OpenAI Unveils Ultrafast Mode, GPT-5.6 Sol 14x Faster

OpenAI released two moves on the same day: first, the Ultrafast preview, with GPT-5.6 Sol running in this new tier at up to 14x faster than standard processing and outputting 750 tokens per second; second, the Computer History feature for ChatGPT desktop. The former is a speed tier, the latter a memory input source, and neither has disclosed pricing.

Ultrafast is positioned for tasks where "a step slow means a step late." According to the source article, OpenAI's internal scenarios include simultaneously reading logs, comparing code changes, and sorting through alert conversations during system alarms; simultaneously analyzing trading records and picking out suspicious actions when market signals jump; and simultaneously answering product questions, checking inventory, and giving recommendations while a user hesitates at the shopping cart. These are usage scenarios given by OpenAI, not third-party-verified benchmark results.

Early users gave positive reviews: Alex Wang, head of applied AI at financial research AI company Rogo, described it as "feels like talking to a person in real time"; Courtland Lykins, head of voice AI product at AI customer service platform Podium, said Ultrafast lets customer service handle tough problems without the call being interrupted to wait for the model to spin. These are subjective user feedback.

This speedup has a key premise: according to the source article, the model's intelligence level is unchanged at all; only the hardware and scheduling method running it differ. Many vendors make fast versions by actually swapping a large model for a smaller-parameter, discounted-capability version, trading intelligence for speed; but Ultrafast still runs the top-tier GPT-5.6 Sol.

What supports the speed is Cerebras's wafer-scale engine: model weights are placed directly into the chip's 44GB SRAM, without repeatedly moving data from memory as traditional GPUs do, thereby bypassing the memory bandwidth bottleneck that has long constrained inference speed. This is the technical explanation given by the source article.

The relationship between the two had long been laid: according to the source article, OpenAI and Cerebras signed a multi-year agreement earlier this year, planning to deploy a 750MW-scale wafer-scale system, with the agreement worth over $10 billion; Ultrafast is a visible output of this agreement. The agreement amount and deployment scale are both relayed by the source article, with no first-party original text corroboration seen.

ChatGPT's original memory feature is called Memories, which records things you actively tell it in conversation—what style you like, what project you're working on, what you talked about last time. Computer History opens an extra input source for it, recording interaction events such as clicks, typing, keyboard shortcuts, and app switches, without you having to explicitly say anything.

It is born from the research preview Chronicle launched in April, and this counts as a refactor. Actual queries sound like schedule conversations, such as "What was I doing before my last break," "Where is that proposal document I was looking for earlier today," "Help me list the tasks and progress I made today." Open the history page in settings, and the timeline is grouped by day and time, with each record carrying a one-line summary, marking the apps involved, and allowing one-click finding of the corresponding memory file in Finder or direct deletion.

On privacy handling, the source article says it does not take screenshots, record audio, or record system audio; temporary captured event files are kept locally for at most 48 hours, and after OpenAI's servers process them, these raw data are not retained and not used to train models, with the final generated memory files saved as local Markdown text. Users can pause collection at any time from the menu bar or settings, and can also exclude certain apps or websites. When encountering repeated workflows, it can save them as a skill—for example, if you organize press releases for several days in a row, it will ask whether to save this set of steps as a directly callable skill.

Over the past half year, several leading model makers have almost simultaneously made "faster" into a standalone product line, no longer just a parameter in a model capability table. According to the source article, Anthropic launched Fast Mode in May this year, cutting median response time from 2.8 seconds to 1.2 seconds under clearly structured prompts; Google prepared Gemini Flash Live for real-time interaction scenarios; xAI opened a separate fast endpoint for Grok.

Competition at the chip layer is more direct. According to the source article, Cerebras can run at 3,000 tokens per second on small and medium models; Groq's strength is stable low latency; SambaNova takes another path, not necessarily fastest on a single request, but more stable under high concurrency when serving many people at once, with total throughput often overtaking the previous two.

Domestic vendors are also advancing along multiple paths: ByteDance's Doubao series relies on self-built Volcano Engine inference infrastructure, emphasizing latency that is on the low side among mainstream platforms; Alibaba's Qwen-Turbo has been positioned from the start as ultra-fast, ultra-long context, and low-cost, specifically for high-concurrency scenarios; Xiaomi's MiMo-V2.5-Pro-UltraSpeed relies on FP4 quantization plus speculative decoding to push a trillion-parameter model's generation speed above 1,000 tokens per second on ordinary 8-GPU general-purpose hardware; Moonshot's Kimi K3 uses a KDA hybrid linear attention architecture, with decoding speed claimed to be 6.3x faster than conventional architectures. These figures all come from the source article's relay and are vendor-claimed.

What is confirmed is the product moves themselves: OpenAI officially announced the Ultrafast preview, and the same day ChatGPT desktop launched Computer History. Specific figures such as 14x, 750 tokens/second, 44GB SRAM, 750MW, over $10 billion, and 48-hour local retention all come from this QuantumBit report's relay, among which the agreement amount and deployment scale have no first-party original text and should be understood according to the source article's wording.

Several questions need to be retained: Ultrafast's pricing is undisclosed; 14x and 750 tokens/second are "up to" figures, and the source article gives no test conditions, task types, or comparison baseline; early user reviews come from partner personnel and are subjective feedback; Computer History's privacy promises (no retention of raw data, no use for training) are OpenAI's statements, and the source article provides no independent audit.

One observable judgment is: when Anthropic, Google, xAI, ByteDance, Alibaba, Xiaomi, and Moonshot all launch speed tiers or speedup solutions in the same time window, "fast" has gone from a differentiated advantage to an entry threshold. Whether Ultrafast can deliver on its promise of "no intelligence shrinkage" depends on whether Cerebras's capacity and scheduling can stably reproduce it under real loads, and on this point there is currently only the preview and user word of mouth, with no public benchmark. This is an inference based on the source article.