DeepSeek V4, GPT-5.5 and the week that shook up the models

Today’s most important AI news

In less than 48 hours, two frontier models launched: DeepSeek V4 and GPT-5.5. Meanwhile, Anthropic had a rough week: Claude 4.7 is ignoring control instructions, Claude Code piled up reports of degraded quality, and a growing number of users are cancelling their subscriptions. The pattern of the day isn’t a single star launch — it’s an active reshuffling of the model board that forces a reassessment of which tool to use for what.

DeepSeek V4 and GPT-5.5: two different bets on what a model needs in 2026

DeepSeek published a preview of V4, its new flagship model, on Friday, and the most relevant detail isn’t the reasoning benchmark but the context: V4 can process substantially longer text windows than its previous generation thanks to an architecture redesigned to handle large volumes of text more efficiently. The Hugging Face team described it as “a million tokens of context that agents can actually use” — an important distinction, because long contexts on paper don’t always survive real use in production pipelines. As with the company’s previous models, V4 is open source, which means any team can deploy, fine-tune or audit it without depending on an external API.

OpenAI, for its part, released GPT-5.5 and GPT-5.5 Pro directly in the API. With more than 1,500 points on Hacker News and over 1,000 comments, the announcement sparked an intense technical conversation, although public information about the model’s specific improvements is still scarce. What is clear is the strategic move: OpenAI keeps betting on intermediate versions that extend the commercial life of its GPT-5 line until the next architectural leap arrives.

For a product manager or developer who has to decide today which model to integrate, the practical question is about access and cost, not just capability. Open-source DeepSeek V4 lowers the barrier to entry for teams that don’t want to depend on external APIs or that operate in regulated environments. GPT-5.5 via OpenAI’s API offers the infrastructure and support many teams already have integrated. They are two models that compete with each other less than it seems — and more against each organization’s status quo.

Anthropic’s tough week: Claude ignores instructions and the community notices

Three warning signs about Claude arrived almost simultaneously. First, a Hacker News thread documented that Claude 4.7 is systematically ignoring stop hooks — instructions that let developers inject determinism into their workflows, such as preventing the model from stopping if it modified a source file without running tests. Second, Anthropic published an official postmortem on the reports of degraded quality in Claude Code during the week of April 23, with 924 points and 708 comments on HN — a discussion that rarely happens when everything is working well. Third, a post titled “I cancelled Claude” about token, quality and support issues racked up 888 points and 526 comments.

Anthropic’s postmortem is the most important signal of the three. The fact that the company published it proactively suggests the problem was serious enough to require public communication, not just a silent fix. Claude Code has positioned itself as one of the most widely used tools for agent-assisted programming, and degradation in that layer directly affects development teams that have it built into their daily workflows.

For anyone using Claude in production, the implication is concrete: if you have workflows that depend on hooks or on Claude Code’s consistency, it’s worth reviewing this week’s logs and reading the postmortem before assuming the current behavior is the expected one. A model’s reliability isn’t measured only in benchmarks — it’s measured in whether it does what you told it to do.

Agents that write their own memory: the wuphf project and the infrastructure being built underneath

Less visible than the model launches, but potentially more important for developers building agentic systems, is a project posted on Hacker News that implements an agent-maintained wiki layer using Markdown and Git as the source of truth, with a BM25 index over SQLite. No vector databases, no graphs — just versioned text files that the agent can read and write between sessions.

The idea the project articulates is the one Andrej Karpathy has been describing for months: an LLM-native knowledge substrate where context accumulates across sessions instead of being re-pasted every morning. The problem it solves is real: today most agents are amnesiac between runs, which forces developers to reinvent memory mechanisms in every project. A Git wiki that the agent actively maintains is portable, auditable and requires no additional infrastructure.

It’s not a packaged enterprise solution — it’s a proof of concept that runs locally in ~/.wuphf/wiki/. But for a developer designing an agent’s memory architecture today, it’s the kind of pattern worth understanding before committing to a heavier solution.

In a nutshell

What looks today like a pile of disconnected news has a common thread: the gap between what models promise and what they deliver in production is becoming the industry’s central theme. DeepSeek V4 and GPT-5.5 compete on paper capability; Claude 4.7’s problems with hooks and the Claude Code crisis compete on real reliability. The model market is maturing enough that “it works in the demo” is no longer enough — and that pressure will define which platforms survive in next year’s production stacks.

---

Sources used: MIT Technology Review — https://www.technologyreview.com/2026/04/24/1136422/why-deepseeks-v4-matters/, Hugging Face Blog — https://huggingface.co/blog/deepseekv4, OpenAI — https://openai.com/index/introducing-gpt-5-5/, Hacker News (Claude stop hooks) — https://news.ycombinator.com/item?id=47895029, Anthropic Engineering Postmortem — https://www.anthropic.com/engineering/april-23-postmortem, Hacker News (Claude cancellations) — https://nickyreinert.de/en/2026/2026-04-24-claude-critics/, GitHub wuphf — https://github.com/nex-crm/wuphf