipated weapon in the industry. Meanwhile, Anthropic has put on this big performance, creating a great contrast. Last week, CEO Dario Amodei was still loudly calling for the AI industry to slow down in a long article, but behind the scenes, they threw away Opus 5.2, jumped two levels to Opus 5.5, and prepared to launch a surprise on Tuesday, slashing the price of top-tier intelligence to $4 per million tokens and hitting OpenAI's territory. Google is also unable to wait any longer. Gemini 3.5 Pro's evolution didn't even reach Flash, and it was internally canceled. Gemini 4 Pro became the village's only hope. In Arena testing, it appeared under the name gemini-3.8-flash, completely dominating all indicators. Logan Kilpatrick even stated that Gemini 4 would be Google's largest pre-training project to date, aiming to bring Google back to its peak.

Looking back, the real showdown among the three companies is right in front of our eyes in the coming days. Anthropic's strike is not only against GPT-6 but also against its own IPO anxiety. Skipping versions, lowering prices, and rushing ahead — everyone can see clearly that before the OpenAI Developer Conference, they want to reclaim developers' wallet share first. This Tuesday in Silicon Valley is destined to be another sleepless night.