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.
SpaceXAI shipped Grok 4.7 at half the price of "comparable models," and its own benchmark table still puts Anthropic's Fable 5.1 ahead on both listed coding tests. Xiaomi open-sourced MiMo-V2.6, a trillion-parameter model its own launch post says performs "on par" with Claude Opus 5, though its own model card shows Opus ahead on eight of eleven agent benchmarks. Linear rebuilt its CI pipeline after agents sped up shipping and its test suite nearly quadrupled since January. Nous Research shipped a Hermes Agent plugin that runs the real Claude Code CLI instead of impersonating it, the route Anthropic blocked in June. And an 80-minute outage overnight sent elevated errors to Claude Code, the API and Claude Cowork.
Price cuts and open weights #
Grok 4.7 ships at half rivals' price, SpaceXAI says
SpaceXAI's Grok 4.7 is priced at $2 per million input tokens and $6 per million output, the same price and speed as its predecessor Grok 4.6, but the company says it's twice as fast and half the price of "comparable models" elsewhere. On SpaceXAI's own benchmark table, though, Anthropic's Claude Fable 5.1 Max scores higher than Grok 4.7 on both listed coding tests, CursorBench 4.0 and Terminal-Bench 4.0, at roughly five times the price. Elon Musk called it "a strong combination of intelligence, speed & low cost." If price-per-token is what you're optimizing, Grok 4.7 is a real cut. If you're picking the model that scores highest on SpaceXAI's own coding benchmarks, its own chart points elsewhere.
Xiaomi open-sources MiMo-V2.6, trailing Opus 5 on most tests
Xiaomi open-sourced MiMo-V2.6 under an MIT license: a 1.02-trillion-parameter mixture-of-experts model with a 1-million-token context window. Its launch post says the flagship Pro checkpoint "performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks." Its own model card says otherwise: Opus 5 scores higher on eight of the eleven agent benchmarks it lists, including a 14-point gap on Terminal Bench 4.0. Artificial Analysis lists it at $0.435 per million input tokens. Read the table, not the tweet.
Tools reshaping the daily grind #
Linear rebuilt its CI after AI coding made it a bottleneck
Linear says agents made it "exponentially faster to ship code," and every change still has to pass CI (the pipeline that validates every pull request). Its test suite nearly quadrupled since January; CI hadn't kept pace: PRs were queuing more than 6 minutes for validation. The fix was faster infrastructure, not fewer tests: moving off GitHub Actions to third-party runners cut average job time 34% on its own. Wait time is now just over 5 minutes despite roughly four times the tests.
Hermes Agent runs the real Claude Code CLI, after June's block
Nous Research shipped a plugin letting its Hermes Agent use a Claude Pro or Max subscription again, by running the actual Claude Code CLI instead of pretending to be it. The earlier route broke in June, when Anthropic's login token endpoint began rejecting requests carrying a spoofed claude-cli/ user agent. This plugin drives the real executable instead, so it no longer needs that workaround. Nous calls it an official Hermes plugin; Anthropic hasn't said anything about it either way.
Overnight #
Claude Code, API and Cowork go down for 80 minutes overnight
An 80-minute incident overnight sent elevated errors to Claude Mythos 5.1, Claude Fable 5.1 and Claude Opus 5, across claude.ai, the API, Claude Code and Claude Cowork. Anthropic says it identified the cause and resolved it by 02:10 UTC; the status page doesn't say what the cause was. If you saw failed requests between 00:50 and 02:10 UTC on the 22nd, that was this, not your code.
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
Also worth your time #
- JetBrains launches Air, a new umbrella for agentic development across its IDEs — mostly positioning ("the era in which the whole software development system can be contained in one window is ending"); the post names no pricing and no concrete new workflow.
- One developer says Fable 5 used far fewer thinking tokens in August than July — one user's own measurement ("Measured five different ways"), not independently checked.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.