OpenAI dropped GPT-5.5 on 23 April. China’s DeepSeek released V4 the very next day. Two different strategies, two different price points. The AI race is fully on!
Two major AI model releases debuted within 24 hours of each other this week. Furthermore, both carry significant implications for the direction of the global AI industry. On 23 April 2026, OpenAI released GPT-5.5 — a model the company describes as its most capable yet, built specifically for complex, autonomous work. The following day, on 24 April, Chinese AI startup DeepSeek released a preview of its V4 model series — the most ambitious open-source AI release since the company’s R1 model shook global markets in January 2025.
Furthermore, the two releases embody competing philosophies. OpenAI is a closed-source commercial model built on massive proprietary compute. DeepSeek is open-source, runs on domestic Chinese chips, and undercuts every frontier model on price. Consequently, understanding both releases together tells you more about where AI is heading than either story alone.
What’s Happening & Why It Matters
GPT-5.5: OpenAI’s Sharpest Thinker Yet
OpenAI released GPT-5.5 — internally codenamed “Spud” — to paid ChatGPT subscribers on 23 April 2026. Furthermore, the release arrived just seven weeks after GPT-5.4, reflecting the breakneck pace of frontier model iteration. The company immediately made GPT-5.5 available in ChatGPT and its agentic coding application Codex for Plus, Pro, Business, and Enterprise users. Additionally, API access followed on 24 April after OpenAI incorporated additional cybersecurity safeguards.
OpenAI President Greg Brockman described the model in clear terms. “This is a new class of intelligence. It is a big step towards more agentic and intuitive computing,” he told reporters at a press briefing. Furthermore, Brockman described GPT-5.5 as “a faster, sharper thinker for fewer tokens” compared to GPT-5.4. He added that users can hand the model “a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going.” Consequently, the design philosophy behind GPT-5.5 is not primarily about raw benchmark scores — it is about autonomous task completion.
What GPT-5.5 Can Actually Do
OpenAI says GPT-5.5’s gains are strongest in four specific areas. Furthermore, each area reflects a real-world work pattern rather than an academic benchmark. The first is agentic coding — writing, debugging, and deploying code across complex multi-step workflows. The second is computer use — navigating software interfaces, interacting with web applications, and completing tasks across multiple tools. The third is knowledge work — reviewing documents, analysing data, and creating spreadsheets without step-by-step user guidance. The fourth is early scientific research — identifying patterns, generating hypotheses, and supporting expert workflows in fields like drug discovery.
Furthermore, OpenAI Chief Research Officer Mark Chen described the scientific research angle with specificity. “The model shows meaningful gains on scientific and technical research workflows,” he said. “We feel it could really help expert scientists make progress.” Additionally, Chen confirmed that GPT-5.5 could assist with drug discovery — an area of increasing commercial significance. Furthermore, early enterprise testers reportedly saved up to 10 hours per week by using GPT-5.5 for document review and complex coding tasks.
The performance profile is notable for what it does not sacrifice. OpenAI says GPT-5.5 matches GPT-5.4’s real-world response latency — despite being significantly more capable. Furthermore, it uses substantially fewer tokens to complete the same Codex tasks. Consequently, it is more intelligent and more cost-efficient at the task level — even though the API price is approximately twice that of GPT-5.4.
At the time of release, OpenAI reported 900 million weekly active users on ChatGPT, 50 million paid subscribers, and 9 million paying business users. Furthermore, Codex has 4 million active users. Additionally, more than 10,000 Nvidia employees received early access to GPT-5.5 through Codex before launch. Nvidia Vice President of Enterprise Computing Justin Boitano called the model capable of acting as “a chief of staff” for enterprise teams.
Safety and the Mythos Context
OpenAI released GPT-5.5 amid heightened industry attention to AI safety. Furthermore, Anthropic‘s Claude Mythos Preview — the AI cybersecurity model with zero-day vulnerability detection capabilities — had generated significant controversy just weeks earlier, including reports of unauthorised access. Consequently, OpenAI made its safety posture unusually explicit at the GPT-5.5 launch.
OpenAI Vice President of Research Mia Glaese stated: “We have a strong and longstanding strategy for our approach to cyber, and we’ve refined a durable approach to rolling out models safely. GPT-5.5 underwent extensive third-party safeguard testing and red teaming for cyber and bio risks, and we’ve been iterating on our cyber safeguards for months.” Furthermore, OpenAI evaluated GPT-5.5 across its full preparedness framework and collected feedback from nearly 200 early-access partners before release. Additionally, GPT-5.5’s chain-of-thought controllability — a measure of how well the model’s reasoning can be shaped to obscure its logic — is lower than GPT-5.4. Consequently, this makes the model’s internal reasoning more transparent to monitors, not less.
DeepSeek V4: The Open-Source Challenger Returns
One day after GPT-5.5’s launch, DeepSeek released preview versions of its V4 model series on 24 April 2026. Furthermore, the Hangzhou-based startup offered two variants: V4 Flash and V4 Pro. Both are Mixture-of-Experts (MoE) models supporting a 1-million-token context window — enough to process an entire codebase, a year’s worth of documents, or a full legislative archive in a single prompt.
The scale of V4 Pro is extraordinary. Furthermore, it contains 1.6 trillion total parameters — with 49 billion active at inference time. That makes it the largest open-weight model ever released. V4 Flash, the smaller variant, contains 284 billion total parameters with 13 billion active. Both support a Thinking mode for deeper reasoning and a Non-Thinking mode for standard queries. Additionally, both are fully open-source — published on Hugging Face with downloadable weights for local deployment and fine-tuning.
The Architecture That Makes V4 Different
DeepSeek‘s technical report highlights two core architectural innovations. Furthermore, both address fundamental efficiency constraints that most AI models struggle with at long context lengths.
The first is the Hybrid Attention Architecture — a combination of two distinct attention mechanisms, Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). This hybrid approach dramatically reduces computational cost at the 1-million-token scale. In the 1-million-token setting, V4 Pro requires only 27% of single-token inference computation and 10% of key-value cache compared to V3.2. Consequently, V4 processes vastly longer inputs at a fraction of the cost of its predecessor. The second innovation is Manifold-Constrained Hyper-Connections (mHC) — a technique that strengthens conventional residual connections between neural network layers, improving signal stability and preserving model expressivity.
DeepSeek V4’s Benchmark Performance
DeepSeek claims V4 has “almost closed the gap” with leading closed-source models on reasoning benchmarks. Furthermore, the technical data supports several specific claims. On competition mathematics benchmarks, V4-Pro-Max scores 89.8 on IMOAnswerBench — ahead of Google’s Gemini 3.1 Pro. On formal mathematics at the hardest Putnam benchmark level, V4 achieves a perfect 120/120 score — matching the best result available. Furthermore, on coding benchmarks, both V4 Flash and V4 Pro perform comparably to GPT-5.4, according to DeepSeek.
However, the models still trail on knowledge benchmarks. Furthermore, DeepSeek‘s own technical paper acknowledges a gap behind GPT-5.4 and Gemini 3.1 Pro on pure knowledge tests. The lab describes a “developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.” Consequently, V4 is not yet the equal of the very best proprietary models — but it is significantly cheaper to use. Furthermore, Neil Shah, Vice President of Research at Counterpoint Research, described V4 as offering “lower inference costs than previous models.” Additionally, Counterpoint Research Principal AI Analyst Wei Sun assessed V4’s benchmark profile as suggesting “excellent agent capability at significantly lower cost.”
The Chip Independence Question
The geopolitical dimension of V4 may ultimately be more consequential than any benchmark. Furthermore, DeepSeek trained V4 using chips from Chinese manufacturer Huawei — specifically its Ascend 950 processors — as well as chips from Cambricon. Huawei confirmed it supported DeepSeek with its “Supernode” technology, combining large clusters of Ascend chips to deliver sufficient compute. Consequently, V4 is the first major DeepSeek model where domestic Chinese hardware plays a confirmed significant role.
Wei Sun described the significance directly. “It allows AI systems to be built and deployed without relying solely on Nvidia, which is why V4 could ultimately have an even bigger impact than R1 — accelerating adoption domestically and contributing to faster global AI development overall.” Furthermore, the release came just one day after the White House accused foreign entities — primarily Chinese — of conducting “industrial-scale” campaigns to extract capabilities from US frontier AI models. The memo, issued by White House science and technology policy director Michael Kratsios, did not name DeepSeek directly. However, both Anthropic and OpenAI have previously accused the startup of “distilling” their models. Consequently, V4’s arrival lands in a politically charged environment.
TF Summary: What’s Next
OpenAI will continue rolling out GPT-5.5 to API users and expanding Codex’s agentic capabilities in the coming weeks. Furthermore, the company confirmed that GPT-5.5 Pro — the highest-reasoning-effort variant — is already available to Pro, Business, and Enterprise subscribers. Additionally, the pace of OpenAI’s releases — GPT-5.4 in early March, GPT-5.5 in late April — suggests another update before the end of Q2 2026. Brockman was explicit: “It is one step, and we expect to see many in the future.”
MY FORECAST: DeepSeek V4 is a preview. Furthermore, the company has not provided a finalisation timeline. The older deepseek-chat and deepseek-reasoner endpoints will be deprecated on 24 July 2026, creating a migration deadline for existing API users. Consequently, the coming months will see DeepSeek refining V4 based on real-world feedback — a process the lab describes as its standard approach before a full production release. Furthermore, V4’s support for Claude Code and other major agent frameworks signals DeepSeek‘s intent to embed its models deeply into developer infrastructure used by Western companies — regardless of the geopolitical climate. Morvan Su’s assessment carries a quiet warning for the West: the open-source AI gap is shrinking, and it is doing so on Chinese hardware.