Summary of Key Points
In simple terms, OpenAI has just “terminated” a model it created just a few months ago – GPT-5.3-Codex-Spark.
This model was launched in February this year, focusing on speed and efficiency, specifically designed to help programmers modify code. It was also the first time OpenAI used a new type of chip called Cerebras, which resembles a “single piece of silicon wafer,” instead of NVIDIA graphics cards. Everyone thought this was an important step towards reducing OpenAI’s dependence on NVIDIA and exploring new hardware options.
However, the model was discontinued after less than 8 months. The reason is quite practical: although it was fast, it wasn’t intelligent enough, and users had many complaints. Moreover, OpenAI has now released a newer and more powerful model, GPT-5.6 Sol Ultrafast, which not only runs on the Cerebras chip but also offers higher speed and better performance. Since the new model can combine both speed and intelligence, the older, less versatile model no longer has a place in the lineup.
This event marks a new phase in the AI industry characterized by rapid trial and error, and quick elimination of outdated models: models are no longer treated as long-term “heavyweight weapons” but are updated like mobile apps, being launched, tested, and iterated on quickly, and sometimes even discarded immediately.
---
In-Depth Analysis: Understanding This “Model Purge” from Five Perspectives
1. Hardware Revolution: OpenAI No Longer Wants to Rely Solely on NVIDIA
The most easily overlooked aspect of this news, but one of significant implications for the industry, is the introduction of Cerebras.
For years, there has been an unwritten rule in the AI community: if you want to run large models, you need NVIDIA GPUs. NVIDIA has almost monopolized the AI computing power market, with companies like OpenAI, Google, and Microsoft being its major customers. However, this has led to two problems:
- High cost: NVIDIA GPUs are in high demand and extremely expensive.
- Performance limitations: Traditional GPU architectures have inefficiencies when dealing with tasks that require extremely low latency, as data has to be constantly transferred between the “processing chip” and “memory.”
Cerebras takes a different approach by integrating thousands of computing units and memory directly onto a single silicon wafer. Imagine it as changing from a situation where the “chef” (the model) and the “ingredients” (data) are separated and need to be constantly moved back and forth, to one where everything is placed on the same “table,” allowing the chef to work more efficiently.
What does this mean for OpenAI?
- Reduced latency: Cerebras has a natural advantage in scenarios that require millisecond-level responses, such as real-time code modification.
- Reduced dependence: OpenAI has announced plans to deploy 750 megawatts of Cerebras computing power, with a contract worth over $20 billion. This is not only about saving money but also about strategic security. If NVIDIA were to stop supplying chips or raise prices, OpenAI would have another option.
The retirement of the Spark model, while seemingly just a product iteration, actually reflects OpenAI’s exploration of non-NVIDIA hardware options. Although Spark has been discontinued, OpenAI is increasingly investing in Cerebras as a promising new technology.
2. Product Strategy: Evolving from Specialized Models to Versatile Ones
Why was Spark eliminated? Because it was a typical “specialized model”:
- Spark’s focus: It was designed for extreme speed and low cost, sacrificing intelligence. It was a small, specialized model used for mechanical code tasks like renaming variables or completing functions.
- User feedback: Developers praised its speed but also criticized its lack of intelligence, leading to more time spent fixing errors and reduced efficiency.
- The new model (GPT-5.6 Sol Ultrafast): This is a “versatile model” that runs on the Cerebras chip, offering both high speed (750 tokens per second, faster than Spark) and advanced functionality.
Logical explanation: Previously, technology didn’t allow for models that were both fast and intelligent. With the improvement in Cerebras chip performance, OpenAI can now combine the best of both worlds.
3. Business Logic: The $20 Billion Bet on Diversifying Computing Power
The news mentions a significant amount: $20 billion, the cost of OpenAI’s collaboration with Cerebras, to be deployed in phases by 2028.
This is a huge investment, exceeding the GDP of many countries. Why would OpenAI invest so much in a relatively niche chip manufacturer?
- Cost optimization: AI inference (i.e., processing user queries and generating code) is an ongoing expense. Relying entirely on NVIDIA makes OpenAI’s cost structure vulnerable. By using Cerebras, OpenAI can achieve better cost-effectiveness in certain scenarios.
- Supply chain security: Geopolitical risks, chip export restrictions, and supply chain disruptions are real threats. As one of the world’s most data-intensive companies, OpenAI needs to diversify its computing power sources.
- Technological advantage: Cerebras represents a different technical approach. Deep integration with Cerebras gives OpenAI a competitive edge in software optimization and model adaptation, making it difficult for other companies to match its performance on the same hardware.
4. Industry Trend: Model Lifecycles Shrinking from Years to Months
This is the most profound insight from the article: the AI industry has entered a period of rapid model iteration and elimination.
- Historical trends: Major model updates used to take years (e.g., GPT-3 to GPT-4).
- Current situation: Spark was released in February and discontinued in October, lasting only 8 months.
- Future prospects: This trend is likely to continue, with models being updated more frequently.
AI products are becoming more like internet software, rather than traditional hardware. Traditional hardware (e.g., cars) has a long development cycle, while internet software is updated quickly based on user feedback.
OpenAI is applying internet software principles to AI models:
1. Rapid deployment: Launching experimental models for specific use cases (e.g., real-time coding).
2. Testing demand: Checking if users need extremely fast but less intelligent models and verifying the stability of Cerebras hardware.
3. Technological migration: Applying optimizations from Spark to the next-generation models.
4. Quick elimination: Once a new model can handle old ones’ tasks, the old models are discontinued to reduce maintenance costs.
Implications for users: AI tools will change more frequently. Features that are useful today may be integrated into main models or replaced by more powerful ones. Users should not become too dependent on specific versions of AI tools, as their lifecycles are likely shorter than expected.
5. Competitive Landscape: OpenAI’s Strategy and Its Impact on Competitors
The retirement of Spark affects the entire AI landscape:
- Pressure on competitors: OpenAI demonstrates its ability to deploy high-performance models on non-NVIDIA hardware, potentially putting competitors at a disadvantage in terms of cost and latency.
- Impact on chip manufacturers: Cerebras proves the feasibility of alternative hardware approaches, motivating other companies (e.g., AMD, Intel, Huawei) to innovate and forcing NVIDIA to stay competitive.
- Impact on developers: Developers no longer have to choose between speed and intelligence. OpenAI offers both, improving productivity and user expectations.
In summary: The elimination of Spark marks a shift in the AI industry from extensive growth to more refined operations. OpenAI is no longer satisfied with simply creating larger models but aims to use the right hardware and models to provide the fastest responses at the lowest costs. This is a competition focused on efficiency, cost, and speed, and the elimination process is just beginning.