AI Evening News: OpenCode Raises Credits to $80, GLM-5.3 Enters Gray Testing, Step 5 Open-Source Countdown

The pace of AI product iteration hasn't slowed. From credit boosts to model gray testing to open-source countdowns, several announcements pieced together reveal a few facets of the current large model competition: inference speed, subscription value, and the openness of agent foundation models.

OpenCode Raises Go Plan Credits to $80

OpenCode announced that in its $10/month Go plan, the model usage value for DeepSeek^Ad V4.1 Flash has been increased from $60 to $80. The plan price stays the same, but available credits jump by a third.

This move deserves a closer look. One of the core selling points of subscription-based AI coding tools is the "sense of credits"—users need to clearly understand how much callable model resource their money buys. Raising credits from $60 to $80 strengthens the value-for-money signal without moving the price anchor. For developers who frequently use DeepSeek V4.1 Flash, this could be a direct reason to renew or convert to a subscription.

Combined with recent generative AI website traffic data, ChatGPT's share dropped from 71.8% to 56.4% within a year, but has stabilized and rebounded in recent months; Gemini rose to 25.0%, firmly in second place; Cloud rose to 9.6%, surpassing DeepSeek for third. The competitive landscape is loosening, and tool-layer moves to fight for users through credit adjustments will become more frequent.

GLM-5.3 High Speed Enters Gray Testing

Zhipu has opened gray testing of GLM-5.3 High Speed to some subscribers. Users can randomly receive trial access in Zcode.

According to user testing, the generation speed is about 493 tokens per second, more than twice the peak of GLM-5.3 Flash. If this number holds steady, it will make a noticeable difference for latency-sensitive scenarios—such as code completion, real-time conversation, and agent loop calls.

The gray test uses a "random access" approach, indicating resources are still within controlled limits. High-speed versions are usually not positioned to replace the main model, but rather to supplement specific latency-sensitive tasks. Whether it can later move to full release depends on load and feedback during the gray test.

What is the tiering between "High Speed" and "Flash": In large model product lines, the same generation of models is usually split into different tiers based on "capability ceiling" and "response speed." Flash-style versions often trade lower per-call latency by reducing inference steps, compressing chain-of-thought length, or using smaller activated parameters; High Speed versions further focus optimization on throughput and time-to-first-token, with common techniques including speculative decoding, KV Cache reuse, and batch scheduling for high-concurrency requests.

It's important to distinguish "generation speed" from "model intelligence": tokens per second measures throughput in the decoding stage, and it does not directly equal the model's performance on complex reasoning tasks. A high-speed version may perform excellently on code completion, formatted output, and short conversations, but may not outperform the main model on math or planning tasks requiring long-chain reasoning. This is why high-speed versions are usually positioned as "supplements for specific latency-sensitive tasks" rather than full replacements for the main model.

Step 5 Preview Enters Open-Source Countdown

Step 5 Preview announced that it will soon be open-sourced, positioned as a foundation model for agent tasks in real-world scenarios, and the official countdown has begun.

The positioning of "agent tasks for real-world scenarios" is worth noting. In the past, many foundation model open-source releases leaned toward showcasing general capabilities, but explicitly targeting agent tasks means training data and evaluation may be closer to needs such as tool calling, multi-step planning, and environment interaction. If the accompanying inference framework and evaluation sets are also released after open-sourcing, it will be a fairly practical supplement for teams building agent applications.

Training orientation of agent foundation models: Foundation models for agent tasks differ significantly from models for general conversation in training objectives. The former emphasizes function calling, multi-step planning, error recovery, and the ability to use environmental feedback, and training data often contains a large number of stateful interaction trajectories rather than single-turn Q&A.

When such models are open-sourced, what truly determines usability is often not just the weights themselves, but also the accompanying inference framework, tool-calling protocols, and evaluation sets. If there are only weights and no reproducible evaluation benchmark, teams will find it hard to judge whether the model is truly stronger than general models in their own scenarios. Therefore, "whether open-sourcing includes evaluation sets and inference frameworks" is a key indicator for judging the value of such releases, and a point worth continuing to track.

Research and Open Source: From Mathematical Proofs to 3D Reconstruction

Two research open-source announcements are also worth noting.

HackingFace turns formalized data into a reinforcement learning environment. The project organizes OpenAI's recently published formalized data display into an open-source reinforcement learning environment containing 369 tasks and 405 math problems, equipped with a Linux automatic scorer that can verify proofs given by models. Researchers can use it directly for training, reducing manual evaluation costs.

NVIDIA open-sources Gator. It can turn casually taken photos into 3D assets for simulation, and can inspect and repair reconstruction results on its own. The team also demonstrated a humanoid robot running in a reconstructed scene. The value of such tools lies in lowering the barrier to 3D content creation and robot simulation training, shortening the distance between "taking a few photos" and "a usable simulation environment."

Formalized data and automatic scoring: Formalized mathematics (such as languages used by proof assistants like Lean and Coq) requires every step of reasoning to be strictly checkable, making it naturally suitable as a source of reward signals for reinforcement learning. Organizing formalized data into a reinforcement learning environment, the core value lies in the "automatic scorer"—proofs given by models can be machine-verified as true or false, without manual item-by-item evaluation, greatly reducing the cost of training and evaluation.

The significance of photo-to-3D assets: Reconstructing 3D assets usable for simulation from casually taken photos is difficult due to geometric consistency and physical plausibility. A tool that can inspect and repair reconstruction results on its own means it can reduce the manual model repair step. For robot simulation training, this can significantly shorten the path from "photographing real scenes" to "usable simulation environment," lowering the barrier to data collection and scene construction.

A Few Scattered but Interesting Updates

NASA used the photo localization model Astrolog in the Artemis-I mission to locate 3,687 lunar photos taken by astronauts. The model finds shooting locations by matching lunar surface basemap tiles, a method derived from an earlier project for locating Earth photos.

Tencent's TRTC team launched a free tool called Nukate. Developers can add an asynchronously loaded script tag to turn a webpage's contact page into a real-time chat entry. Visitor messages can go directly to the site owner's Telegram, and replies can return to the webpage in real time. The tool supports integrating custom AI models and is still in early development.

There are also reports that Google had planned to update Gemini 3.X Flashlight, but after three model checkpoints appeared in early September, no new internal release was seen. The update may be shelved later, or it may shift to version 4. There is currently no definitive conclusion.

---

Putting it together, today's information centers on three things: the tool layer uses credits to retain users, the model layer uses speed to grab scenarios, and the open-source layer uses vertical positioning to gain adoption. The directions differ, but the logic is consistent—all are digging deeper into "more specific usage scenarios."

One-line Summary

OpenCode boosts plan credits, GLM-5.3 High Speed enters gray testing, Step 5 Preview open-sourcing soon; large model competition digs deeper into specific scenarios.

Paragraph Summary

Today's information centers on three main threads: the tool layer uses credits to retain users, with OpenCode raising the usage value of DeepSeek V4.1 Flash in its $10 Go plan from $60 to $80; the model layer uses speed to grab scenarios, with GLM-5.3 High Speed in gray testing for some subscribers, measured at about 493 tokens per second, more than twice the Flash peak; the open-source layer uses vertical positioning to gain adoption, with Step 5 Preview announcing an upcoming open-source release explicitly aimed at agent tasks in real-world scenarios. In addition, HackingFace organizes OpenAI formalized data into an open-source reinforcement learning environment with 369 tasks and 405 math problems, and NVIDIA open-sources Gator to turn photos into simulated 3D assets. Overall, the logic across directions is consistent—all are digging deeper into more specific usage scenarios.