Three things worth your time today, plus five that matter less but will come up.
Grok 4.7 Ships at Half the Price
SpaceXAI released Grok 4.7 on 21 September at $2 per million input tokens and $6 per million output, with a fast variant at double both for roughly double the speed.
It scores 71.0% on DeepSWE v1.1 and 46.3% on CursorBench 4.0, which is solid coding performance. It also scores 38.0% on Terminal-Bench 4.0 and 19.6% on the Harvey legal agent benchmark - a model tuned hard for bounded code tasks and noticeably weaker on long-horizon and professional-domain work.
Several launch-day reviews flagged high token consumption per task, so a lower per-token price may not produce a lower bill. Measure cost per completed task.
Full pricing and benchmark breakdown
ZCode Was Uploading Your Git History
A developer clearing disk space on 17 September found 700MB in ~/.zcode. It turned out to be staged workspace snapshots. One commercial project came to 42,411 files in a 313MB encrypted archive, of which 86.6% was Git data - every commit, every branch, every deleted file.
Two UI settings that looked like privacy controls governed training use and server-side indexing, not capture. The pipeline ran unconditionally at startup. Z.ai fixed it in 3.14.0 and open-sourced the harness under Apache-2.0 on 21 September.
What was sent, and what to rotate
Harvey Left OpenAI and Anthropic
Bloomberg reported that Harvey, the legal AI company valued at $15.6B, moved off closed frontier models. The reason is arithmetic: gross margins ran from roughly 50% to minus 50% by June as token usage jumped twentyfold under usage-based pricing.
In August Harvey launched an in-house model post-trained on Moonshot Kimi K3. Margins swung positive after the migration. Abridge, Decagon and Ramp are reported to be pursuing the same open-weight pivot, with Sequoia and General Catalyst backing the trend.
This is the first serious wave of enterprise AI companies leaving closed-model dependence, and the driver is unit economics rather than capability.
Also Today
- Anthropic IPO moves to November. Pacing above $100B annualized revenue in 2026, up 50% from the $65B disclosed in July. The IPO slipped from October to November to include Q3 financials, targeting a valuation around $2 trillion.
- StepFun Step 5 Preview. A 600B sparse MoE with 27B active parameters and 1M context, scoring 44 on the Artificial Analysis Intelligence Index at roughly one-seventh the price of GPT-5.6 Sol. Full open weights promised for 15 October.
- UN panel calls for agent safeguards. The 40-expert Independent International Scientific Panel on AI invoked the precautionary principle on 21 September, referencing the OpenAI Hugging Face incident. Co-chair Yoshua Bengio: the traditional model of safeguarding is unravelling.
- California signs seven data centre bills. Governor Newsom signed a package on 21 September covering separate CPUC tariffs for large data-centre loads and mandatory water disclosure before local approval. He had vetoed similar measures previously.
- Cloudflare Python Workers hit GA. FastAPI, Django and Flask run natively, with OpenAI, LangChain and MCP libraries working without JavaScript glue, plus a new TCP socket implementation for Postgres and MySQL through Hyperdrive.
The Thread Running Through Today
Grok 4.7 competes on price. Harvey left the frontier labs over price. StepFun is undercutting GPT-5.6 Sol by seven times. Three unrelated stories, one direction of travel: for a growing set of production workloads, the frontier model is no longer the default answer, and the reason is the invoice rather than the benchmark.
Sources