// the wire
what just happened in AI coding.
launches, model drops, pricing changes, and the occasional drama — reported fast, explained like you're a builder, not a shareholder.
- ·4 minkimimoonshot-aiopen-weightsmodelscodingbenchmark
kimi k3 weights are out — 2.8 trillion parameters, modified MIT, and a 51% hallucination rate you should know about
Moonshot AI released the full Kimi K3 weights on Hugging Face on July 27, 2026 — the world's first open 3T-class model at 2.8T parameters. Independent testing by Artificial Analysis found a hallucination rate of 51%, up from 39% on the previous generation. Here's what's real and what to watch.
read it → - ·3 minclaudeanthropicsonnet-5modelsswe-benchcodingpricing
claude sonnet 5 is the new default — 72.7% swe-bench, intro pricing through august
Anthropic released Claude Sonnet 5 on June 30, 2026 as the new default for Free and Pro plans. It scores 72.7% on SWE-bench Verified and 76.1% on Terminal-bench — up from 62.3% and 55.4% on Sonnet 4.6, respectively. Intro API pricing is $2/$10 per million tokens through August 31, then moves to $3/$15.
read it → - ·4 minawskiroidespec-drivenagentstoolingamazon-q
aws kiro is generally available — spec-driven development replaces amazon q developer
AWS Kiro launched GA on May 7, 2026, replacing Amazon Q Developer as AWS's primary coding IDE. The differentiator is spec-driven development: before any code is written, Kiro formalizes your intent into a requirements document (EARS notation), a technical design with data flow diagrams and API schemas, then generates and iterates on code. Hooks automate tasks like test regeneration on file save — and they're free, not counted against your quota.
read it → - ·4 mincursorspacexxaiacquisitionmarketanysphere
spacex is buying cursor for $60 billion — the largest startup acquisition ever
SpaceX announced a $60 billion all-stock deal to acquire Anysphere (Cursor) on June 16, 2026 — expected to close this quarter. Cursor had $4B in ARR as of early June. The deal gives xAI, now merged with SpaceX, its first serious foothold in developer tools.
read it → - ·5 minfable-5anthropicsecurityexport-controlsjailbreakmodels
the US government pulled claude fable 5 for 18 days. here's what actually happened.
Three days after launch, Claude Fable 5 was suspended globally under a U.S. export-control order triggered by a jailbreak Amazon researchers found in the model's cybersecurity safety classifier. The model returned July 1 with a new classifier and a trade-off: fewer exploits generated, more false positives on routine coding tasks.
read it → - ·3 minclaude-codeanthropicbrowserdesktoptoolingupdates
claude code desktop can now browse the web — and /doctor can actually fix things
Claude Code v2.1.202–206 (week 28, July 6–10) added a sandboxed built-in browser to the desktop app — Claude can now pull up docs, click through pages, and interact with external sites the same way it reads your local dev server. The same release upgraded /doctor from a read-only report into a tool that diagnoses and fixes setup problems.
read it → - ·4 mingrokxaispacexaicursormodelscodingbenchmark
grok 4.5 is trained on cursor session data — and it shows in the benchmarks
xAI launched Grok 4.5 on July 8–9, 2026 — a 1.5-trillion-parameter MoE model built jointly by SpaceXAI and Cursor, trained on real developer coding sessions. It lands within 1% of Fable 5 on Terminal-Bench 2.1 at $2/$6 per million tokens — and CursorBench was honestly excluded from the results due to training data contamination.
read it → - ·3 minanthropicclaudemodelscodingbenchmark
anthropic shipped claude opus 5 — top of frontier-bench at half fable 5's price
Anthropic released Claude Opus 5 on July 24, 2026 — its fourth model in under two months. On Frontier-Bench v0.1 it posts 43.3%, roughly double Opus 4.8's 21.1%; ARC-AGI-3 is 30.2%. Pricing holds at $5/$25 per million tokens, the same as Opus 4.8 and half Fable 5's rate.
read it → - ·4 minkimiopen-weightmodelscodingbenchmarkchina
kimi k3 is the largest open-weight model ever — and it tops the frontend code arena
Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model that leads the Frontend Code Arena at 1,679 points, ahead of Fable 5 and GPT-5.6 Sol. API pricing is $3/$15 per million tokens. Full weights drop July 27.
read it → - ·4 minopenaigpt-56codingmodelscodex
openai launched three models at once — sol is the one you'll care about
OpenAI's GPT-5.6 family — Sol, Terra, and Luna — went live on July 9, 2026 in ChatGPT, Codex, and the API. Sol is the flagship: state-of-the-art on coding with Programmatic Tool Calling that lets the model write and run in-memory JavaScript to coordinate tools. Terra is GPT-5.5 quality at half the cost. Luna is the cheap/fast option.
read it → - ·4 minwindsurfdevincognitioniderebrand
windsurf is now devin desktop — and cascade is already dead
Cognition quietly rebranded Windsurf to Devin Desktop via an OTA update in June. Cascade reached end-of-life on July 1. Devin Local, its Rust-based replacement, is 30% more token-efficient and supports sub-agents. if you have Windsurf installed, you're already running Devin Desktop.
read it → - ·3 mingithub-copilotdesktop-appbyokagents
copilot's desktop app is now on every plan — free tier included, bring-your-own-key supported
the GitHub Copilot desktop app opened to all plans on July 7 — Free, Education, and paid. you don't even need a Copilot subscription: bring your own model provider API key and the agent runtime is free to use.
read it → - ·3 mingithub-copilotkimiopen-weightenterprise
copilot added its first open-weight model — and enterprise teams have to opt in
Kimi K2.7 Code is the first open-weight model selectable in GitHub Copilot's model picker. it is off by default for Business and Enterprise — admins must flip the policy switch. the open-weight distinction matters more for auditing than for pricing.
read it → - ·3 mincursormodel-routingenterprise
cursor shipped a model router so you stop overpaying for simple tasks
cursor router picks the right model for each request automatically — intelligence, balance, or cost mode — and claims 60% savings versus always reaching for the frontier model. it is live on teams and enterprise plans.
read it → - ·3 mingithubcopilotenterprise
github's new copilot dashboard tells you which users are still just autocomplete people
github shipped a copilot impact dashboard for enterprise admins on july 22 that groups users into three adoption phases — code-first, agent-first, multi-agent — and surfaces per-cohort PR throughput and merge velocity so admins know which levers to pull.
read it → - ·3 minclaudeanthropicsecurityclaude-codevulnerabilityplugins
claude security is now a claude code plugin — multi-agent vulnerability scanning from your terminal
Anthropic released Claude Security as an installable Claude Code plugin in public beta on July 21, 2026. The plugin runs a multi-agent vulnerability scan of your repo, then turns selected findings into patch files — all inside an existing Claude Code session. Previously Enterprise-only; now available to all Claude Code users.
read it → - ·3 mingithubcopilotgemini
copilot's model picker now comes with a credit meter
github added gemini 3.6 flash to copilot on july 21, one day after shipping clearer ai-credit usage and cost-center controls. the model menu is getting richer, but the meter is getting closer too.
read it → - ·3 minopenaicodexhardware
openai made a $230 keyboard for codex power users
openai's codex micro turns agent status, voice prompts, and common codex actions into physical controls. it is already out of stock, which says as much about developer-tool fandom as it does about the hardware.
read it → - ·3 mingooglegeminiantigravity
gemini cli users now have an antigravity migration deadline
google says gemini cli and gemini code assist stopped serving consumer-tier requests on june 18, with antigravity cli replacing the old path. the agent stack is consolidating around bigger workspaces, not single-purpose terminals.
read it → - ·3 mingeminigooglemodelsflashapipricingcoding
gemini 3.6 flash: $1.50 per million tokens, 1m context, and three new models at once
Google released Gemini 3.6 Flash alongside two sibling models on July 21, 2026 — 3.5 Flash-Lite and 3.5 Flash Cyber. The flagship of the batch: 1M-token context, 64K output cap, pricing at $1.50/$7.50 per million tokens (input/output), and knowledge cutoff updated to March 2026. It landed in GitHub Copilot on the same day.
read it → - ·3 minkilo-codeanacondaacquisitionvscodeopen-sourceenterprise
anaconda acquires kilo code — open-source, model-agnostic coding agent gets enterprise backing
Anaconda acquired Kilo Code on July 15, 2026, picking up the open-source, model-agnostic VS Code and JetBrains coding agent used by more than 3 million developers. Kilo routes nearly 10 trillion tokens per month across 500+ models from 60+ providers. Anaconda's pitch: enterprise governance for the trillion-token era. Kilo stays open-source and unchanged for existing users.
read it → - ·3 mincursorpricingteamsenterprisesubscriptions
cursor teams splits into standard ($40) and premium ($120) seats — two usage pools, 5x the limit
Cursor restructured its Teams plan starting July 1, 2026 for renewing customers. The old single-tier Teams seat is now two: Standard at $40/user/month and Premium at $120/user/month with 5x the usage. Usage is split into separate pools for first-party models (Auto, Composer 2.5) and third-party API (OpenAI, Anthropic, Google billed through Cursor). Billing is per active seat, not pre-allocated.
read it → - ·3 mingithubcopilotvisionbrowser-toolsagentsvscodemultimodal
github copilot vision and browser tools are generally available — attach images, pdfs, and navigate live pages
GitHub shipped two GA features on July 1, 2026: Copilot Vision lets every subscriber attach images and PDFs directly to chat prompts so Copilot can reason about what it sees alongside code. Browser tools for VS Code agents let agents navigate pages, capture screenshots, fill forms, and validate web apps without leaving the editor. Both are on by default, no admin action required.
read it → - ·4 minopenaicodexgitpodonaagentscloudacquisition
openai acquires ona (formerly gitpod) — codex gets persistent cloud sandboxes for multi-day agent work
OpenAI announced on June 11, 2026 that it will acquire Ona, the Berlin-based startup formerly known as Gitpod, which runs AI agents in isolated cloud sandboxes. The deal shifts Codex from a tool you supervise keystroke by keystroke to a worker you hand a job and check on later — agents can now run for hours or days unattended, inside an enterprise-grade cloud environment, without your laptop being open.
read it → - ·5 minagentsclaudemilestone
bun just got rewritten in rust — by an agent swarm, in 11 days
jarred sumner reportedly used claude's new dynamic workflows to migrate the bun runtime from zig to rust — 750,000 lines, 11 days, 99.8% of the test suite passing. the real hero of the story isn't the model.
read it → - ·3 mingithubpricingcopilot
copilot's new token billing has developers doing math at 2am
github completed its switch to usage-based AI credits on june 1 — and the backlash was immediate, with some developers reporting bills jumping from $29 to $750. the metering of vibe coding is officially here.
read it → - ·3 mingithubgooglemodels
gemini 3.1 pro and 3.5 flash land across copilot
github's changelog quietly added google's gemini 3.1 pro (preview) and 3.5 flash to copilot CLI, the cloud agent, the copilot app, and the SDK. copilot keeps becoming a model marketplace — which is the most interesting thing about it.
read it → - ·4 minmicrosoftmodelsbusiness
microsoft stops renting its coding brain
at build, microsoft unveiled MAI-Code-1-Flash — its first in-house coding model, shipping straight into github copilot and vs code — plus a budget reasoning model. the subtext: less openai, lower costs, and a price war that's good news for builders.
read it →