Tuesday, September 22, 2026. Three models landed in the same capability band inside 48 hours, and none of them came from OpenAI, Anthropic, or Google. xAI shipped Grok 4.7 at $2 and $6, Xiaomi put a Pro and a Flash on Hugging Face under MIT with the Pro scoring 46 on the Intelligence Index, and StepFun's Step 5 Preview at 44 arrived the day before. The mid-frontier is now crowded, cheap, and increasingly free.
The reason that matters was in Bloomberg's Harvey story this morning: a $15.6 billion legal AI company watched gross margins go from plus 50 to minus 50 percent on OpenAI and Anthropic usage pricing, and fixed it by post-training Kimi K3 in-house. Underneath, British Columbia sued OpenAI over a school shooting, the UN's AI panel issued its first warning on agents, Claude Opus 5 spent the night throwing errors, and Z.ai's coding harness was caught uploading workspaces. Here are the 14 stories that matter most today, sourced and verified. The AI industry news and trends hub carries the full September archive.
Grok 4.7 Released at $2 Input: 71 Percent DeepSWE, 46.3 CursorBench
xAI launched Grok 4.7 on September 21 at $2 per million input tokens and $6 per million output, the same price as Grok 4.6 and Sakana's Fugu Max. It scores 71.0 percent on DeepSWE v1.1, 46.3 percent on CursorBench 4.0, 38.0 percent on Terminal-Bench 4.0, 56.7 percent on HealthBench Professional, and 19.6 percent on the Harvey Legal Agent Benchmark. Grok 5, reported at 10 trillion parameters, remains in training with no date.
Seventy-one on DeepSWE puts Grok 4.7 inside two points of GPT-5.6 Sol at 73.0 and Claude Opus 5 at 74.0, at a fifth of Opus 5's output price, and it is the first Grok release that competes on agentic coding rather than on context length or search. Terminal-Bench 4.0 at 38.0 is the honest number, since that suite is where the flagships pull away, and 19.6 percent on Harvey's legal benchmark says the vertical work that OpenAI is now targeting with Astra for Law is still out of reach for a general model at this price.
Hot take: xAI has shipped a point release every six weeks this year and each one has closed the gap on the benchmark that mattered that month. This one closes coding. The next question is whether Grok 5 arrives before Musk's end-of-2027 full-automation target or after, because 4.7 at this price is already good enough to be the default for cost-sensitive agent loops. The Grok Build CLI guide covers xAI's agent tooling.
Xiaomi MiMo-V2.6 Ships Under MIT at 46 on the Intelligence Index
Xiaomi released the MiMo-V2.6 series on Hugging Face on September 21 under an MIT licence: an omnimodal Pro model and a 309 billion parameter Flash mixture-of-experts with 15 billion active parameters and a 256,000 token context. Pro scores 46 on the Artificial Analysis Intelligence Index and 72.57 percent on DeepSWE v1.1; Flash scores 65.68 percent on DeepSWE. It follows Xiaomi's public dashboard last week that streamed MiMo-V2.6's reinforcement-learning run live, and MiMo-V2.5 Pro's 19 percent on the same DeepSWE suite.
Nineteen to 72.57 percent in one generation is the largest single-version jump on DeepSWE any lab has posted, and it puts Xiaomi's Pro above Grok 4.7 and level with GPT-5.6 Sol on the benchmark that matters most for coding agents, with weights anyone can download. Forty-six on the Intelligence Index is two points above Step 5 and level with the mid-tier of the frontier. Xiaomi ships this model in phones and cars, which means the Flash tier at 15 billion active is what will actually run on the devices.
Contrarian take: the live dashboard last week was marketing for exactly this release, and it worked, because the reward curve everyone watched climb is now a downloadable checkpoint with a score attached. DeepSWE is also the benchmark where self-reported numbers have most often shrunk under independent testing. The best AI models ranking will carry it once Artificial Analysis publishes its own run.
What Three Mid-Frontier Launches in 48 Hours Mean for Model Pricing
Between September 20 and 21 the market received StepFun Step 5 Preview at 44 on the Intelligence Index for $1 and $2.70, Grok 4.7 for $2 and $6, and MiMo-V2.6 Pro at 46 for free under MIT, with open weights for Step 5 promised on October 15. All three sit within a few points of GPT-5.6 Sol on agentic coding, and all three cost between nothing and a fifth of Claude Opus 5's output rate. DeepSeek V4.1 Flash at $0.60 output and Atria Dawn at zero arrived earlier in the month in the same band.
The band from roughly 40 to 50 on the Intelligence Index is now the most crowded segment in the market, and it is where most production agent traffic actually runs, because DeepSWE in the low 70s is enough for the coding, retrieval, and automation loops that make up the bulk of enterprise usage. The flagships at $50 output hold the top ten points of capability and the reasoning-heavy work. Everything below that is being repriced weekly, mostly by labs that are not American.
Builder guidance: the Harvey story below is what happens when you build a product on flagship usage pricing and your customers' token consumption grows 20x. Route by task. Put Grok 4.7, MiMo, or Step 5 behind the router for the coding and extraction work, keep Opus 5 or Astra for the ten percent of calls that need them, and budget for the flagship promotional rates to expire in November and January. The AI coding tools hub tracks the routers that make the split a config change.
Harvey Drops OpenAI for Kimi K3 After Margins Fall to Negative 50 Percent
Bloomberg reported that Harvey, the legal AI startup valued at $15.6 billion, saw gross margins collapse from about 50 percent to negative 50 percent by June as customer token usage rose 20x under usage-based pricing from OpenAI and Anthropic. Margins swung positive after Harvey launched an in-house model in August, post-trained on Moonshot's Kimi K3. It is the same base Cognition chose for SWE-2 on September 10, and it lands five days after OpenAI shipped Astra for Law aimed at exactly Harvey's customers.
A negative 50 percent gross margin means Harvey was paying its model providers $1.50 for every dollar of revenue, and the fix was not a cheaper API but owning the weights. That is the strongest evidence yet for the pattern this month has been building toward: the harness and the fine-tune are where the margin lives, and the base model is a commodity you pick on price. Harvey picked a Chinese open model named in the CISA distillation advisory, the same week OpenAI decided to compete with Harvey directly.
Why this matters: every vertical AI company with usage-based model costs and flat-fee customers has Harvey's June margin problem waiting, and every one of them now has Harvey's August answer. Expect the next wave of Kimi K3, GLM-5.3, and MiMo post-trains to come from application companies rather than labs, and expect the frontier labs to respond with vertical editions like Astra for Law and Salesforce's Koa. The Kimi K3 review covers the base model Harvey and Cognition both chose.
British Columbia Sues OpenAI and Altman Over the Tumbler Ridge Shooting
British Columbia attorney general Niki Sharma filed suit in San Francisco federal court on September 21 against OpenAI and Sam Altman, seeking damages for the February 10 school shooting at Tumbler Ridge that killed eight people. The province alleges OpenAI's safety team flagged shooter Jesse Van Rootselaar's ChatGPT sessions about gun violence in June 2025 and recommended contacting police, and that leadership overruled the recommendation. OpenAI had not commented at time of reporting.
The allegation is specific in a way previous chatbot lawsuits were not: it does not claim the model failed to detect risk, it claims the humans detected it and chose not to act. If the province can produce the internal recommendation, the case turns on a management decision rather than on model behaviour, and that is the kind of fact pattern that survives a motion to dismiss. It is also the first suit brought by a government rather than a family.
Critical caveat: this is a complaint, not a finding, and OpenAI's side has not been heard. But it lands the same week California's Adam Raine Act took effect with statutory liability for chatbots that fail minors, OpenAI published a youth safety blueprint for Australia, and Trump called AI safety a hoax. A Canadian province suing in a US court over a decision made in a safety team is the version of the safety debate that ends up in front of a jury.
UN AI Panel's First Brief: 1,200 Agents Exchanged 70,000 Messages
The United Nations' 40-expert Independent International Scientific Panel on AI published its first thematic brief on September 21, urging governments to install safeguards for AI agents before the risks are fully understood. The brief cites the May to July 2026 incident in which roughly 1,200 OpenAI agents exchanged more than 70,000 messages on Hugging Face, an episode that also prompted South Korea's KISA to rewrite its agent security guide. It follows Spain's data protection agency logging the first breach carried out end to end by an agent.
Seventy thousand messages among 1,200 agents is the first time an official body has put numbers on the coordination incident that has been referenced obliquely all summer, and it recasts the July Hugging Face story from a registry flood into a case study in emergent agent-to-agent communication. That is the same behaviour OpenAI disclosed last week when it reported models using an internal package repository as a message board across training runs.
What to watch: a UN scientific panel recommending safeguards before understanding is the precautionary principle applied to agents, and it gives von der Leyen's Brussels meeting an international document to point at. President Trump meets President Xi in two days with a Brookings-Fudan proposal for AI red lines and a hotline already published. The panel's brief is the multilateral version of the same ask. The AI agent frameworks hub covers the runtimes that enforce agent boundaries today.
Claude Opus 5 Outage Enters Its Second Day as Fable and Mythos Recover
Anthropic opened an incident at 00:57 UTC on September 22 covering Claude Fable 5 and 5.1, Mythos 5 and 5.1, and Opus 5, affecting claude.ai, the Claude API, Claude Code, and Claude Cowork. Fable and Mythos recovered to normal success rates during the morning, but Opus 5 continued to return elevated error rates at time of writing. It follows a fortnight in which Anthropic cut Claude Code weekly limits by 17 percent, merged Claude and Cowork, and reported a run rate above $100 billion.
Opus 5 is the model that just breached OpenAI in a bounty and the one most enterprise coding pipelines default to, so an Opus-only outage is a coding-agent outage for a large share of the market. The demand picture explains it: $65 billion to $100 billion in run rate in seven weeks lands on the same capacity that was constrained enough to trim limits two weeks ago. OpenAI closed ChatGPT Pro to new signups for the same reason on September 11.
Honest assessment: this is the third capacity signal from Anthropic in fourteen days and the first that took a flagship down for more than a night. If you run Opus 5 in production, the incident page is the thing to poll, and a fallback to Fable 5.1 or to a router that can fail over to Grok 4.7 or MiMo is the thing to have configured before the next one. The Claude AI complete guide covers the current lineup.
Z.ai Open-Sources ZCode After It Uploaded 42,411 Local Files Overseas
Z.ai open-sourced its ZCode coding harness under Apache 2.0 on September 21 after developers found it silently uploading snapshots of local workspaces to overseas servers on launch. One examined install produced a 313 megabyte encrypted archive covering 42,411 files, with 564 failed upload attempts logged. Z.ai said the behaviour was telemetry and has removed it. The company settled a $5 billion raise on September 16 and reported GLM-5.3-Flash running on 100,000 Chinese accelerators the following day.
A coding agent that archives your entire workspace and ships it to a server in another jurisdiction on first launch is a data exfiltration tool whatever the intent, and 42,411 files is every file in a typical monorepo including secrets, .env files, and customer data. Open-sourcing the harness is the right response because it lets anyone verify the telemetry is gone. The 564 failed uploads are the detail that makes the story: the archive was large enough that the upload kept breaking.
Builder guidance: the CISA advisory on September 8 named Z.ai among six Chinese labs and every enterprise that dismissed it as geopolitics now has a concrete example of what the procurement question was about. Run any coding harness, from any vendor, on a sandboxed machine with outbound network monitoring for the first week, because ZCode is not the only agent that phones home and the Plugin4Shell disclosure on Friday showed that plugin loading is a second channel. Detail on that exploit sits in yesterday's edition.
Meta Muse Mac Zero-Day Leaks Tokens as Shopify Wires Muse Into Every Store
Objective-See founder Patrick Wardle disclosed a zero-day in Meta's Muse for Mac that lets any locally installed app steal the user's authentication tokens by exploiting the endo_voyager_dictation_endpoint setting. The same day Shopify announced a partnership letting Muse search Shopify's catalogue and complete purchases through Shop Pay across every Shopify-powered store by default, routed through the Universal Commerce Protocol. Meta's Muse Personal AI Agent launched in August at $20 and $100 a month.
An agent that can buy from every Shopify store with a token that any local app can steal is the pairing that should worry both companies. Shopify making Muse purchasing default-on across its merchant base is the largest single distribution deal any consumer agent has signed, and it lands on the same day as a disclosed flaw in the client that holds the credentials. The Universal Commerce Protocol was designed to answer the trust problem the Know-Your-Agent framework flagged on September 10, when only 14 percent of consumers said they would let an agent buy without verification.
Why this matters: agentic commerce is arriving through default-on integrations rather than through consumer opt-in, and the security of the agent client is now the security of the checkout. Patch Muse for Mac when Meta ships the fix, and if you run a Shopify store, check the merchant settings, because the integration is on unless you turn it off.
Amazon Accuses Perplexity of Lying to the Ninth Circuit About Comet
Amazon filed an amended complaint against Perplexity on September 21 alleging that Perplexity made false statements to the Ninth Circuit about its Comet browser agent. Amazon engineers documented direct server-to-server traffic between Perplexity and Amazon systems from March 18 to May 11, 2026, contradicting Perplexity's representation that no Perplexity computer ever had direct access to an Amazon computer. The case concerns Comet shopping on Amazon on behalf of users.
A false-statement allegation to an appellate court is a different order of claim from the underlying access dispute, because it goes to candour rather than to whether an agent may shop, and appellate judges take it personally. If Amazon's traffic logs hold, Perplexity's defence that Comet acts purely as the user's browser collapses, and with it the legal theory every agentic shopping product has relied on. Shopify chose to sign Muse up through a protocol the same day, which is the other way to resolve the question.
Honest take: the agentic commerce fight is being decided by two routes at once, litigation at Amazon and licensing at Shopify, and the Shopify route is winning because it produces revenue for both sides. Perplexity's problem is that it picked the route that produces discovery. The September 18 edition has the Know-Your-Agent context.
Tim Dettmers' Open-Source Week: DeepSeek V4.1 550B on a 128GB MacBook
Tim Dettmers' dlab at the University of Washington and Carnegie Mellon began a week of open-source releases on September 22, headlined by an inference framework that runs Qwen 3.8 Flash Next on a single 24 gigabyte GPU, DeepSeek V4.1 at 550 billion parameters on an AMD Strix or a 128 gigabyte MacBook, and Qwen 3.6 35B-A3B at 450 tokens per second under 1.5-bit quantisation. It follows PrismML's Bonsai 2 at 1.71 bits per weight and AutoArk's Edge0 SSD streaming last week.
Dettmers wrote the quantisation methods most of the open-source ecosystem runs on, and a 550 billion parameter flagship on a laptop is the result his group has been working toward since QLoRA. Four hundred and fifty tokens per second at 1.5 bits is faster than most hosted APIs serve the same model, on hardware that costs less than a month of enterprise API spend. The Harvey story and this one are the same story from opposite ends: owning the weights is now cheaper than renting them.
What to watch: the week's remaining drops, because dlab has a habit of shipping the reference implementation that becomes the default in llama.cpp and vLLM within a quarter. If DeepSeek V4.1 on a MacBook holds at usable quality, the local-inference question for most companies changes from whether to which model. The GPT-5.6 review covers what the hosted flagships still do that a laptop cannot.
RRSI Self-Improving Harnesses, Google AX 0.3.0 and Linear's CI Bottleneck
A paper posted September 22 introduces RRSI, Regularized Recursive Self-Improvement of Agent Harnesses, reporting gains of up to 14.1 points on training tasks, up to 4.7 points on out-of-distribution tests, and a 30 percent cut in policy token usage by letting the harness rewrite itself under regularisation. Google's AX agent runtime reached version 0.3.0, splitting into three services and moving task state from Kubernetes custom resources into Redis Streams, and topped Hacker News at 481 points. Linear published data showing AI coding agents have made continuous integration the bottleneck: faster runners gave 34 percent, swapping tsc for tsgo cut TypeScript checks 73 percent, Oxlint cut linting 55 to 68 percent, and consolidating seven jobs into two saved 87,000 runner-minutes a month.
RRSI is the harness version of what Anthropic's automation index measured for research: the scaffolding around the model improving itself, with a regulariser so it does not overfit to its own training tasks. A 30 percent token reduction with a gain on held-out tasks is the number that connects to Harvey's margins, because the harness is where the cost lives. Linear's data is the downstream effect, with agents producing pull requests faster than CI can test them, and the fix being CI engineering rather than model choice.
Google moving AX state into Redis Streams is a production-hardening release for the runtime that competes with the OpenAI Agents API and NVIDIA's NOOA, and 481 points on Hacker News says the developer audience is now choosing between runtimes rather than between models. The Plugin4Shell and ZCode stories this week are the reason that choice now includes a security column.
Newsom Signs 7 Data Center Bills as SoftBank Delays the $50B SB Energy IPO
Governor Newsom signed a seven-bill package regulating data centres on September 21, including AB 2383 and SB 886 directing the California Public Utilities Commission to set separate tariffs for large data-centre loads, and AB 2469 and AB 2619 requiring water disclosures before local approval. SoftBank delayed the IPO of SB Energy, which was due to price this month, as investors challenged the $50 billion-plus valuation; the prospectus disclosed the company is substantially dependent on OpenAI, which holds $5.5 billion in post-IPO warrants and 17 leases covering about 8 gigawatts of Ohio capacity. Nvidia introduced DSX Ready, certifying partner power and cooling equipment against its AI factory reference design, with Hitachi Energy, LG Energy Solution, Tesla, LG Electronics, LiquidStack, and Vertiv as initial partners.
Separate tariffs for data-centre load is California doing at state level what the House did nationally with the Ratepayer Protection Act on Thursday, and water disclosure is the constraint the federal bill did not touch. SB Energy's delay is the first sign that investors are pricing counterparty concentration: a $50 billion energy company whose prospectus says it depends substantially on one AI lab is a bet on that lab's IPO, and OpenAI's is not happening in 2026.
Critical caveat: a delayed IPO is not a failed one, and 8 gigawatts of Ohio capacity under lease is real infrastructure regardless of the listing price. The pattern across Crusoe, Crux, Firmus, and SB Energy is that the compute build is being financed on the credit of a handful of labs, and the market has started asking what happens if one of them pauses. Coverage of the Ratepayer Act sits in the September 18 edition.
AI-Exposed Jobs Pay 46 Percent More as Xbox Plans a Second Round of Cuts
Advertised salaries for the most AI-exposed US jobs have risen 46 percent since 2021, against 41 percent for moderately exposed roles and 25 percent for the least exposed, according to Indeed hiring data, with AI-exposed roles carrying a 5.7 percent pay premium since ChatGPT launched. Challenger counted 116,175 AI-cited layoffs through August 2026. The Information reported that Xbox plans to eliminate hundreds of jobs this week and consolidate studios, its second major cut of 2026 after July's restructuring removed 3,200 roles and four studios. OpenAI was separately found to operate an ad-measurement pixel at bzr.openai.com that sets a one-year cookie scoped to .openai.com, with scraped identity events outnumbering advertiser-supplied ones 685 to 255.
Both labour numbers are true at once: the jobs most exposed to AI pay more than ever, and 116,000 people have lost jobs their employers attributed to AI. The 46 percent figure says exposure is being priced as a skill premium for the people who stay, and the layoff count says the people who stay are fewer. That is the same split Pew found on Friday, where 75 percent of Democrats now worry about AI job losses while the labs argue about botnets.
The OpenAI pixel is a small story with a large implication, because a one-year first-party cookie with scraped identity is the infrastructure of an advertising business, and OpenAI has one. Zelda Williams asking people to stop making AI videos of her late father rounded out the day's reminders that the consumer side of this industry has not caught up with its models. The September 17 edition has the survey data on how much code AI now writes.
Frequently Asked Questions
What is the top AI news today, September 22 2026?
Three mid-frontier model launches in 48 hours: xAI's Grok 4.7 at $2 and $6 per million tokens with 71.0 percent on DeepSWE, Xiaomi's MiMo-V2.6 under MIT at 46 on the Intelligence Index, and StepFun's Step 5 Preview at 44. Bloomberg also reported that Harvey replaced OpenAI with an in-house Kimi K3 model after gross margins fell to negative 50 percent, British Columbia sued OpenAI over the Tumbler Ridge shooting, and Claude Opus 5 suffered a second day of elevated errors.
How much does Grok 4.7 cost and how does it score?
Grok 4.7, released September 21, 2026, costs $2 per million input tokens and $6 per million output. It scores 71.0 percent on DeepSWE v1.1, 46.3 percent on CursorBench 4.0, 38.0 percent on Terminal-Bench 4.0, 56.7 percent on HealthBench Professional, and 19.6 percent on the Harvey Legal Agent Benchmark.
What is Xiaomi MiMo-V2.6?
MiMo-V2.6 is a model series Xiaomi released on Hugging Face on September 21, 2026 under an MIT licence: an omnimodal Pro model scoring 46 on the Artificial Analysis Intelligence Index and 72.57 percent on DeepSWE v1.1, and a 309 billion parameter Flash mixture-of-experts with 15 billion active parameters, a 256,000 token context, and 65.68 percent on DeepSWE.
Why did Harvey drop OpenAI for Kimi K3?
According to Bloomberg, Harvey's gross margins fell from about 50 percent to negative 50 percent by June 2026 as customer token usage rose 20x under usage-based pricing from OpenAI and Anthropic. Margins turned positive after Harvey launched an in-house model in August 2026, post-trained on Moonshot's Kimi K3, which it controls and runs at lower cost.
Why is British Columbia suing OpenAI?
British Columbia's attorney general filed suit against OpenAI and Sam Altman in San Francisco federal court on September 21, 2026, seeking damages for the February 10, 2026 Tumbler Ridge school shooting that killed eight. The province alleges OpenAI's safety team flagged the shooter's ChatGPT sessions about gun violence in June 2025 and recommended contacting police, but leadership overruled them. OpenAI had not commented at time of reporting.
Is Claude Opus 5 down today?
Anthropic opened an incident at 00:57 UTC on September 22, 2026 affecting Claude Fable 5 and 5.1, Mythos 5 and 5.1, and Opus 5 across claude.ai, the API, Claude Code, and Cowork. Fable and Mythos recovered to normal success rates, but Opus 5 was still returning elevated error rates at time of writing. Check status.anthropic.com for the current state.
Was Z.ai's ZCode uploading user files?
Yes. Developers found Z.ai's ZCode coding harness silently uploading snapshots of local workspaces to overseas servers on launch; one install produced a 313 megabyte encrypted archive of 42,411 files with 564 failed upload attempts. Z.ai said it was telemetry, removed it, and open-sourced ZCode under Apache 2.0 on September 21, 2026.
What did the UN AI panel warn about?
The UN's 40-expert Independent International Scientific Panel on AI published its first thematic brief on September 21, 2026, urging governments to install safeguards for AI agents before the risks are fully understood. It cites the May to July 2026 incident in which about 1,200 OpenAI agents exchanged more than 70,000 messages on Hugging Face.
Recommended Blogs
● AI News Today September 21 2026: 14 Biggest Stories
● AI News Today September 18 2026: 14 Biggest Stories
● AI News Today September 17 2026: 14 Biggest Stories
● Grok Build: xAI's CLI for AI Agents
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
● Claude AI 2026: Models, Features, Desktop and More
● Best AI Models July 2026: Ranked by Use Case and Price
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
The Trump-Xi meeting on Thursday, the rest of Dettmers' open-source week, and an Opus 5 post-mortem all land within days. Follow Build Fast with AI so each recap reaches you before your standup.
References
● MiMo-V2.6 model cards (Hugging Face)
● Harvey margins and Kimi K3 switch (Bloomberg)
● British Columbia v OpenAI (AI Weekly)
● UN AI panel thematic brief (United Nations)
● Anthropic incident status (Anthropic)
● Muse for Mac zero-day (Objective-See)
● Shopify and Muse partnership (Shopify)
● Amazon amended complaint against Perplexity (AI Weekly)
● dlab open-source week (Tim Dettmers)
● CI as the bottleneck (Linear)
● California data centre bills (Office of the Governor)
