{"id":216,"date":"2026-09-23T10:06:21","date_gmt":"2026-09-23T10:06:21","guid":{"rendered":"https:\/\/scoy.ai\/guides\/ai-news-roundup-september-23\/"},"modified":"2026-09-23T10:06:21","modified_gmt":"2026-09-23T10:06:21","slug":"ai-news-roundup-september-23","status":"publish","type":"post","link":"https:\/\/scoy.ai\/guides\/ai-news-roundup-september-23\/","title":{"rendered":"AI News Roundup for September 23: Anthropic and OpenAI Cut Prices 90 Minutes Apart the Day Before Briefing the UN"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">Yesterday two frontier labs repriced their flagship models inside the same afternoon, and today both of them sit in front of the UN Security Council to explain why the industry should slow down. Most of the coverage I read treated those as two stories. They are one story, and the operator&#8217;s version of it is a lot less flattering than the press-release version.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The Two Price Sheets, Side by Side<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Here is what actually changed. Anthropic shipped Claude Opus 5.5 on Tuesday at $4 per million input tokens and $20 per million output, down from Opus 5&#8217;s $5 and $25. About ninety minutes later OpenAI answered with <a href=\"https:\/\/openai.com\/index\/introducing-gpt-6-sol-and-luna\/\" target=\"_blank\" rel=\"noopener\">GPT-6 Sol and Luna<\/a>, Sol at $2 and $10, Luna at $0.10 and $0.50. VentureBeat got an OpenAI spokesperson on the record confirming those are permanent rates rather than introductory pricing, which matters more than the numbers themselves.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Now the part almost nobody reported correctly. Anthropic&#8217;s headline claim is that Opus 5.5 costs 40% less than Opus 5, and the company&#8217;s own <a href=\"https:\/\/www.anthropic.com\/news\/claude-opus-5-5\" target=\"_blank\" rel=\"noopener\">Opus 5.5 announcement<\/a> is careful about what that means: &#8220;Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads.&#8221; That is a blended run-cost estimate, not a price cut. The actual sticker reduction is 20%. The gap between those two numbers is made up of token efficiency and, mostly, cache reads, which fell from $0.50 to $0.20 per million.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I have been running this content engine against Anthropic&#8217;s cache pricing long enough to say the obvious thing out loud: the 40% is real only if your workload looks like their workload. If you are making one-shot calls with cold context, you get 20% and nothing more. If you are running long agentic sessions against a stable system prompt, you might beat 40%. The headline number is an average of a distribution you are not necessarily in, and you should go measure your own cache hit rate before you put that 40% in a budget.<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\"><p>Sticker price is what you negotiate. Cache behavior is what you actually pay. Those have been separate numbers for a while now, and this week they drifted further apart.<\/p><\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">On raw price OpenAI is the cheaper option by a wide margin, and Sol at $2 and $10 undercuts Opus 5.5 by half on input. That is not the whole decision. Anthropic is claiming Fable 5.1 level output at the Opus tier, which is a capability-per-dollar argument rather than a dollar argument. My read: if your work is high-volume and well defined, Luna at ten cents per million input is close to impossible to argue against. If your work is long-horizon agentic coding where a failed run costs more than the tokens, the sticker price is the least interesting number on the page.<\/p>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<iframe loading=\"lazy\" title=\"Anthropic unveils Claude Opus 5.5 model\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/NSb7-UqGeTA?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe>\n<\/div><figcaption class=\"wp-element-caption\">CNBC Television, September 22: the Opus 5.5 launch and the 40% figure as it was reported on the day.<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">The Opus 5.5 Change That Is Not About Money<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Buried under the pricing noise, Anthropic says Opus 5.5 posted the best scores of any of its models to date on its internal automated behavioral audit, and improved resistance to prompt injection.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">That second one is the line I care about. I have written about agent security on this site more than once this month because it keeps being the thing that actually breaks production systems, and Cisco&#8217;s State of AI Security 2026 report puts the gap plainly: 83% of organizations plan to deploy agentic AI and only 29% feel ready to do so securely. OWASP still ranks prompt injection as the number one LLM risk. So a measurable improvement in injection resistance on a flagship model is worth more to me this quarter than a dollar off the input price.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic also raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and added a rate limit reset that subscribers can bank and spend when they choose. If you run Claude Code as your daily harness, as I do, that is the change you will feel first.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Sora 2 Dies Tomorrow, and Nothing Replaces It<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">This is the item to act on today. OpenAI&#8217;s <a href=\"https:\/\/developers.openai.com\/api\/docs\/deprecations\" target=\"_blank\" rel=\"noopener\">deprecations page<\/a> shows the Videos API and every Sora 2 alias and snapshot shutting down on September 24, which is tomorrow. That covers <code>sora-2<\/code>, <code>sora-2-pro<\/code>, and the dated snapshots. The recommended replacement column is empty. There is no migration target, because OpenAI is not replacing the product.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">If you built video generation on that API, you do not have a version bump ahead of you, you have a vendor swap.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A second date worth putting in your calendar: GPT-5.5 leaves ChatGPT, ChatGPT Work and Codex on October 14. The precision point that a lot of write-ups are getting wrong is that this retirement does not hit the API. If you call GPT-5.5 over the API you are fine for now; if your team has it pinned in workspace defaults, custom agents, saved settings or scheduled tasks, those break. And there is a nice irony in the official migration path, which points Codex users at GPT-5.6 Sol. As of yesterday that tier costs twice what GPT-6 Sol does. Migrating onto the recommended target means migrating onto a price tier that was superseded before the deadline arrived.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Check your pinned model strings this week. Not next month.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">The Slowdown Pledge Meets the Price War<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Today France, which holds the Security Council presidency for September, convened the 15-member council on AI and international security during the General Assembly&#8217;s high-level week. Sam Altman is addressing it. Dario Amodei is expected by remote link. DeepSeek and Moonshot were both invited to speak, which is the first time the council has hosted frontier Chinese and American developers together on shared safety concerns, and on that narrow point the session genuinely matters.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Here is the calendar, though. Amodei published his roughly 3,800-word &#8220;We Must Pace the Frontier&#8221; essay on September 12, arguing the industry should deliberately slow the rate of capability gain, and Altman agreed publicly within hours. Ten days later both companies shipped cheaper, faster flagship models ninety minutes apart. The next day they brief the UN on slowing down.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">I want to be fair about the technical distinction, because it is real: pacing frontier capability and cutting inference price are different axes, and you can honestly do the second while restraining the first. Opus 5.5 is positioned as matching an existing model more cheaply rather than as a capability jump, which is exactly what a pacing commitment would predict.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">But cheaper inference is how capability reaches production. Every price cut expands the set of tasks it is economically rational to hand to an agent, and Amodei&#8217;s own essay warned about rogue agent swarms threatening large parts of the internet within six to twelve months. You cannot cite agent proliferation as the risk and then halve the cost of proliferating agents in the same fortnight without someone asking about it. I will call the safety commitment sincere and the pacing framing incomplete. What got paced was the benchmark number. Deployment accelerated.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Alphabet Open-Sourced the Robotics Stack, and Almost Nobody Noticed<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The most underrated release of the week got a fraction of the attention the price war did. Intrinsic, Alphabet&#8217;s robotics software company, <a href=\"https:\/\/www.intrinsic.ai\/blog\/posts\/introducing-intrinsic-core\" target=\"_blank\" rel=\"noopener\">open-sourced Intrinsic Core<\/a> at ROSCon 2026 in Toronto under Apache 2.0, on GitHub, running on local hardware.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">What is in it is the interesting part. Not a demo, but the components Intrinsic says it uses in real manufacturing deployments: a real-time control framework that adjusts a robot&#8217;s path mid-move from sensor feedback, pose estimation built on Nvidia&#8217;s FoundationPose, collision-free motion planning, grasp planning that adapts to how an object is sitting, plus simulation, calibration and ROS 2 drivers.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Apache 2.0 on production industrial robotics components is a genuinely permissive license on genuinely hard software. If you are anywhere near physical automation, this is a bigger week than the token prices.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What I Am Watching<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The price floor is not done falling, and Luna at ten cents per million input tokens is the number that will reset expectations for everything built on high-volume classification and extraction. The UN session will produce a communiqu\u00e9 with no enforcement behind it, and the thing to watch is not what gets said today but whether the third-party evaluator access Amodei promised METR actually materializes. And if you take one action out of this roundup, make it the boring one: grep your codebase for hardcoded model strings before Sora 2 goes dark tomorrow morning.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Anthropic and OpenAI repriced their flagships 90 minutes apart, one day before briefing the UN on slowing down. The 40% cut is not what it looks like.<\/p>\n","protected":false},"author":1,"featured_media":215,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-216","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news"],"_links":{"self":[{"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/posts\/216","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/comments?post=216"}],"version-history":[{"count":0,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/posts\/216\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/media\/215"}],"wp:attachment":[{"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/media?parent=216"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/categories?post=216"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scoy.ai\/guides\/wp-json\/wp\/v2\/tags?post=216"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}