Opus 4.7 and GPT-5.5 a week apart, Cursor 3.0 makes parallel agents the interface, DeepSeek-V4 brings 1M context as open weights, and Anthropic dissects its own regression.
In April the whole stack reorganized around agents running in parallel: frontier labs shipped models for long coding horizons, the IDEs made multi-agent orchestration the default interface, and the open-weights labs closed the gap at a fraction of the cost. Every link goes to the primary source.
Built for long-running engineering tasks: vision up to 2576px for computer-use agents reading dense screenshots, and a new xhigh effort level. Pricing unchanged at $5/$25 per million tokens.
More literal instruction following costs existing prompts and harnesses a re-tune.
A new flagship with gains in agentic coding and reasoning: 82.7 percent on Terminal-Bench 2.0 and 51.7 on FrontierMath tiers 1 to 3.
Rolled into Codex the same month, which shifts the default target of most agent harnesses.
Two open mixture-of-experts models, V4-Pro at 1.6T total and V4-Flash at 284B, with a 1M-token default context and compatibility with both OpenAI and Anthropic API formats.
A deadline comes attached: the legacy deepseek-chat and deepseek-reasoner retire on July 24.
A 1T mixture-of-experts model with 32B active under a modified MIT license, built for long agentic coding sessions, with an Agent Swarm system reported to scale to 300 sub-agents.
Reported to match GPT-5.5 on SWE-Bench Pro at much lower cost.
Four variants under Apache 2.0, from E2B to a 26B MoE, built from Gemini 3 research with native support for agentic workflows and edge deployment.
A permissively licensed local model for planning and tool use, if that is what you need.
The new Agents Window runs many agents in parallel, locally, in git worktrees, in the cloud, and over SSH, plus /worktree, /best-of-n across several models, Design Mode for browser-level UI annotation, and an Await tool that lets agents block on background processes.
Orchestrating several agents stops being a side feature and becomes the window you work in.
Computer use on macOS, GitHub PR reviews, SSH to remote devboxes, memory of preferences, and thread automations for recurring work. OpenAI cites over 3 million weekly developers.
The built-in browser means frontend work gets verified rather than assumed.
Build your own agents on the same runtime, harness, and models that power Cursor.
Direct competition with the Claude Agent SDK and Codex as an embeddable agent platform.
A research preview that generates prototypes, decks, and UI mockups from prompts. Exports to HTML, PPTX, Canva, or straight into Claude Code.
It reads a linked repo's actual React components, so output lands inside your design system rather than beside it.
Google repositioned Vertex AI around the Gemini Enterprise Agent Platform for building and governing autonomous multi-step agents, and announced eighth-generation TPUs sized for agentic inference loads.
Vertex AI is now organized around governing autonomous agents, not around serving models.
An unusually detailed postmortem: a lowered reasoning-effort default, a caching bug that cleared reasoning history every turn, and a system-prompt verbosity cap that measurably cut coding quality.
Required reading on how small harness changes degrade agent output.
About 3.5 gigawatts of additional TPU capacity from 2027, alongside the disclosure that Anthropic's run-rate revenue passed 30 billion dollars.
A capacity signal: Claude inference and rate limits are being underwritten years out.
