OpenCode 2.0 adds background subagents and an SDK, and 1.0 plugins won't work

OpenCode's 2.0 overhaul runs subagents and shell commands in the background on your machine, adds an SDK, and won't run 1.0 plugins. Three more stories inside.

Anthropic replaces git internally after its merges hit 30,000 a day, from 2,500 in May; a team of seven built Scion, the in-house system

Anthropic outgrew git at 30,000 merges a day, so a team of seven built Scion, an in-house system it says isn't for general use. Three more stories inside.

Cursor ends new $20 Pro signups as it joins SpaceX's X Subscription; staff say the comparable tier is $30, and current Pro users keep $20 for now

Cursor now sells only inside SpaceX's X Subscription, bundled with Grok and X Premium; staff say the $30 Premium tier is comparable to the $20 Pro, and current Pro users keep $20 for now. Also: Codex CLI's default background daemon, and a Harvard study on AI coding agents.

OpenAI's Codex raises GPT-6.1 Sol's default context about 80%, from 272k to 500k tokens, Sottiaux says

OpenAI's Tibo Sottiaux says Codex now gives GPT-6.1 Sol 500k tokens of context by default, up from 272k, with no long-context fees on subscription usage. Also: Claude filed a false murder tip during Anthropic's own testing, and Zod 4.6.6 on npm is clean.

Codex Ultrafast runs GPT-6.1 Sol up to 8x faster, OpenAI says, and costs up to 8x more

OpenAI says Codex's Ultrafast mode runs GPT-6.1 Sol up to 8x faster, and it draws plan usage at 8x the standard rate per token. Also: Anthropic pauses its startup perks, and Google's new agent runs on Claude too.

Anthropic ships Claude Haiku 5.5, says it costs about 75% less to run than Haiku 4.5

A new small model Anthropic says costs about 75% less to run than Haiku 4.5, plus cheaper Sonnet 5.5 cache reads and API credits for Max and Team plans. Also: Codex refills usage limits again.

Mistral ships a trillion-parameter model, Large 4, in API preview, with weights due at month's end

A trillion-parameter model in preview, scoring 61.7% on the DeepSWE coding benchmark, with weights due this month. Plus Codex users vote for a usage reset.

OpenAI makes Codex's Auto-review free, drawing no usage from your plan

Day two of the Codex team's 28-day pledge: Auto-review, a second agent that approves risky actions, no longer draws on your plan.

Hit Codex limits? OpenAI vows relief: an upgrade or a reset every day for 28 days

OpenAI's Codex team says each day brings either an upgrade or a full usage reset. Day 1 brought a faster default speed. Plus default spending caps, Copilot's retired models and Claude Code's mod patches.

Claude Code adds mods: plugins that reach into its interface

Claude Code's new mods can rewrite its interface and behavior; a built-in one flags what you or Claude might miss. Three more stories inside.

Google announces Gemini 4 Argon, invite only for now

Gemini 4 Argon tops one coding benchmark in Google's own table but is invite only for now. GPT-6.1 Sol's speed complaints draw a promised fix from OpenAI's Codex team. Two more stories inside.

OpenAI ships GPT-6.1 Sol and dots, its always-on agents

GPT-6.1 Sol lists at Sonnet 5.5's price and leads it at medium effort in an independent test; dots are agents that run around the clock. Four more stories inside.

Anthropic ships Claude Sonnet 5.5: 70.6% on Terminal-Bench, up from 10.3%, it says

Same price as Sonnet 5, and Anthropic says up to 30% less per task. An outside firm measured about 50% more at maximum effort. Three more stories inside.

Hit the limit? Claude Code now wraps up instead of cutting off mid-edit

It spends a small, fixed allowance from your weekly limit: once a week on Pro, every time on Max and Team Premium. Three more stories inside.

OpenAI planning $500/mo Pro Max? A ChatGPT tier leaks, TestingCatalog says

A $500 tier with "Fastest Work and Codex" sits in unreleased ChatGPT plan text, TestingCatalog says; OpenAI hasn't confirmed it. Two more stories inside.

Claude Code cloud sessions go live, with a $100 or $250 credit to try them

Existing Pro and Max subscribers get a one-time $100 or $250 credit, spent before plan usage. Two more stories inside.

Opus 5.5 costs 37% of Opus 5 per task

At each model's default setting, two independent benchmarks put Opus 5.5 at 37 to 38% of Opus 5's cost per task, a bigger saving than Anthropic claims. At max effort it costs more.

Claude Opus 5.5 and GPT-6 Sol ship hours apart, each cheaper than before

Anthropic shipped Claude Opus 5.5 and OpenAI shipped GPT-6 Sol and Luna on Tuesday, each cheaper per token, and Artificial Analysis has first coding numbers for both. One more story inside.

Grok 4.7 ships at half rivals' price, SpaceXAI says

SpaceXAI prices Grok 4.7 at $2 per million input tokens; its own table puts Anthropic's Fable 5.1 ahead on both coding benchmarks.

Claude Code adopts AGENTS.md, the shared format rivals already read

Claude Code 2.1.277 now reads AGENTS.md, the instructions file Codex, Jules and other agents already use, whenever a project has no CLAUDE.md.

Bend 2 blocks AI mistakes with proof, its maker says. The demo took 442 lines

Taelin's Bend returns as a proof checker for AI code, and the first outside read found the demo needs 442 lines of proof for 58 lines of rules. Five more stories inside.

OpenAI says its models told themselves to hide mistakes, in six new misalignment reports

OpenAI's own report says the instructions "were often followed," and gives a rate for two models it has since shipped.

Ex-OpenAI researcher launches Jev, a decision model he says runs 20-200x faster than LLMs

Diogo Almeida, who says his OpenAI work became the research behind ChatGPT, launched Jev: a model that returns typed decisions instead of text. His blog, his post and his home page each quote a different speed multiple.

Claude Mods land behind a flag: 14 of 31 public plugins can run host code

Claude Code's plugin hooks are switchable today behind an environment variable, and a community scan of 31 public mods found 14 can run processes on your machine.