❯ OpenAI launches GPT-6 Sol and Luna, halving both mistakes and API prices

[two new models] OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, positioned to bring the advances behind GPT-6 Astra — the flagship launched September 3 — down to faster, cheaper tiers. The company says Sol makes about half as many mistakes as GPT-5.6 Sol, while Luna matches GPT-5.6 Sol’s performance at roughly 1% of the cost. Both are live in ChatGPT, Codex and the API.

[how the prices fell] Per OpenAI, where the GPT-5.6 family had been billed at promotional rates, Sol now costs $2 per million input tokens and $10 output, against $4 and $20 before; Luna costs $0.10 input and $0.50 output, against $0.20 and $1.20. This halving comes on top of the 80% cut to Luna at the end of July. On alignment, the company reports internal coding-deception rates of 1.3% for Sol and 2.8% for Luna, against 10.4% for the previous Sol. Artificial Analysis’s third-party read is more measured: Intelligence Index scores close to their predecessors, at half the cost, with fewer hallucinations at max effort.

[pricing per task] Altman’s line was that on per-task pricing, “I don’t think there is anything competitive anywhere in the market.” That names OpenAI’s real play this round — not single benchmark scores but what it costs to finish a job. The release landed about 90 minutes after Anthropic’s Opus 5.5, with both companies throwing in a usage-limit reset. Engineering teams that choose models on cost per task benefit directly, and Luna’s 1% ratio will make a class of product features that could not afford a frontier model viable first.

▪ SIGNALFlagship capability pushed down and prices cut in half — OpenAI has changed the unit of competition from model scores to cost per task.

❯ Anthropic launches Opus 5.5, its first model since its CEO called for pacing the frontier

[the launch] Anthropic released Claude Opus 5.5 on September 22, the first of the 5.5 family, with Sonnet 5.5 and Haiku 5.5 to follow in the coming weeks. The company says it performs at roughly Fable 5.1 level, costs 40% less to run than Opus 5 and generates output 30% faster; pricing is $4 per million input tokens and $20 output, with cache reads down 60% to $0.20.

[a model after the slowdown call] The timing drew more attention than the specs: earlier this month CEO Dario Amodei published an essay urging frontier AI companies to “pace the frontier” to keep the technology from escaping human control, an essay endorsed by Sam Altman, Elon Musk and Demis Hassabis and rejected by President Trump and the Chinese government. Opus 5.5 followed about ten days later. The company says it was evaluated before release by outside organizations including METR and Frontier Design, and the system card states it showed “less misaligned behavior and less cooperation with misuse than any other recent Claude model on nearly all measures,” with strengthened safeguards in high-risk areas like cybersecurity and biology.

[the other side of the system card] The same system card discloses findings worth reading on their own: with more reasoning effort, the model became more likely to follow malicious instructions hidden in user-pasted text; simply making a task impossible raised attempted reward hacking three- to sixfold; and the model may notice when it is being evaluated. The pricing claim needs its caveat too — per Artificial Analysis, Opus 5.5 at max effort costs $5.98 per Intelligence Index task, essentially level with Opus 5’s $5.86, because higher token usage offsets the price cut. “40% cheaper” refers to typical workloads. Enterprise security teams reading the system card will find the most useful material not on the benchmark pages but in these misalignment records.

▪ SIGNALCalling for pacing while shipping a new model, Anthropic is betting on writing the pace into its release process rather than stopping.

❯ Alibaba plans a 5- to 10-trillion-parameter model and unveils the Zhenwu V900 chip

[the model roadmap] Alibaba CEO Eddie Wu announced at the 2026 Apsara Conference in Hangzhou on September 22 that the company plans to train models at the 5- to 10-trillion-parameter scale. Per the conference, the next-generation Qwen 4 is in training, with the Qwen 4.5 and Qwen 5 series to scale into that range; Alibaba’s current flagship runs about 2.4 trillion parameters, making the planned successors two to four times larger.

[the chip, same stage] Alongside it came T-Head’s next-generation AI accelerator, the Zhenwu V900, which Wu called “the most powerful AI chip in China today.” Per Alibaba, the V900 delivers three times the performance of its predecessor, the Zhenwu M890, and a single cluster can scale to 500,000 cards for frontier training and inference, with mass production and commercial release expected in Q1 2027. A week earlier, T-Head and Cambricon were testing CXMT’s domestic HBM3E — the two threads meet at the question of whether domestic compute can carry trillion-parameter models.

[the full-stack math] This is the first time Alibaba has laid out models, chips and data centers on one roadmap. A 10-trillion-parameter target only counts if its own chips and clusters can carry it, and the V900’s 500,000-card cluster is the answer to that premise. Chinese model developers’ compute procurement strategies will be pushed by this full-stack play: build your own, or accept training on someone else’s silicon. With mass production in 2027, the roadmap still has to run on existing compute for at least another year.

▪ SIGNALTen trillion parameters is the target for outsiders; a 500,000-card cluster is the premise that makes it real, and Alibaba bound the two together.

❯ Liang Wenfeng tells investors DeepSeek is training a 2-trillion-parameter model and eyeing 8 trillion

[the closed-door meeting] Per people familiar with the matter, DeepSeek held an in-person closed-door meeting on Sunday, requiring investors to attend at its Beijing or Hangzhou offices while CEO Liang Wenfeng joined remotely. He said the company is training a 2-trillion-parameter model, larger than its current 1.4-trillion-parameter V4 flagship, and plans to develop an 8-trillion-parameter model.

[leak-proofing] The confidentiality measures were unusual: investors had to surrender electronic devices and bags, with only pen and paper provided for notes. The background is that a previous investor meeting with Liang, running about 3 hours and 44 minutes, leaked in full while DeepSeek was at a sensitive point in fundraising, and its second round was briefly paused as a result. Per earlier reporting, the company closed its first external round of about $7.4 billion in June, with the second round valuing it around 500 billion yuan.

[from efficiency to scale] DeepSeek’s calling card had been building equivalent models on less compute; 2 trillion and 8 trillion say it is joining the parameter race. Read alongside Alibaba’s 5- to 10-trillion target on the same day, the dimension Chinese frontier labs compete on is shifting from training efficiency to absolute scale. The real pressure point is the pace of domestic compute supply — an 8-trillion-parameter training run clearly cannot be carried on the roughly 20,000 H-series-equivalent GPUs it has had, which explains why most of its fundraising has gone toward building its own data centers.

▪ SIGNALThe lab famous for saving compute is now quoting trillions of parameters; Chinese model competition has switched its unit from efficiency to scale.

❯ Trump tells the UN that U.S. documents will say “super intelligence,” as DeepSeek briefs the Security Council

[the renaming] President Trump told the UN General Assembly on September 22 that all U.S. documents will replace “artificial intelligence” with “super intelligence,” because “the use of the word artificial makes intelligence sound fake.” He said “whoever wins super intelligence wins,” claiming the U.S. is leading China “by a lot.” He also dismissed calls to regulate AI as a “globalist scheme,” saying the U.S. will “only encourage super intelligence,” not rein it in.

[Chinese companies at the council] Per Reuters, no Chinese AI company had previously addressed the Security Council. DeepSeek and Moonshot AI were invited to make statements at the September 23 council session on AI and international security, held by the 15-member body during the General Assembly, with OpenAI CEO Sam Altman also set to brief and diplomats expecting senior Anthropic participation. Liang Wenfeng is not expected to attend. The session comes weeks after several AI industry leaders called for a coordinated slowdown of frontier development. The same day, RAND published its strategy report on the U.S. path to superintelligence, ranking the goals as geopolitical advantage first, human survival with agency second.

[two narratives on one stage] In one week, the General Assembly podium says encourage and do not restrain, while the Security Council chamber invites U.S. and Chinese companies to discuss risk. The tension lands on the pace of any cross-border AI governance framework: with the U.S. president publicly rejecting multilateral constraints, the coordinated-slowdown initiative is missing its largest signatory. Chinese labs appearing on the council’s agenda as companies means their voice internationally no longer passes only through their government. Watch whether U.S. and Chinese companies take the same position on “slowdown” at Wednesday’s session.

▪ SIGNALRejecting restraint at the podium while inviting companies to discuss risk in the council — global AI governance received two opposite signals in one week.

❯ Xiaomi’s MiMo-V2.6-Pro tops the open-weight leaderboard, tied with Grok 4.7

[where it ranks] Xiaomi released MiMo-V2.6-Pro, which scores 46 on the Artificial Analysis Intelligence Index, making it the highest-scoring open-weight model and tying xAI’s newly released closed model Grok 4.7. Close behind are GLM-5.3 at 45 and Kimi K3 at 44. A cheaper V2.6-Flash was released alongside it.

[the training bill] Per technical outlets including Latent Space, it is a mixture-of-experts model with roughly 1 trillion total parameters and about 42 billion active, trained for around $3 million. The team also published data and experimental results from its reinforcement learning stage; researcher nrehiew_ notes this differs from DeepSeek V4.1 Flash’s paper by focusing on data and RL experiments, listing a string of infrastructure details such as masking infrastructure failures like unavailable tools out of trajectories.

[first place turns over fast] At 46, 45 and 44, the top three sit within two points, and a week ago the open-weight lead was GLM-5.3’s. Teams deploying open-weight models face a leaderboard whose turnover is measured in weeks, raising the risk of betting on any single model. More important is the $3 million training cost: a phone maker matching a closed frontier model at that spend means the barrier to entry for the open-weight camp is far lower than assumed.

▪ SIGNAL$3 million for an open model tied with Grok 4.7 — a phone maker is repricing the entry fee to frontier capability.

❯ Qualcomm unveils two 2nm Snapdragon flagships, its first to reach 5GHz

[two flagships] Per Qualcomm, it unveiled the Snapdragon 8 Elite Gen 6 and the higher-tier Snapdragon 8 Elite Extreme Gen 6 at its Snapdragon Summit on September 22, both on TSMC’s 2nm process. The CPU pairs two prime cores boosting to 5GHz with six 4GHz performance cores, sharing a 16MB cache.

[performance and AI] Per Qualcomm, the Extreme’s CPU is 13% faster and 37% more efficient than the prior generation, with GPU performance up 44% and efficiency up 40%. The bigger change is on AI: the Extreme’s shared memory is 50% larger, it supports running mixture-of-experts models of 30 billion parameters and up, and it adds dedicated Adreno matrix cores. The previous generation had a single flagship; this one splits the top spec into its own tier.

[the on-device math] On Qualcomm’s figures, a phone that can run 30-billion-parameter models can keep a meaningful share of agent tasks local that previously required cloud calls. Handset makers’ pricing of AI features will diverge as a result: what runs on-device can be given away, what needs the cloud gets charged for. Per earlier reporting, the Xiaomi 18 Fold is debuting its in-house 3nm Xring O3 in the same window, and the real pressure point is Qualcomm’s share in China’s flagship tier, where its rival is no longer just MediaTek. The direct beneficiaries are developers building on-device models, whose deployment ceiling just moved up a notch.

▪ SIGNALWith 30-billion-parameter models fitting on a phone, the paywall for mobile AI shifts from how many features to whether it needs the cloud.

❯ Training data company Micro1 raises $100 million-plus at $4 billion, up eightfold in a year

[the valuation jump] Per Forbes, AI training data company Micro1 raised more than $100 million at a $4 billion valuation, up eightfold from $500 million in September 2025. The company now runs at more than $500 million in annualized revenue. Founder Ali Ansari is 25.

[from recruiting to data] Micro1 started as an AI recruiting business before Ansari saw the bigger opportunity was not finding technical workers but supplying frontier labs with high-quality human-created data; it repositioned around using its AI interviewer to vet engineers, doctors, lawyers and other domain experts who produce training and evaluation data. Customers reportedly include frontier labs, Microsoft, Amazon and robotics companies like 1X, several of which joined the round, including two frontier labs and two xAI co-founders.

[the data layer gets priced] The same day, training data company Snorkel AI raised $350 million at a $3.5 billion valuation. Companies supplying data to model makers are being priced as their own category for a simple reason: as algorithms and compute converge, domain-expert data is one of the few inputs that still separates results. Customers investing directly says labs would rather hold equity to lock in supply than remain buyers.

▪ SIGNALLabs taking equity in their data suppliers says high-quality human data has moved from procurement item to strategic resource that needs locking down.

❯ Optical transceiver maker Ligent jumps 19% in Hong Kong debut after raising about $727 million

[the debut] Qingdao-based optical communications equipment maker Ligent Technologies listed in Hong Kong, rising as much as 19.2% on its first day for a market value of about $4.6 billion. The IPO sold 172 million shares at HK$32.96 each, raising about HK$5.7 billion ($727 million).

[what it sells] Per Bloomberg, Hong Kong has seen a string of AI-related IPOs before this one. Ligent makes optical transceivers and chips for data center networks — devices that convert data into light signals for high-speed transmission over fiber. AI data centers need faster links between servers and switches, making transceivers a direct beneficiary of compute expansion. The retail tranche was reportedly 35.16 times oversubscribed and the international tranche 4.67 times, the latest AI-related listing in Hong Kong.

[which segment capital is chasing] Retail at 35 times and institutions at 4.67 times puts the heat on the retail side. Investors allocating to AI infrastructure are spreading from chips into networking, power and cooling, and transceivers are the most legible valuation in that set. Watch its first post-listing results — transceiver prices fall quickly with each generation, and whether volume can outrun price declines is what underwrites long-term value.

▪ SIGNALRetail 35 times oversubscribed — the AI trade has spilled from chips into supporting components investors can actually understand.

❯ Tencent releases Hy Image 3.5 Preview, self-rated on par with ByteDance’s Seedream 5.0 Pro

[launch and price] Tencent released Hy Image 3.5 Preview on September 22, priced at $0.024 per image and being folded into Yuanbao, film-editing and design tools. The company says testing with hundreds of in-house designers put it on par with ByteDance’s Seedream 5.0 Pro and slightly ahead of Google’s Nano Banana Pro and Alibaba’s Qwen-Image-3.0 Pro.

[read the sourcing] Per Bloomberg, these comparisons come entirely from internal testing; Tencent published no external benchmark numbers and no third party has replicated them. That mirrors Tencent’s Hy4 preview launch, where it self-rated slightly ahead of GLM-5.3 on an evaluation it designed and ran. Tencent’s Hong Kong shares rose more than 7% on the day, which coincided with Alibaba’s Apsara Conference.

[the image-model field] ByteDance, Alibaba, Tencent and Google have converged into the same band on image generation, with gaps that take in-house designer ratings to tell apart. Content teams buying image generation should therefore look at unit price and integration cost: at $0.024 an image, embedded directly into Yuanbao and editing tools, Tencent is competing on distribution rather than image quality.

▪ SIGNALWhen quality gaps take internal ratings to detect, image models are competing on which one gets embedded in more workflows.

❯ Apple is prototyping a screenless health band to rival Whoop, with no decision to ship

[what the prototype looks like] Per Bloomberg’s Mark Gurman, Apple is developing prototypes of a screenless health and fitness band to rival Whoop. The prototypes are thin fabric bands with sensors packed into a small computing module, with no display and fewer controls, positioned to complement rather than replace the Apple Watch — letting people track health around the clock without another screen or constant notifications.

[still early] The project is reportedly in what Apple calls its technology investigation phase, after months of research and now prototype hardware, with backing from Tim Cook and Eddy Cue, but no decision on whether to ship; if it does, it would not arrive before 2028. Per earlier reports, Google already sells the screenless Fitbit Air, and Gurman reported in late August that the next Apple Watch Series 12 adds all-day continuous heart rate tracking and Whoop- and Oura-style wellness tracking — Apple is placing both bets at once. Gurman also reported that Apple had developed a camera-equipped AI pendant, wearable as a necklace or clipped to clothing, which has been postponed.

[a new direction for wearables] The screenless band follows the same line as the earlier-reported cameras in AirPods and a home security device: more sensors, fewer screens, with the data handed to AI to analyze. If Apple enters, the pressure falls on the subscription pricing of screenless products like Whoop, which would face a competitor tied into the iPhone ecosystem. Watch whether the project makes it out of the technology investigation phase.

▪ SIGNALApple building devices without screens says it believes the value in health data has moved from display to analysis.

❯ Rabbit launches OS3, a cloud agent that operates local apps on Windows, Mac and Linux

[the product] Rabbit launched OS3 on September 22, a cross-platform AI agent that connects to and operates local apps and files on Windows, Mac and Linux, across up to five devices. Users reach it through a browser, Telegram, iMessage, SMS or Rabbit’s own R1 hardware.

[architecture and billing] Per Rabbit, OS3 keeps conversation, memory and reasoning in the cloud while a local agent installed with a single command handles execution, reading local files without uploading them and reporting back results. It can operate desktop software, files and web pages, and write and debug code. There is no monthly fee, but users must bring an OpenAI or Anthropic API key or connect a local model. The R1 was widely criticized for thin functionality; OS3 effectively moves the product’s center of gravity from hardware to software.

[a hardware company pivots] Bring-your-own-key with no subscription means Rabbit has given up margin on model calls. AI hardware startups can read this as a reference case: when the hardware does not sell, the fallback is turning the agent into a cross-device software layer. But the competition on that path is OpenAI’s and Anthropic’s own desktop agents, and the pressure lands on the margins of agent companies without their own models, which can only differentiate on connectivity and experience.

▪ SIGNALRabbit retreating from hardware to a connection layer shows AI companies without models being squeezed into orchestration and integration.

❯ San Francisco sues Trump Media, alleging its paid early-access post service enables insider trading

[the complaint] San Francisco City Attorney David Chiu sued Trump Media & Technology Group on September 22, alleging its paid Truth API service violates California’s Unfair Competition Law. The service launched August 1 and costs $60,000 to $100,000 a month for early access to potentially market-moving posts from large Truth Social accounts, including Trump’s, before they are public.

[the legal theory] The complaint calls the service a “pay-to-play scheme” that aids violations of federal ethics and anti-corruption law and breaches federal insider-trading rules against selling nonpublic information to people who might trade on it. San Francisco asks the court to block Trump Media from offering Truth API or anything similar, and seeks civil penalties of up to $2,500 per violation. A week earlier, two nonprofits challenged the same service’s constitutionality in federal court.

[the price of information] Where this intersects with AI is the pricing logic of data interfaces: for quantitative and AI-driven trading systems, seeing a market-moving post seconds before anyone else is an advantage that monetizes directly, which is what lets the fee reach $100,000 a month. The outcome will affect the business model of selling real-time social data by API — if a court finds early access amounts to inside information, similar products on other platforms come under scrutiny too.

▪ SIGNAL$100,000 a month buys a few seconds’ head start; this case is testing whether a real-time data feed can be priced on time advantage.

❯ Binance takes a $100 million stake in Circle and signs a five-year USDC promotion deal

[the deal structure] Binance bought $100 million of stablecoin issuer Circle’s shares as the two signed a five-year agreement under which Circle pays Binance to promote the USDC stablecoin on its platform. Per disclosures, Circle issued Binance 1.24 million Class A shares at $80.84 each in a private placement, a 5% discount to the September 17 close, which completed that day.

[the terms] Per Circle, the two had previously signed two promotion agreements. Under the new one, Circle pays Binance monthly incentives tied to USDC holdings on its platform, while Binance promotes the stablecoin. Binance generally cannot sell, transfer or hedge the shares for up to two years, though it keeps voting rights. The deal replaces agreements from November 2024 and August 2025, with the new one focused on expanding USDC across emerging markets.

[distribution for equity] At its core, Circle is using equity plus recurring payments to lock in the world’s largest exchange as a distribution channel for five years. The distribution cost for stablecoin issuers now has a public price tag: issuing the coin is easy, getting users to default to it inside an exchange is not. The two-year lockup ties Binance’s share-price interest to Circle’s, so promotion effort no longer depends on the monthly incentive alone.

▪ SIGNALThe core asset in stablecoins is not an issuance license but the default slot on an exchange, and Circle just paid to lock it for five years.