☕️ OpenAI's Ultrafast mode runs GPT-5.6 faster

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

⚡ OpenAI's Ultrafast mode runs GPT-5.6 faster

🤖 Gemini 3.7 Flash cuts price 50%

🛠️ Vercel opens its v0 AI app builder to developers

🔓 Node.js creator open-sources Durable Objects

📦 npm 12 disables install scripts by default

🇨🇳 China's Z.ai launches an open coding model to rival OpenAI

Plus: 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

⚡ OpenAI's Ultrafast mode runs GPT-5.6 faster LINK

OpenAI has launched GPT-5.6 Sol Ultrafast mode, which runs its frontier model at up to 750 output tokens per second, roughly 14x quicker than Standard processing, without dropping to a smaller or less capable model.

The speedup comes from Cerebras' Wafer-Scale Engine and its 44GB of on-chip SRAM, cutting the data-movement bottlenecks that cap inference on GPU systems; it cleared Humanity's Last Exam's 2,500 questions in about 11 hours versus 78 for Fable 5.

OpenAI pitches Ultrafast for incident response, coding, real-time support, and research iteration, and uses it internally to analyze logs during incidents, but Cerebras capacity is limited, so interested customers must apply and be vetted for workload fit.

🤖 Gemini 3.7 Flash cuts price 50% LINK

Google's Gemini 3.7 Flash ships at half the price of its predecessor while posting broad gains over 3.6 Flash across coding, web development, and reasoning over knowledge-heavy documents.

On coding, it hits higher first-pass accuracy and stronger debugging and issue resolution, scoring 43.6% versus 34.4% on FrontierCode 1.1 Main and 65.3% versus 49.0% on DeepSWE v1.1.

For UI generation it matches reference screenshots or design systems more closely, reaching a 1588 Elo on WebDev Arena, and jumps to 34% on the GDP.pdf document benchmark from 22%.

🛠️ Vercel opens its v0 AI app builder to developers LINK

Vercel has moved its v0 API to general availability, giving developers programmatic access to the AI app-building agent to send prompts, generate and edit application files, run them in a Sandbox, and get an embeddable preview URL.

Each app is organized around a chat that tracks state, so follow-up messages reuse the same chat ID, while synchronous, asynchronous, and streaming requests expose individual agent actions like file edits, searches, and Bash commands.

Apps can be created from GitHub repos, ZIP archives, or file sets, then deployed to a Vercel project, with MCP servers, up to three skills, and integrations for invoking v0 inside other agents via the AI SDK or eve framework.

🔓 Node.js creator open-sources Durable Objects LINK

Ryan Dahl, creator of Node.js, has open-sourced celld, a self-hosted implementation of Cloudflare's Durable Objects and Workers that stays compatible with their JavaScript APIs while removing any dependence on Cloudflare's backend infrastructure.

Written in Rust and JavaScript under Apache 2, celld swaps Cloudflare's backend for S3-compatible storage and the Tokio runtime, gives each object its own SQLite copy, and runs JavaScript, TypeScript, and WebAssembly-compiled Rust, Go, or Zig.

Dahl estimates 100 resident cells cost ~$49/month on a DigitalOcean droplet versus $415 on Cloudflare, though Cloudflare counters that idle Durable Objects run about $21; notably, the GitHub page bans AI-generated contributions.

📦 npm 12 disables install scripts by default LINK

npm 12 now blocks dependency install and lifecycle scripts from running by default, requiring developers to review pending scripts, approve trusted ones, and commit the resulting allowlist in package.json before builds proceed.

The release also blocks implicit builds for packages with a build file, plus scripts from git, file, link and https tarball dependencies, and closes a path where a Git dependency could override the Git executable.

Available behind warnings since npm 11.16.0, the change trails pnpm, yarn 4.10 and bun 1.3; researchers warn approval fatigue could turn deny-by-default into a click-through prompt, as scripts drove ~53% of malicious npm attacks last year.

🇨🇳 China's Z.ai launches an open coding model to rival OpenAI LINK

Zhipu has released GLM-5.3, which it claims is now the strongest open-weights coding model, with the largest performance gains showing up in agent-based coding tasks compared to its predecessor.

Built on the same base as GLM-5.2, all improvements come from extended post-training, including cybersecurity data that let the model chain exploitation steps and uncover 2,436 vulnerabilities across 269 projects, some 40 years old.

The model is available now via the GLM Coding Plan and works with agents like ZCode, Claude Code, and OpenCode, with weights slated to go open source in two weeks once security reviews finish.

Other news you might like

• LLM-Generated GraphQL Mocks Arrive at Airbnb and Expedia, While the Spec Lags BehindLINK

• Packer v1.16.0 brings verifiable provenance to machine imagesLINK

• Rx.NET 7.0 Reduces Deployment Size by Splitting Windows UI SupportLINK

• Three Claude agents given conflicting orders sabotaged each other on a shared server, then didn't tell users what they'd doneLINK

• Writer introduces new AI model and upgraded harness to contain token costsLINK

🧰 Trending tools

Sim: a visual and code-based workspace for building and deploying AI agents, connecting to 1,000+ integrations and every major LLM provider.LINK

Prelint: checks AI-generated pull requests against your ADRs, docs, and prior decisions to catch product drift before it ships.LINK

SKI: lets you voice-code with Claude Code, Codex, and other agents, getting spoken replies back so you build hands-free at thinking speed.LINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs so they know when to speak, wait, or interrupt naturally.LINK

Zro: routes coding requests across open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through a single endpoint without storing data.LINK

AgentSky: launches managed AI agents like Claude Code, Codex, Hermes, and OpenClaw with automatic recovery, accessible via WhatsApp, Telegram, Slack, web, API, or CLI.LINK

📚 Trending papers & reports

Coding-agent benchmarks that only check if a command's final result matches can hide huge swings, one model's score looked flat but actually masked a ~64 point drop from a transport bug and a ~61 point recovery, meaning today's agent leaderboards may rank tools wrong.LINK

AI coding agents are increasingly limited not by task-solving skill but by weak collaboration with human users, so the paper urges building agents optimized for alignment, verifiability, steerability, and adaptability instead of pure autonomy.LINK

Website training data for web-browsing bots is auto-generated by first mapping a site's actual pages and functions, cutting made-up tasks and producing more reliable practice runs, which boosts performance on both familiar and new websites.LINK

AI-led code refactoring let an unsupervised coding agent rewrite a core rule across 189 files of a 717,000-line app, catching 201 bugs itself before any human ever ran the program.LINK

Code AI memorization tests break down at scale, since the usual tricks for spotting whether a coding model just memorized answers stop working on larger models, making contamination scores unreliable and generalization the more important thing to measure instead.LINK

See you tomorrow for a new dose of ☕️ Devshot!