AI Daily — Apr 24, 2026

Two frontier model drops in one day used to be unthinkable. Today it happened before lunch. OpenAI released GPT-5.5 six weeks after GPT-5.4; DeepSeek answered within hours with V4-Pro, a 1.6-trillion-parameter open-weights model running entirely on Huawei silicon. The competition is no longer just about who builds the smartest model — it's about who owns the workflow, the chips, and the distribution layer underneath.

• GPT-5.5 (OpenAI): 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, 1M context; $5/$30 per million input/output tokens. Live in ChatGPT and Codex; API delayed pending safeguards.
• DeepSeek-V4-Pro (DeepSeek): 1.6T parameters (49B active), 1M context, MIT license. Now the largest open-weights model available. $1.74/$3.48 per million tokens.
• DeepSeek-V4-Flash (DeepSeek): 284B total / 13B active. $0.14/$0.28 per million tokens — the cheapest frontier-adjacent model on the market.
• Nemotron 3 Super (NVIDIA): 120B / 12B active MoE, 1M context, open weights, tuned for long-horizon agent coherence.
• Codex app (OpenAI): Adds browser control, Sheets/Slides, Docs/PDFs, OS-wide dictation, and an auto-review "guardian" agent that cuts approvals on long runs.

The most important thing about GPT-5.5 isn't the benchmark numbers — it's what OpenAI is building around them.

Greg Brockman called it a step toward the company's "super app" vision, saying "what is really special about this model is how much more it can do with less guidance." The numbers back that framing: 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and 51.7% on FrontierMath Tier 1–3. More striking, Artificial Analysis placed GPT-5.5 medium on par with Claude Opus 4.7 max — at roughly a quarter of the cost.

But the real shift is that Codex shipped the same day with browser control, Sheets/Slides/Docs integration, OS-wide dictation, and an auto-review "guardian" agent. This is the chat-as-background-software play: agents that research, operate tools, and continue across long tasks without step-by-step handholding. OpenAI disclosed 4 million active Codex users and 900 million weekly ChatGPT users, per Fortune.

What changes if it works: the competitive surface shifts from model quality to who owns the work loop — scheduling, tools, integrations, governance. What failure looks like: the API delay stretches past a week because OpenAI classified GPT-5.5 at its internal "High" cyber-risk threshold, per the company's own safety disclosures. Watch the API release date. If it slips, the cyber gating is a harder problem than the launch suggested.

Timing is never accidental in this industry.

DeepSeek released V4-Pro (1.6T parameters, 49B active, 1M context) and V4-Flash (284B/13B) under MIT license, with weights live on Hugging Face within hours. Flash pricing is $0.14/$0.28 per million tokens — undercutting even GPT-5.4 Nano. Pro is $1.74/$3.48, roughly a third of GPT-5.5.

The technical detail that matters: DeepSeek Sparse Attention (DSA) plus token-wise compression materially cut KV-cache costs, making million-token contexts economical. vLLM provided day-zero support. Community reports claim ~4x compute efficiency gains and order-of-magnitude KV-cache reductions versus prior DeepSeek stacks.

The part that matters more: V4 is the first DeepSeek model with zero Nvidia CUDA dependency. It runs on Huawei's Ascend 950 clusters via the CANN framework. Counterpoint Research analyst Wei Sun argued: "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."

What changes if it works: the hardware chokepoint the US spent three years building collapses. What failure looks like: V4-Pro throughput remains compute-constrained, and DeepSeek's self-reported agentic benchmarks don't survive independent evaluation. Morningstar's Ivan Su already called V4 "competent" but "not as big a breakthrough as R1." Third-party benchmarks over the next 72 hours will settle it.

OSTP Director Michael Kratsios wrote Thursday that "the U.S. has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI." He said Beijing is using "tens of thousands of proxies and jailbreaking techniques in coordinated campaigns" to extract American breakthroughs.

The Trump-Xi summit is scheduled for May 14. V4 shipped the morning after this memo — on Huawei silicon, with open weights, MIT-licensed.

What changes if Washington acts: the first line of defense moves from chips to models themselves, which means export controls on weights, API access restrictions, and potential criminal liability for distillation activity. What failure looks like: Jensen Huang warned of a "horrible outcome" — China optimizing around Nvidia entirely — a scenario V4 makes more plausible. Watch the Trump-Xi summit readout.

For two years, DeepSeek said no to China's biggest venture firms. That's changing.

Per The Information, DeepSeek is raising over $300 million at a valuation of at least $20 billion. Tencent has proposed acquiring up to a 20% stake; DeepSeek isn't keen on ceding that much control. The purpose, according to 36Kr reporting cited by BigGo Finance, is to fund employee option repricing — countering poaching offers from tech giants that are 2–3x higher — while also supporting the complete rewrite of DeepSeek's stack from CUDA to Huawei's CANN framework.

The company that proved you could build frontier AI cheaply is discovering that staying there isn't cheap. What to watch: if Tencent closes above 10%, DeepSeek's "independent research lab" identity is effectively over.

Shopify CTO Mikhail Parakhin said on Latent Space this week that after the December model-quality inflection, AI adoption at Shopify went effectively vertical — approaching 100% daily active usage across engineering. But the bottleneck moved. It's no longer generation. It's review, CI/CD, and deployment stability.

"AI writes code on average with fewer bugs than the average human," Parakhin said, "but since they write so much more of it, more bugs make it into production." Shopify built its own PR review flow because off-the-shelf tools don't use expensive-enough models. Their solution: fewer, deeper critique loops between top-tier models, not massive parallel agent swarms.

What changes if this generalizes: the winners in enterprise AI won't be whoever has the best model — they'll be whoever has the best workflow hygiene around it. What failure looks like: companies that didn't invest in review infrastructure hit a wall where deployment cycles lengthen faster than AI speeds up coding. Watch for public metrics on rollback rates and review loads.

Overshadowed by the OpenAI-DeepSeek collision: NVIDIA released Nemotron 3 Super, a 120B/12B active hybrid MoE model with a 1M token context window, explicitly tuned for agent memory and multi-step planning. Per NVIDIA's release, it delivers over 50% higher token generation than leading open models.

This is NVIDIA trying not to be just the shovel seller. If Nemotron 3 Super gets serious traction in robotics stacks and persistent-agent deployments, NVIDIA has a seat at the application layer — and a hedge against the DeepSeek-style Huawei-native stack that explicitly routes around it. Watch GTC follow-ups.

While everyone watched GPT-5.5 and V4, Mozilla.ai's "cq" — a shared, reviewable knowledge commons for agents, framed as "Stack Overflow for agents" — spiked across developer channels this week. The idea: agents query prior "knowledge units," contribute new ones, and route items through human review before reuse.

Combined with Hugging Face's ml-intern (an open-source agent that runs the full post-training research loop — reads papers, collects datasets, launches training jobs, iterates on failures), a new layer is forming above any single vendor: cross-agent, cross-model institutional memory.

What changes if this crystallizes: the moat stops being the model and becomes the memory layer — which means governance, audit, and IP questions that enterprises haven't begun to answer. What to watch: whether a standard emerges before a vendor locks it down.

A superapp pitch sharing a release window with a model that runs on the chips America tried to ban; a memo about industrial-scale theft posted the day before the theft shipped on Hugging Face; a Shopify CTO explaining that the real AI problem is not writing code but surviving what was written.

Somewhere in Beijing, a trillion-parameter food-delivery model is learning to order itself lunch.

Stay skeptical of the benchmarks.

Forward this to the one person on your team who still thinks "the model" is the moat.