August 24, 2026·6 min read
AWS releases aws-bench for testing agents
AWS agent benchmark, OpenAI's price cut, JDK 27 previews, and more.
Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
☁️ AWS releases aws-bench for testing agents
💸 OpenAI cuts GPT-5.6 Sol price 20%
🔌 Espressif open-sources IoT cloud platform
⚡ VoidZero launches Vite+ beta
☕ JDK 27 and 28 early features revealed
🕵️ Anonymous coding model Ox Alpha appears
Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.
☁️ AWS releases aws-bench for testing agents LINK
AWS has open-sourced aws-bench under Apache-2.0, a benchmark that evaluates how accurately AI agents complete real AWS tasks like diagnosing misconfigurations, provisioning infrastructure, and operating live cloud environments in disposable accounts.
Each scenario deploys AWS resources via CDK stacks in isolated accounts, then runs the agent in a sandboxed container with scoped credentials before scoring the result using either an LLM judge or programmatic checks against live state.
Built on the Harbor framework, it ships adapters for Claude Code, Codex, Kiro CLI, and Mini-SWE-Agent, but requires management-account access, is pinned to us-east-1, may incur costs, and lacks published baselines or a leaderboard.
💸 OpenAI cuts GPT-5.6 Sol price 20% LINK
OpenAI cut API pricing for its frontier GPT-5.6 Sol model by more than 20% for the next three months, responding to competition from Anthropic and Chinese AI models.
Standard short-context Sol now costs $4 per million input tokens and $20 per million output tokens, down from $5 and $30, with the discounts also applying to ChatGPT Work and Codex credits.
Subscription pricing for Pro, Plus, and Business plans stays unchanged, and the move follows last month's cuts of 20% on mid-tier GPT-5.6 Terra and 80% on the cheaper Luna model.
🔌 Espressif open-sources IoT cloud platform LINK
Espressif has open-sourced its entire IoT stack with ESP RainMaker Neo, releasing the previously proprietary cloud backend under Apache 2.0 so teams can deploy it inside their own AWS account.
The GitHub release spans four components: a Go cloud backend and admin dashboard, a device firmware SDK for ESP32-series SoCs, a TypeScript phone-app SDK, and a reference mobile app for iOS and Android.
Neo drops Classic's custom infrastructure for native AWS IoT services, Device Shadow, Thing Groups, AWS IoT OTA, and IAM access control, running serverless with pay-as-you-go pricing, though the two platforms aren't deployment-compatible and migration isn't recommended.
⚡ VoidZero launches Vite+ beta LINK
VoidZero, the company behind Vite from Evan You, has shipped the beta of Vite+, an MIT-licensed toolchain that puts the runtime, package manager, and core frontend tools behind a single command.
It ties together Vite 8, Vitest, Rolldown, tsdown, Oxlint, and Oxfmt with a cache-aware, dependency-aware task runner, so one workflow handles dev server, format, lint, type check, tests, builds, and monorepo tasks.
Since the alpha the team merged over 500 pull requests across a dozen releases, adding smarter caching, enterprise features, and 180+ fixes, with more than 1,300 public repos already depending on it, including Dify and BlockNote.
☕ JDK 27 and 28 early features revealed LINK
JDK 27 has reached its first release candidate with a locked set of nine JEPs, while JDK 28's early feature list is taking shape, with the non-LTS JDK 27 due for formal release on September 15, 2026.
JDK 27 makes G1 the default garbage collector in all environments via JEP 523, and advances previews including primitive types in patterns and switch, structured concurrency, PEM cryptographic encodings, and a twelfth Vector API incubation.
JDK 28, targeted for March 2027, adds a standard JSON parsing API without external libraries, deprecates the macOS/x64 port, defaults Shenandoah GC to generational mode, and previews value objects and strict field initialization.
🕵️ Anonymous coding model Ox Alpha appears LINK
Since August 20, an anonymous provider has offered free access to a frontier-class coding model called "Ox Alpha" on OpenRouter, and no company has claimed to have built it.
The model handles text, images and video with a ~1M-token context window and 131,072-token output cap, and fingerprinting by developer unclecode's modelprint tool matched it to GLM-5.3 on six of nine infrastructure probes.
OpenRouter's free route ran through today, August 24, while OpenCode advertised 100 trillion tokens a day with near-unlimited rate limits, though full DeepSWE benchmark runs put Ox Alpha roughly level with GPT-5.6-sol rather than ahead.
Other news you might like
- 80% of developers find AI coding more addictive than helpfulLINK
- Walk through a 3D cyberpunk city built purely from ASCII characters — a text-based metropolis runs on a 283KB Rust WebAssembly engine feeding a WebGL rendererLINK
- Java’s String.indexOf can be slow (quadratic)LINK
- Nvidia just showed that the harness, not the AI model, is now the real heroLINK
- Linux kernel maintainers overwhelmed by surge of AI-generated patchesLINK
- Canonical backs quest to translate mountains of C into safe Rust with AILINK
🧰 Trending tools
Sim: build and deploy AI agents via visual or code workspace, with 1,000+ integrations and support for all major LLM providers.LINK
Prelint: validates AI-generated pull requests against your ADRs, docs, and past decisions to catch product drift before it reaches production.LINK
SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate and hear responses while coding hands-free.LINK
Dograh: an open-source voice agent platform with a visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30+ model integrations.LINK
Zro: routes coding requests to open-source models like GLM-5.2, DeepSeek V4 Flash, and Kimi K3 through one endpoint without storing your data.LINK
AgentSky: runs managed AI coding agents with auto-recovery, controllable from WhatsApp, Telegram, Slack, web, API, or CLI for remote development workflows.LINK
📚 Trending papers & reports
Coding-assistant memory tracks when a fact changes during a session, so a renamed function or moved endpoint gets the current answer 91% of the time versus ~58% for standard retrieval, which serves the outdated value roughly a third of the time.LINK
Prompt formatting can make coding assistants faster and more stable just by structuring requests cleanly, with JSON working best, while letting the model rewrite its own prompts actually hurt results.LINK
AI coding assistants readily invent fake software libraries, accepting made-up names in up to 99% of tasks and inventing them after a single typo in up to 26%, opening the door to supply-chain attacks.LINK
Vibe coding security shows that AI agents building software from prompts produce mostly insecure code, with a top setup getting 57% of tasks working but only ~12% secure, and safety hints don't help.LINK
AI code graders can be swayed by cosmetic tweaks like variable names, comments, or formatting, handing functionally identical programs inflated or unfairly low scores across five languages, a reliability gap that persists even with added safeguards.LINK
See you tomorrow for a new dose of ☕️ Devshot!
