☕️ Google launches Gemini 4 Argon

Hi there, this is your daily ☕️ DevShot.

In today's DevShot:

🔷 Google launches Gemini 4 Argon

🤖 OpenAI turns ChatGPT into an app platform

📦 Cloudflare Containers now start 6x faster

🧠 OpenAI launches Decisions API

🦆 AWS embeds DuckDB in Aurora PostgreSQL

🗃️ Cursor hits 300 Git pushes per second

Plus: 🎁 8 more stories you might like, 🛠️ 5 tutorials, 🧰 7 tools, and 📚 5 papers.

• Google has rolled out Gemini 4 Argon, its next-generation frontier model, first to cybersecurity partners under its Fairwinds program, with Sundar Pichai citing strong internal use across coding, cyber defense, and quantum computing.
• Pricing undercuts rivals at $2 per million input tokens and $10 per million output tokens, with cached input heavily discounted, while Google claims parity or an edge over Anthropic's and OpenAI's top models.
• Google highlighted use cases including quantum algorithmic optimization of qubits and gates, fleetwide data-center telemetry analysis, and large-scale codebase migration, with Argon already deployed in its security business and at Wiz.

• OpenAI opened ChatGPT to third-party apps, letting developers build app-like experiences via plugin extensions that get a dedicated home in the sidebar, unveiled at its Dev Day presentation on Tuesday.
• Beyond existing plugins that wire ChatGPT to tools like Slack and Google Drive, the extensions add interactive panels and file viewers for custom formats, so users work inside their tools while chatting; a new Plugin Creator handles building.
• Submission runs through a redesigned directory flow with clearer feedback and improved ranking, users approve each access request individually, ChatGPT Sites can host the plugins, and the proposed MCP Events spec lets plugins trigger automations from connected-app events.

• Cloudflare rearchitected its Containers for on-demand agent workloads, so code can now pick each sandbox's image and instance type at runtime, startup is roughly 6x faster, and filesystem snapshots are in public beta.
• The new durable_object scheduling policy starts Containers wherever the Durable Object already runs, favoring hosts that cache the image, and restores a prepared VM instead of booting fresh, cutting median startup from about four seconds to 648ms in ComputeSDK's benchmark.
• Rollouts become plain code, canary, pin, or roll back by choosing which image future starts use, with no config push or drain, and the new cloudflare/debian-trixie image plus snapshots let agents skip Dockerfile builds and resume saved workspaces across sessions.

• OpenAI unveiled its Decisions API at DevDay on Tuesday, built on its smallest and cheapest Luna model, offering a purpose-built alternative to general chat models for classification and routing tasks.
• Instead of prose, the API returns a set of developer-supplied answers with realistic confidence scores, and it responds in about 150 milliseconds versus roughly 1.6 seconds for GPT-6 Luna on the same kind of task.
• It sits between prompt-based classification and training a small classifier, taking new labels in the prompt without retraining, but it's only in limited preview, and per-call pricing, candidate limits, and fine-tuning support remain unannounced.

• AWS has embedded the DuckDB analytical engine inside Aurora PostgreSQL, letting applications query Apache Iceberg and Parquet data in S3 alongside live transactions without copying data or building reverse ETL pipelines.
• DuckDB runs analytical scans within Aurora, so a single query can join operational records, including uncommitted writes, with data lake tables, while Aurora filters columns, caches hot data, and exposes metrics like rows scanned and S3 bytes read.
• Enabled via the aurora_analytics extension and an IAM role, the feature works on Aurora PostgreSQL 17.11 and 18.6 in all commercial regions at no extra charge, with customers paying only for the compute and S3 requests their queries use.

• Cursor built Continuity, a Git storage architecture that treats an S3-backed write-ahead log as the source of truth instead of coordinating consistency across replicas, hitting more than 300 Git pushes per second in its tests.
• Pushed data lands in S3 and the reference update is recorded in the WAL, with a push acknowledged only after persistence; local NVMe repositories act as warm caches, and any server can accept a push via atomic compare-and-swap on S3.
• Synthetic tests on Cursor's everysphere monorepo reached ~120 pushes per second on S3 Standard and over 300 on S3 Express One Zone, where Git compaction became the bottleneck, though the published figures remain independently unverified.