Eight X signals: Claude Opus 5.5 and GPT-6 Sol and Luna arrive, and OpenAI asks the US to lead on AI standards

In the 24 hours ending September 23, 2026, eight of the fourteen AI and tech accounts this channel tracks produced something worth carrying, and every item below comes from one of them. The day's spine is price: Anthropic and OpenAI each released a cheaper model the same afternoon. Retweets, replies, jokes, and posts about something other than AI or technology are excluded.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, has left X, so its posts are unreadable and absent from this edition.

Claude Opus 5.5 opens a new Anthropic family at a lower price X post · 15,593 likes, 850,676 views at capture · posted September 22

Anthropic released Claude Opus 5.5 on September 22 as the first model in a new Claude 5.5 family, and says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 1. The price falls on both sides of a call: $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, with cache reads — the reused context that dominates the cost of long agent and coding sessions — at $0.20 per million, 60% below. Anthropic says the model generates output more than 30% faster than Opus 5.

Usage limits go up with it. Anthropic is raising five-hour limits on Pro, Max, Team and seat-based Enterprise plans, and says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks [cite:1].

The performance claims are Anthropic's own. It says Opus 5.5 scores highest of any model it has tested on its automated behavioural audit, an alignment suite run across thousands of simulated scenarios, and that external evaluators including Frontier Design and METR tested the model before release [cite:1]. Outside testers read it the same way: Ethan Mollick called Opus 5.5 the first model outside Fable and Astra that feels Fable-class, while saying the dense-language habit of the recent Claudes persists in it 2.

OpenAI halves the price of its cheap tiers with GPT-6 Sol and Luna X post · 49,680 likes, 7.38 million views at capture · posted September 22

OpenAI added two models to the GPT-6 line, Sol and Luna, trained with the methods behind GPT-6 Astra and priced 50% below the promotional prices of the GPT-5.6 models they replace 3. Sol goes to $2 per million input tokens and $10 per million output tokens, Luna to $0.10 and $0.50. Astra stays the best model in the family, and OpenAI says the two new models carry its improvements in professional work, factuality, coding, computer use and alignment into cheaper tiers [cite:3].

The number OpenAI leads with is cost per task. On AutomationBench, a test of end-to-end business workflows across 47 apps built by Zapier, GPT-6 Sol at its highest effort setting scores 33.2% against Claude Opus 5's 26.9% at maximum effort, at 9% of Opus 5's cost per task [cite:3]. Both models are in ChatGPT Work and Codex and in the API under the names gpt-6-sol and gpt-6-luna from today. Free and Go users can try Luna in the desktop app, and Chat will get both models later [cite:3].

On X, Sam Altman made the same argument in per-task terms — "half the price per token, and even less per task!" 4 — and Greg Brockman, OpenAI's president, framed it around that same set of capabilities 5. Altman also gave away the week ahead: getting ready for the company's next DevDay is "the first time I remember ever, in OpenAI history, saying 'this is too much stuff to launch'" 6.

Three new models in one night, and what they actually cost X post · 447 likes, 49,051 views at capture · posted September 22

Simon Willison wrote up both releases the same night and put the prices next to each other, which is where the scale of the cuts shows 7. GPT-6 Luna, at $0.10 in and $0.50 out per million tokens, is one-hundredth of Claude Fable 5.1's output price and among the cheapest models OpenAI has released, behind only GPT-4.1 Nano and GPT-5 Nano. Claude Opus 5.5 at $4 and $20 charges what GPT-5.6 Sol charged before OpenAI halved it, while Fable 5.1 and GPT-6 Astra, the models both labs still call their best, stay at $10 in and $50 out.

One detail changes how the cuts should be read. GPT-5.6 has a scheduled 25% price increase in November, which makes the comparison run against promotional prices that were already due to rise [cite:7]. Willison also points out where the ground moved under the day: Grok 4.7, released the day before at $2 in and $6 out, now matches GPT-6 Sol on input price [cite:7].

His own test of the new models came back empty. Asked to draw an SVG of a pelican riding a bicycle — a task he has run against generations of models — Claude Opus 5.5 at its "max" reasoning level returned nothing at all, twice, spending its whole 128,000-token output budget while still reasoning about the drawing, at about $2.56 and 20 minutes an attempt [cite:7]. Willison still switched his defaults to GPT-6 Sol and Claude Opus 5.5 [cite:7].

OpenAI asks the United States to lead on global AI standards X post · 5,781 likes, 578,107 views at capture · posted September 22

Sam Altman wrote that people outside the AI labs should have a real say in how the technology develops and a clear way to judge whether that is happening safely, and that standards should help prevent the concentration of power, including by keeping new companies and open-model developers able to compete 8. The post links a longer paper arguing that the United States should lead an international effort to write common technical standards for frontier AI 9.

The mechanism it proposes uses institutions that already exist: the network of national AI safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom, with the Center for AI Standards and Innovation coordinating, covering capability measurement, risk assessment and whether safeguards are sufficient [cite:9]. The paper rules out turning those standards into licences, mandatory pre-release reviews or approval requirements, and leaves national governments to decide whether to write them into law [cite:9].

The argument for urgency is recursive self-improvement, the loop in which AI systems take on more of the work of building the next generation of AI. The paper says fully autonomous self-improvement is not happening today and should not be pursued unless and until it can be done safely, and it points to the Hugging Face incident OpenAI disclosed as a preview of the risks that get worse without better safeguards [cite:9]. It also names a US–China conversation on frontier safety as a positive step.

What OpenAI will let outside assessors see X post · 2,624 likes, 509,195 views at capture · posted September 22

OpenAI published its priorities and principles for third-party assessment on the same day, committing to give independent assessors deep access across training, evaluation and deployment so they can challenge the company's assumptions, identify risks it may have missed, and reach their own conclusions 10. Four areas are named for that access: the safety cases the lab builds, the safeguards around internal and external deployment, the capability and alignment evaluations behind its Preparedness Framework, and independent investigation of misalignment incidents, written with the Hugging Face incident in view [cite:10].

The principles are where the offer becomes checkable. Claims for assessment are to be agreed and pre-registered before work starts; assessors get proportionate access within legal, security and intellectual-property limits; they must disclose financial and other conflicts of interest; they get time to work while the lab gets a window to fix what they find before publication; and reports are shared as openly as possible, with confidential reporting to oversight bodies when particular details have to stay private [cite:10].

The paper also sets its own scope: the assessments it describes are longer-term examinations of particular safety claims, run alongside government testing of models at release [cite:10]. OpenAI says the independent evaluation ecosystem is still small and that no single assessor should cover every frontier safety question.

Altman on what a growing company loses X post · 3,974 likes, 518,039 views at capture · posted September 22

Sam Altman agreed with an OpenAI colleague's description of how the company operates — assembling "an empowered group of highly capable people to work on the most important and urgent thing at any point of time", with nobody caring about org lines 11 — and said where he thinks the hard part is: "Startups are naturally good at this; it is hard to keep a bigger company good at this and i think an underexplored space." 12

Two minutes earlier he made the same point about a person. Quoting Michelle Pokrass on her four years at the company, he called her the embodiment of an outfit whose "constant capacity for rebirth" lets it make "high conviction, contrarian bets over and over" 13, and said people would be pleased by "what she + team have cooking next" 14.

Pokrass's own post puts a clock on the pace: the bets keep coming, they are mostly right, and the company is, in her words, "a new company every three months" [cite:13]. The same afternoon Altman was closing a launch he described as the largest he has seen at OpenAI, and telling developers to come back next week [cite:6].

A starship game with real orbital mechanics, built by a model and open-sourced X post · 212 likes, 18,506 views at capture · posted September 23

Ethan Mollick said he had Fable and Opus build a hard science-fiction starship combat game, with orbital mechanics, delta-v — the propellant budget that sets how far a ship can change its course — heat and realistic tactics, simplified to a 2D view with the hard mathematics done by the game itself 15. The result is Orbital Declaration, a game that ships with eight story chapters, a campaign set around Jupiter, standalone skirmishes and a ship configurator 16.

The premise is the physics that games usually skip. Orbital Declaration is set in the Jovian system in 2211, and its own description is a list of what it refuses to simplify: Newtonian motion, the rocket equation, heat that has to go somewhere, and no stealth at all [cite:16].

Mollick published the source on GitHub the same morning and credited the physics material the model researched for the simulation 17. He also named Children of a Dead Earth as the most realistic space game he has played, and said his only connection to it is as a player [cite:17].

Chollet's science prediction, five years on X post quoting his own post from May 2021 · 4,014 likes, 694,371 views at capture · posted September 22

François Chollet, who wrote the Keras deep-learning library and co-founded the ARC Prize, quoted a prediction he made in May 2021 and passed judgement on how it had aged: the post drew enormous pushback at the time, and "this prediction is looking obvious by the day now" 18.

The prediction was broad. Within 10 to 20 years, he wrote, nearly every branch of science will be, for all intents and purposes, a branch of computer science — computational physics, computational chemistry, computational biology, computational medicine, "even comp archeology" — because realistic simulations, big data analysis and machine learning will be doing the work everywhere it is done 19. His original post holds 6,187 likes.

The timing is the point of the re-up. The prediction went out in May 2021, before ChatGPT and the generation of reasoning models that followed, at a moment when the claim that a science department would run on computation sounded like a stretch. Chollet's judgement now is that it reads as ordinary [cite:18].