Tech • AI • Robotics • Game
Huge Opus 5.5 Leaks + Cheaper? Qwen 4, Kimi K3.1, MiniMax M3.1 & Step 5 Preview! AI News
9.4/10TL;DR
A new wave of AI model launches appears imminent, led by Anthropic’s Claude Opus 5.5, while Chinese labs including StepFun, MiniMax, Alibaba, and Moonshot AI prepare major updates across language, coding, and image generation.
KEY POINTS
Anthropic tests Claude Opus 5.5
Anthropic is reportedly testing a new checkpoint labeled Claude Opus 5-5 under the codename Claude Wafer EAP, replacing an earlier Opus 5.2 track. The model is said to be targeted for release as soon as Tuesday, with indications that Opus is the most likely near-term launch among Anthropic’s upcoming systems. Early comparisons place it around the level of GPT-6 Astra on coding-heavy visual generation tasks.
Lower pricing could sharpen competition
Reported pricing would make Opus 5.5 notably aggressive for a frontier model: $4 per 1 million input tokens and $20 per 1 million output tokens, with cached reads at $0.20 and cached writes at $5 per million tokens. If accurate, that would position it as a cheaper high-end alternative while preserving top-tier performance. The trade-off appears to be a slightly smaller context window than the latest prior checkpoint.
Visual and simulation outputs show stronger capability
Internal tests suggest the new Opus 5.5 checkpoint outperforms the earlier Opus 5.2 variant in one-shot scene generation and simulation-style tasks. Examples include procedural SVG scenes, a coffee-machine demo, and a detailed 3D BMW M5 environment. Some functions reportedly remain imperfect, but the overall jump points to stronger multimodal reasoning and code generation.
StepFun unveils Step 5 Preview
Chinese AI lab StepFun has introduced Step 5 Preview, a flagship model focused on agentic work, software engineering, professional knowledge tasks, and especially finance. The system uses a 600 billion parameter architecture with only 27 billion active parameters at a time, combines that with vision support, and offers a 1 million token context window. The design aims at long-horizon execution on complex tasks.
Step 5 targets lower cost per task
StepFun’s own benchmarking places Step 5 Preview around the level of GLM 5.3 and Kimi K3 while costing roughly 65% less per task. The preview is already available through the company, while open weights are planned for October 15. Even so, the model’s sheer size is likely to make local deployment difficult despite sparse activation and possible quantization.
MiniMax appears close to M3.1 launch
References to MiniMax M3.1 have surfaced in test files tied to a newly found commit in the company’s code repository, pointing to late-stage preparation. Comments from MiniMax leadership during an August 26 interim results call also indicated that M3.1, M3 Pro, and H3.1 were nearing completion. The expected emphasis is less on raw scaling and more on reliability, stability, output quality, inference efficiency, and agent generalization.
MiniMax also signals a much larger M3 Pro
Alongside M3.1, MiniMax has described M3 Pro as a model nearing 3 trillion parameters with a major architectural upgrade. The company has framed it as a push into larger pretraining scale, better training efficiency, and stronger performance on extreme long-horizon tasks. That suggests a two-track strategy: a practical production model and a far larger frontier system.
Alibaba may use Aspara to reveal Qwen 4
Alibaba appears to be preparing the Qwen 4 family, potentially for unveiling at its upcoming Aspara conference. That timing would fit prior launches, as earlier Qwen generations were introduced around the same event cycle. Conference materials point to a session on advancing Qwen toward agentic AI, with planned discussion of a new flagship model, stronger coding and office capabilities, an omni model, and real-time multilingual translation.
Qwen Image 2.1 arrives as a lightweight open model
Ahead of that event, Qwen Image 2.1 has already launched as an open-weight image model with only 7 billion parameters. Despite its size, early reactions describe it as highly competitive, including against systems such as Google Nano Banana 2. It supports image generation and editing, up to 10 reference images, high-fidelity edits, and native RGBA transparency, making it unusually capable for a compact openly available model.
Moonshot AI hints at Kimi K3.1
Moonshot AI is also signaling a likely near-term release of Kimi K3.1. A teaser posted by the official Kimi account used a string of numbers that aligned with pi after the opening 3.1, a clue widely interpreted as a reference to the next version. No formal launch details have been confirmed, but the teaser suggests that another major Chinese model announcement may be close.
CONCLUSION
The next AI release cycle is shaping up to be unusually dense, with Anthropic pushing price-performance at the top end and Chinese labs accelerating across open and closed models. The main competitive themes are clear: cheaper inference, longer context, stronger coding, and more capable multimodal systems.
Ask a question