OpenAI cuts GPT-6 token prices in half

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

💰 OpenAI cuts GPT-6 token prices in half

🤖 Anthropic launches Claude Opus 5.5

📌 Google open-sources AI agent orchestrator

🧩 JetBrains bets big on agentic coding

🔀 Cloudflare adds Git branch previews

Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

💰 OpenAI cuts GPT-6 token prices in half LINK

OpenAI released GPT-6 Sol and Luna on Tuesday, cutting API token prices 50% versus their GPT-5.6 counterparts, landing at $2 and $0.10 per million input tokens and $10 and $0.50 per million output tokens.

The new models raise default prompt cache-hit rates and let developers change reasoning effort and tool availability without breaking the cache, so agents reuse earlier context, cutting latency and keeping the 90% cached-input read discount.

A new Prompt Caching Dashboard exposes how much input is cached over time and flags missed caching opportunities; OpenAI says GitHub cut prompt tokens needing fresh processing by over 50% across billions of requests.

🤖 Anthropic launches Claude Opus 5.5 LINK

Anthropic released Claude Opus 5.5 today, claiming Fable 5.1-level performance on most work while running about 40% cheaper than Opus 5, generating output over 30% faster and using far fewer tokens per task.

Token prices dropped 20% versus Opus 5, and subscription users get 20% larger five-hour limits that, combined with slower token burn, add up to roughly 50% more effective usage capacity on both five-hour and weekly caps.

Anthropic routes risky requests down to Opus 4.8 for most cybersecurity work and Opus 5 for biology and LLM development when safeguards fire, while approved teams can apply to its Cyber and Life Sciences Verification Programs for fuller access.

📌 Google open-sources AI agent orchestrator LINK

Google has released AX, an Apache 2.0-licensed orchestrator and declarative runtime for running autonomous AI agent workloads, hosted at agentexecutor.io and on GitHub as google/ax and built on its Agent Substrate execution layer.

AX treats each agent session as a stateful, isolated actor sandbox with strict CPU and memory limits, checkpointing and suspending idle agents awaiting model or tool responses, then resuming them in sub-second intervals with no cold-start delay.

Its control plane exposes four Kubernetes-style primitives, Task, Workspace, Gateway, and Model, managed through the Go-based ax CLI, with commands like ax apply, ax watch, ax ssh, and ax suspend or resume to control execution state.

🧩 JetBrains bets big on agentic coding LINK

JetBrains has launched Air, an open system of products for agentic software development that operates both inside and beyond its IDEs, consolidating several of the company's recent AI efforts under a single banner.

Air spans three layers: Air in JetBrains IDEs for directing and reviewing agents like Claude Code, Codex, and Junie; Air Teams for shared cloud environments and automations; and Air Governance for model access, permissions, and spending limits.

Developers reach Air through IDEs, the browser, or the command line via Air Gateway, supervising multiple agent sessions at once and inspecting changes as native IDE diffs, with Teams and Governance still in early access.

🔀 Cloudflare adds Git branch previews LINK

Cloudflare launched Worker Previews, giving every Git branch a production-like environment with its own code, configuration, URL, observability, and state, so changes can be tested before merging to production.

Running `npx wrangler preview` deploys an isolated Preview with dedicated variables, secrets, and bindings, and automatically creates a new Durable Object namespace and Container per branch to keep state changes contained.

Each branch gets a stable Preview URL, servable on a custom domain and protectable with Cloudflare Access, plus full Workers Observability traces so agents and humans can catch failures, patch, and verify iterations autonomously.

Other news you might like

• We just shipped support for the ugliest part of HTTP: VaryLINK

• Google Adds Cycle-Level Kernel Profiling to XProfLINK

• Terraform provider for Google Cloud 8.0 now generally availableLINK

• Sublime Text stays fast and focused with Python 3.14 upgradeLINK

• Console graphicsLINK

• What if your app could fix itself?LINK

• Article: Modular Edge Computing at Multi-Tenant SaaS Scale on Cloudflare WorkersLINK

🧰 Trending tools

tiun: an AI-native backend unifying authentication, payments, customer database, and analytics, letting teams ship and scale without stitching together services.LINK

Coldtea.ai: automates QA and monitoring using AI agents that catch post-deployment bugs, flag issues, and generate actionable fix tasks for your teamLINK

Kilo Code for JetBrains: an open-source AI coding agent for JetBrains IDEs that lets you build and iterate with flexible model choices across projects.LINK

Kane CLI: drives real Chrome using plain-English commands to test web apps and extract data, replacing brittle selectors and scripts for developers, QA, and AI agents.LINK

Openbase: helps developers choose npm packages by comparing popularity, maintenance activity, and reliability metrics alongside real user reviewsLINK

Harden: a free, local security tool that vets AI coding agent tool calls before execution, keeping your repo and output on your machine.LINK

📚 Trending papers & reports

Parallel coding agents often break each other's work when their changes are merged, failing in 97% of controlled tests, but sharing a note about each change recovered 82% of runs.LINK

AI agent benchmarks can be skewed by the software that runs a model, not the model itself, with scores swinging up to ~55 points and some models wrongly rated at zero when their tool-call attempts are silently blocked.LINK

Runtime guardrails boost how often AI agents reliably repeat a working solution, lifting top-tier repeated success from ~64% to ~74% without retraining, turning capabilities an agent already has into dependable delivery.LINK

WatchPoint gives coding agents a stand-in tester that clicks through and probes a live web app to diagnose failures, recovering ~58% of broken tasks, matching human testers' ~55%.LINK

Coding assistants lean too heavily on variable and function names when reading code, so misleading or stripped names push them toward wrong answers even when the actual program logic is unchanged.LINK

See you tomorrow for a new dose of ☕️ Devshot!