Anthropic shipped Opus 4.7 and withheld Mythos. OpenAI pushed GPT‑5.5. Alibaba and DeepSeek turned up the open‑weight pressure. The through‑line this week: long‑horizon agents, coding, and the operational scaffolding around them.
Anthropic released claude-opus-4-7 across all Claude products and the API. The update emphasizes stronger coding, better performance on long‑running software tasks, and higher‑resolution vision. Pricing holds steady at $5 / $25 per million input/output tokens — same as Opus 4.6.
Anthropic built — and chose not to broadly release — Claude Mythos Preview. Citing cyber risk (notably the model's ability to identify vulnerabilities in software), Anthropic is restricting access to a limited group of cybersecurity partners via Project Glasswing.
This is the clearest signal to date that frontier labs will publicly withhold capabilities on security grounds — and name the program.
A new hosted service on the Claude Platform designed for long‑horizon agent work. It provides stable interfaces for sessions, harnesses, and sandboxes, with durable state and safer tool access — aimed at teams running agents beyond a single turn.
Claude Design is a new experimental surface for generating prototypes, slides, and one‑pagers. Alongside it: Claude can now render custom charts and diagrams in‑line in responses, and the mobile app can display fully interactive, shareable assets in chat.
Claude Code shipped a broad CLI update with a native binary launcher, stronger sandbox and permission safeguards, and smoother workflows. Separately, there was some noise over Claude Code pricing on the $20 Pro tier — Anthropic clarified it is not moving to a flat $100/month; see Simon Willison's write‑up for the plain‑English version.
Anthropic also reaffirmed that Claude will remain ad‑free, arguing advertising incentives are incompatible with a genuinely helpful assistant. Read that as a positioning shot across the bow.
OpenAI released GPT‑5.5 to paid tiers (Plus, Pro, Business, Enterprise) in ChatGPT and Codex. It's positioned as an agentic model designed to work through complex tasks by switching between multiple tools — better at coding, computer use, and deep research. It arrives under two months after GPT‑5.4, underscoring the current pace of frontier releases.
OpenAI expanded Codex from “code assistant” toward a general agentic coder across surfaces. Paired with GPT‑5.5 two days later, this sets up a narrative around ChatGPT as a multi‑tool super app — a framing TechCrunch and others picked up immediately.
Alibaba released its most powerful model to date. Qwen 3.6‑Max‑Preview tops six major coding benchmarks and posts gains in world knowledge and instruction following over its predecessor. Max‑Preview is proprietary; the wider Qwen 3.6 family includes open‑weight variants (Apache 2.0 for the 35B‑A3B model).
Alibaba's Qwen team dropped a 27B‑parameter dense open‑weight model that outperforms a 397B MoE on agentic coding benchmarks. It's a pointed argument that architecture and training quality still beat parameter count — and an easier deploy target for teams that can't serve a trillion‑parameter router.
DeepSeek launched the V4 Preview as two open‑weight MoE models: V4‑Pro (1.6T total / 49B active) and V4‑Flash (284B total / 13B active). Both ship with 1M‑token context. V4‑Pro leads all open rivals on math and coding, and trails only Gemini 3.1‑Pro on world knowledge. Available on Hugging Face, the DeepSeek API, and chat.deepseek.com (Expert = Pro, Instant = Flash).
| Model | Released | Params | License | Notes |
|---|---|---|---|---|
| Claude Opus 4.7 | Apr 22 | — | closed | Coding + long‑horizon + vision; $5/$25 per M |
| GPT‑5.5 | Apr 23 | — | closed | Agentic tool switching; Plus / Pro / Biz / Ent |
| Qwen 3.6‑Max‑Preview | Apr 20 | — | preview | SOTA on 6 coding benchmarks |
| Qwen 3.6‑27B | Apr 22 | 27B dense | open | Beats 397B MoE on agentic coding |
| DeepSeek V4‑Pro | Apr 24 | 1.6T / 49B active | Apache 2.0 | 1M context; math/coding leader among open models |
| DeepSeek V4‑Flash | Apr 24 | 284B / 13B active | Apache 2.0 | 1M context; the cheap‑and‑fast tier |
Jones's thesis, from his breakdown of this week's Codex push: the real product isn't the model anymore — it's the harness around it (sessions, sandboxes, tool access, computer use). OpenAI and Anthropic are converging on the same architecture and splitting sharply on the business model. One sells the runtime. The other gives it away and taxes the tokens.
Jones's second — and sharper — point: the two labs disagree on how an agent should reach into the world. It's not a pricing fight; it's a trust model.
The rough translation of Jones's take: OpenAI bets the model is smart enough to operate human interfaces safely. Anthropic bets the interfaces themselves should be redesigned for models — structured, permissioned, intermediated — because the blast radius of a wrong click in a general agent loop is too high.
The biggest behavior change in Opus 4.7 isn't a benchmark number — it's that the model no longer silently fills in what you meant. If your Opus 4.6 prompt worked partly because the model inferred follow‑on intent, that intent now needs to be written down. Hedges like “try to,” “if possible,” and “you might want to” used to read as firm guidance. In 4.7 they read as hedges — and the model obliges.
Claude Design is an Anthropic Labs surface that turns a prompt into a polished prototype, slide deck, wireframe, or one‑pager — on‑brand, interactive, and shareable. Powered by Opus 4.7, available in research preview to Claude Pro, Max, Team, and Enterprise. It is not trying to replace Figma for dedicated designers. It's aimed at everyone who currently blocks a designer for twenty minutes to get a mockup, or fakes it in Keynote.
Claude Design is the first Anthropic product that treats design as an outcome, not a tool category. It's a direct bet that non‑designers generate 80% of the visual artifacts inside a company — and that the current Figma‑shaped hole in their day is worth a product. Figma's response will tell you whether “design tool” survives as a distinct SaaS category or collapses into the chat window.
Available on desktop and iOS in Australia, Indonesia, Japan, the Philippines, Singapore, South Korea, and Vietnam (iOS excluded in Japan). The agentic browser‑control feature remains US‑only on AI Pro / Ultra.
At Cloud Next '26 (Apr 22), Deloitte announced a dedicated agentic transformation practice built on Gemini Enterprise — a sign of the big consultancies aligning around Google's agent stack.
Anthropic's decision to hold Mythos has become a reference point for government and industry discussions on dual‑use frontier models — expect policy language to cite it within weeks.
The week's ambient backdrop: infrastructure leaders gathering to talk power, cooling, and capacity — the physical tax on every model above.
The agent stack is becoming a product category. Managed Agents (Anthropic), Codex‑for‑everything (OpenAI), Qwen's dense agentic coder, Gemini's browser agent, Deloitte's Gemini practice — five vendors, same framing. Expect the 2026 developer conversation to shift from “which model” to “which agent runtime.”