Top Stories
2 stories + 4 briefsGrok 4.7: xAI reports faster, cheaper coding and knowledge work
The xAI news page, dated September 21, describes Grok 4.7 as its strongest model for coding and knowledge work and claims twice the speed at half the price of comparable models. API prices start at $2 per million input tokens and $6 per million output tokens. A faster variant offers twice the output speed at twice the price. The article was opened; first discovery was September 25. Browser failures from September 19 to 21 kept this release out of Issue 15, so this is a late-discovered September 21 release rather than new September 22 news.
The page identifies the company as SpaceXAI while retaining the x.ai domain. It describes longer reinforcement-learning training aimed at tasks that take hours, a new safeguards stack, and the strongest refusal and jailbreak resistance the company has measured. These are company claims. Availability includes Cursor, Grok Build, the Grok API, third-party coding harnesses, model routers and cloud platforms. Its CursorBench button reports 46.3% for Grok 4.7 Extra High at an average $6.01 per task. Original source: https://x.ai/news/grok-4-7
Claude Opus 5.5: Anthropic reports lower running costs and new safeguards
Anthropic calls Claude Opus 5.5 the first model in the Claude 5.5 family, matching Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. The official page is dated September 22, with no publication time; it was opened and first discovered on September 25. This gives an official product-page follow-up to the previous issue’s Reuters report that Anthropic was weighing a release. The price and audit claims come from the product page.
Anthropic says this is its first release after calling for a slowdown at the frontier, with pre-release testing by external evaluators including Frontier Design and METR. It describes the model as its strongest on automated behavioral audits and says it carries safeguards designed for its most capable models. Examples include a tester migrating 680,000 lines of code in under a day and successful page-loading-time reductions in 39 of 40 web-app tasks. These examples and rankings are reported by Anthropic. Original source: https://www.anthropic.com/claude-opus-5-5
Hugging Face describes loading and running GGUF through from_pretrained for local generation. The article was opened; its page date is September 22, 2026. Source: https://huggingface.co/blog/transformers-llama-cpp-quants
Source: Archived source · huggingface.co · Original issue: 2026-09-22 · Archived edition only; original reading scope is retained in the article. No fresh source verification is claimed.
The article calls for explicit requirements, enforceable controls, named owners and evidence that safeguards work. It was published September 21 at 14:51 UTC and read in full. Source: https://blogs.nvidia.com/blog/ai-security-agent-stack/
Source: Archived source · blogs.nvidia.com · Original issue: 2026-09-22 · Archived edition only; original reading scope is retained in the article. No fresh source verification is claimed.
The Hugging Face article describes using EvalEval infrastructure to make evaluation results reproducible and checkable. It was published September 22 at 00:00 UTC and opened. The earlier collaboration began with a joint workshop at NeurIPS 2025. Source: https://huggingface.co/blog/evaleval-aisi
Source: Archived source · huggingface.co · Original issue: 2026-09-22 · Archived edition only; original reading scope is retained in the article. No fresh source verification is claimed.
The RSS timestamp for Priorities and principles for effective third party assessments is September 22 at 00:00 UTC. The article page was not opened; this brief uses only the RSS summary about independent third-party evaluations of frontier models and safeguards. No figures absent from that summary are added. Source: https://openai.com/index/priorities-principles-third-party-assessments
Source: Archived source · openai.com · Original issue: 2026-09-22 · Archived edition only; original reading scope is retained in the article. No fresh source verification is claimed.
Industry Insights
Editorial analysisNo separate items were recorded in this section of the original manuscript.
Research Briefs
4 papersThe paper uses an LGBTQIA+ community case to explain participatory curation of speech datasets and records concern about harmful technologies such as gaydar. For arXiv 2609.25496, the API publication timestamp is September 21 at 23:55:24 UTC. Title, abstract and time were checked through export.arxiv.org, rather than inferred from the HTML listing batch date; the individual abstract HTML page was not opened. Source: https://arxiv.org/abs/2609.25496
The abstract examines why quantization errors cannot be treated as isotropic when post-training quantization methods designed for unimodal language models are applied to vision-language models. For arXiv 2609.25492, the API publication timestamp is September 21 at 23:42:52 UTC. Only the API title, abstract and timestamp were reviewed; the abstract HTML page was not opened. Source: https://arxiv.org/abs/2609.25492
EvoOntology appeared on the September 21 Hugging Face Daily Papers list. The API abstract for arXiv 2609.15779 was read; the abstract HTML page was not opened. Listing date is a discovery date, not a claim of first publication. Source: https://arxiv.org/abs/2609.15779
CodeMidas appeared on the September 21 Hugging Face Daily Papers list. The API abstract for arXiv 2609.22068 was read; the abstract HTML page was not opened. Source: https://arxiv.org/abs/2609.22068
Open Source Spotlight
0 projectsNo separate items were recorded in this section of the original manuscript.
Historical coverage record · 16/18 checks recorded
The historical GitHub daily ranking could not be replayed. No September 22 ranks or daily star gains were invented; the open-source coverage retained only the Hugging Face article already listed above. Editorial context is included with the news stories rather than supplied as a separate new analysis.
The 24-hour window ended September 22 at 05:52:00 UTC. The original run recorded 18 planned sources: 16 ok, one partial (Microsoft), one failed (DeepMind), and no core source below ok after the xAI browser follow-up. DeepMind’s RSS URL returned non-XML content both initially and on retry. It was not a core source; successful parsing of the same address in the next two issues was not treated as a reason to change the configuration.
Microsoft redirected to https://news.microsoft.com/source/topics/ai/; titles were visible but dates were not, so membership in the window was not established article by article. Day-only dates for Opus 5.5 and the Hugging Face blog could fall after 05:52 UTC. Unknown times remain unknown; the issue neither shifts them to another day nor asserts they preceded the cutoff.
Additional checks of StepFun, Moonshot AI, Z.ai and Tongyi Lab were made September 25 and do not enter the 18-source denominator. Step 5 Preview had already appeared in Issue 15, with no new release in this window; the other three homepages or news pages showed no new announcement within it. Readable, checkable originals from Caixin, Yicai and The Paper were not obtained, so no Chinese brief was added solely from relay sites.
The Chinese narration script is daily/scripts/2026-09-22.txt. The historical Chinese episode is Issue 16, 10 min 55 sec: https://www.xiaoyuzhoufm.com/episode/6ab5ef68195d838e2ae9f03a . Its AI-generated-audio declaration was enabled and the web player’s direct link was filled. This English edition is a later translation of that reconstructed record.
Coverage gaps reflect incomplete checks; they do not establish that a source had no updates.
- Google DeepMind — Access or collection failed. The DeepMind feed returned non-XML content on both the initial attempt and retry.
- Microsoft AI — Partially checked. The accessible Microsoft list did not provide reliable publication dates or establish complete coverage of the original discovery window.
Reviewing a source does not constitute independent testing. Paper-reading scope is stated with each item; experiments were not reproduced.