The AI Landscape: May 2026
Major activity from all of the leading labs, with significant consequences across the board.
In the second half of April, OpenAI released GPT-5.5, DeepSeek released V4 in preview, and Anthropic shipped Opus 4.7 and Claude Design within the same week. OpenAI also shut down Sora, a Discord group breached Anthropic’s restricted Mythos cybersecurity model, and the Musk v. OpenAI trial opened in an Oakland federal courthouse. Each of April’s major moves advanced its lab’s competitive position while exposing a different kind of constraint.
Consolidation mode
OpenAI released GPT-5.5 on April 23, the company’s first ground-up base model retrain since GPT-4.5. Each intervening 5.x version had layered post-training on the same foundation. OpenAI positions GPT-5.5 as agent-first, describing a system that plans, uses tools, checks its work, and continues through ambiguity without step-by-step prompting. On the Artificial Analysis Intelligence Index, GPT-5.5 scored 60, three points ahead of Claude Opus 4.7 and Gemini 3.1 Pro Preview, though the composite flattens a mixed picture in which GPT-5.5 placed second to Gemini on three of the ten underlying evaluations. API pricing doubled to $5 per million input tokens and $30 per million output tokens, but OpenAI argues that token efficiency offsets the increase, with GPT-5.5 generating roughly 40 percent fewer tokens than 5.4 on equivalent tasks and leaving net costs around 20 percent higher rather than double.
The company discontinued Sora’s web and app experiences on April 26, with the API scheduled to follow on September 24. Reports had placed Sora’s inference costs at roughly $15 million per day, with individual Pro users retaining at less than 8 percent after 30 days. The shutdown unwound a billion-dollar Disney partnership announced in December that had included a three-year contract and a planned Disney+ feature for generating content with Disney characters. Disney’s tech team reportedly learned of the change on the night of the announcement. The decision to kill a flagship product and exit consumer video generation entirely aligns with OpenAI’s positioning for a potential IPO as early as late 2026, where recurring enterprise revenue matters more to investors than consumer experiments.
OpenAI separately announced GPT-5.5-Cyber for restricted distribution to a handpicked group of cyber defenders, with Altman writing that the company would “work with the entire ecosystem and the government to figure out trusted access for cyber.” The UK’s AI Security Institute called the model one of the strongest it has tested on cyber tasks and noted that it completed one of the Institute’s multi-step attack simulations end to end, only the second system to do so. The restricted-access approach mirrors the gated distribution that Anthropic used for Mythos through Project Glasswing. On April 21, Altman had called Anthropic’s approach “fear-based marketing” designed to “keep AI in the hands of a smaller group of people.” Ten days later, OpenAI adopted the same framework.
High-velocity friction
Anthropic released Opus 4.7 on April 16 as its most capable generally available model, with benchmark gains of roughly 10 percent on SWE-Bench Verified for software engineering and 13 percent for visual reasoning. The release introduced a redesigned tokenizer that improved handling of non-Latin scripts by 20 to 35 percent in token efficiency but increased token counts on typical English workloads by 12 to 18 percent, a change that raised per-task costs even where the model performed better. Anthropic set Opus 4.7 as the new default in Claude Code and recommended an xhigh effort setting for most coding work.
Within 24 hours of launch, complaints appeared across Reddit, Hacker News, and developer forums describing Opus 4.7 as a regression in creative writing, conversational warmth, and web research quality. The model defaults to structured output such as tables and headers where 4.6 held flowing narrative, and responds more literally to prompts, punishing the imprecise instructions that 4.6 had tolerated. Some of the dissatisfaction reflects real training tradeoffs. Source attribution accuracy dropped measurably, and the model’s conversational register shifted toward something that heavy users described as “corporate”. Boris Cherny, the lead engineer on Claude Code, acknowledged that even he needed several days to adjust his workflow. That concession suggests that the model’s capabilities genuinely shifted rather than simply degraded.
Anthropic launched Claude Design on April 17, an experimental product for creating visual outputs such as prototypes, slides, and one-pagers, positioned as a complement to tools like Canva rather than a direct competitor. On the security front, a Discord group accessed Mythos through a third-party vendor on the same day Anthropic announced the model, reportedly by guessing its URL based on familiarity with Anthropic’s naming conventions for other model endpoints. The group described its intentions as curiosity-driven, but the breach undermined the controlled-distribution rationale of Project Glasswing before the initiative’s first month had ended. Separately, Anthropic removed Claude Code from its Pro plan pricing page in what it described as a test affecting 2 percent of new signups. The change appeared on a public-facing page visible to the entire internet and prompted alarm among developers before an executive clarified that existing subscribers would not be affected. Anthropic’s product velocity in April outpaced its operational communication, leaving the company to clarify retroactively rather than to set expectations in advance.
The digital decoupling
DeepSeek released V4 in preview on April 24, open-sourcing both V4-Pro and V4-Flash on Hugging Face. V4-Pro carries 1.6 trillion parameters and a 1-million-token context window. DeepSeek’s own technical report acknowledged that V4-Pro “falls marginally short of GPT-5.4 and Gemini 3.1 Pro,” characterizing the gap as a trajectory trailing frontier models by three to six months. The company launched V4 with a 75 percent price cut promotion and API compatibility designed to minimize switching costs, including drop-in compatibility with OpenAI’s API format.
DeepSeek trained V4 on Huawei Ascend and Cambricon chips, with Huawei publicly confirming on launch day that its Ascend 950 clusters had supported the model. R1, the release that put DeepSeek on the map last year, ran on NVIDIA hardware. V4 is the company’s first frontier-class release built on primarily domestic Chinese silicon. NVIDIA CEO Jensen Huang, speaking on the Dwarkesh Podcast before V4 launched, called this outcome “a horrible outcome for our nation,” arguing that pushing China outside the American hardware ecosystem would accelerate the development of an alternative stack rather than prevent it.