Anthropic released Opus 5.5 on September 22 at $4 input and $20 output per million tokens, 20 percent under Opus 5, with cache reads down 60 percent to $0.20 per million, per Anthropic's announcement. Cache reads carry most of the cost in agentic and coding work, same source, so that second number moves your invoice further than the headline rate does. Opus 5.5 posts 66.4 percent on Terminal-Bench 4.0 against 55.8 percent for Claude Fable 5.1 on Anthropic's own table, and one early tester finished a 680,000-line code migration in under a day. Five-hour usage limits go up on paid plans too.

GPT-6 Sol drops to $2 and $10 per million tokens and GPT-6 Luna to $0.10 and $0.50, half the GPT-5.6 rates they replace, per OpenAI's pricing table. Cached input reads now carry a 90 percent discount, on the same page, and GitHub reports that OpenAI's caching work cut the share of prompt tokens needing fresh processing by more than half across billions of requests. Sol at $2 and $10 sits at half of Opus 5.5 on the rate card, which is exactly the comparison worth distrusting. Point both at a task set you grade yourself and compare pass rate per dollar before you reroute anything.

xAI kept Grok 4.7 at $2 and $6 per million tokens while lifting CursorBench 4.0 from 40.4 to 46.3 percent over Grok 4.6, per VentureBeat. Then count tokens. Artificial Analysis measured Grok 4.7 at xHigh effort burning roughly 81,000 output tokens per Intelligence Index task against 36,000 for Grok 4.6, landing near $3.74 per task versus about $1.99 for GPT-5.6 Sol Max, a model that lists at double the token rate, per the same report. Rank candidates on cost per finished task, not per million tokens, or the cheap model wins the spreadsheet and loses the invoice. Grok 4.7 is live in Cursor if you want it against your own repo this week.

Intrinsic released Intrinsic Core at ROSCon 2026 on September 22: the same ROS-compatible capabilities it runs for real manufacturing deployments, a hardware-agnostic real-time control framework, and a native digital twin, per Intrinsic. An open machine-tending reference solution ships alongside it, so you start from a working application. Apache 2.0 is the clause that matters commercially: this can go inside a product you sell without a licensing conversation, which is not true of most robotics middleware. Code is on GitHub.

France chairs a high-level briefing on AI and international security today, the Council's first meeting aimed specifically at safety risks from increasingly capable systems and the potential loss of human control, per Security Council Report. Yoshua Bengio, Sam Altman, Dario Amodei, and Clement Delangue are the anticipated briefers, and France's concept note asks how evaluation and verification can keep pace with capability. Text drafted at that table resurfaces in procurement requirements a year later. Anyone selling AI to a government or a regulated buyer should read those four questions this week.

Microsoft's Digital Crimes Unit took down EvilTokens with court authorization in the Eastern District of Virginia and partners including Cloudflare, OpenAI, Coinbase, and TRM Labs, seizing 50 websites and disabling more than 150 supporting domains, per Microsoft. Launched in February 2026, the service was linked to more than 12,000 compromised inboxes across over 10,000 organizations, according to the same post, and its chatbot read stolen mailboxes to find payment authority and draft impersonations of people those targets trust. Act on the mechanism rather than the headline: victims approved a device code on Microsoft's real sign-in page, so attacker access outlived password resets wherever sessions and tokens went unrevoked. Restrict the device code flow with a Conditional Access policy, and revoke sessions on every reset.

Our picks, published September 22 with pricing and licenses cited in a single table: Claude Code for the solo developer who wants the best score on a hard change, OpenCode for the team that wants the vendor-lock question to disappear, Cursor if the full IDE is why you showed up and month-to-month billing through the November 12 OpenAI model shutoff suits you. Opus 5.5 landing across Anthropic's platforms only sharpens the first pick. See the updated Q4 2026 buy decision for Claude Code, Cursor, and OpenCode.