OpenAI is previewing "Ultrafast," an API-only tier that runs full-size GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second (14x standard). Intelligence stays the same; only speed is carved out into a separately priced tier. It helps agents' back-and-forth loops, but this developer-focuse
Pillar Security's "Week of Sandbox Escapes" disclosed seven holes across Cursor, Codex, Gemini CLI, and Antigravity. In every case the sandbox itself was never broken — tools outside the box trusted and ran the files the agent wrote. Here's a developer-focused look at the four weakness patterns and
Since August, new Claude models have begun embedding an invisible statistical watermark in generated text. A developer's-eye look at how the SynthID-Text approach works, which writing it marks and which it slips through, and the danger of confusing "AI processed it" with "AI wrote it."
On August 13, Google released Gemini 3.7 Flash, a new model for coding agents. It arrives three weeks after the prior generation at an introductory price of roughly half—but with a tag set to double in January 2027, and performance still resting on Google's own claims. A look at the economics of the
On August 16, DeepSeek shifted its API pricing to a two-tier peak/off-peak structure. Billed as "off-peak half price," the baseline is the new peak rate—compared with current pricing, output runs up to about 4.7x higher. Here's what it means for coding-agent execution costs and how to think about th
Cursor's newly unveiled "builds" prepares development environments in the background—dependencies installed and setup done—then forks an already-successful machine to launch. Time to first token is roughly 3x faster, and if an environment breaks, agents keep running on the most recent healthy build.
Claude Code v2.1.233 is out. We break down the changes that firm up unattended, self-hosted operation: GitLab merge request support, faster self-hosted runner startup, a memory cap to stop runaway builds on Linux, and a fix for a credential-leak path on Windows.
Claude Code v2.1.232 changes the defaults. Subagents now fork by inheriting the entire conversation and cache, and you can name another running session with @. Parallel work gets lighter, but watch out for blind spots and changes to permissions and settings.
DeepSeek released V4 Pro "0813" as GA. It touts breakthrough pricing of $0.87 per million tokens and top-tier coding scores — but most of that is self-reported, with third-party verification still pending. Here's what developers should confirm before adopting it.
Claude Code v2.1.229 removes destructive git flags—such as --force, --amend, and --no-verify—from auto-approval in the commands generated by /commit-push-pr. It also covers new safeguards for unattended operation and cost optimizations for workflows.
Grok Bot, released by SpaceXAI on August 11, is an always-on agent that owns a dedicated cloud computer and logs into your apps just like a human to get work done. We break down its pricing and mechanics as a passenger on the Cursor subscription, along with the blind spot of shared credentials.
On August 11, Anthropic expanded its Compliance API for auditing so that session records from Claude Code and Cowork running on employees' machines can be read out. Agent conversations that once stayed local now fall within the scope of corporate audits — here's how it works, and the tension between
