Tuesday was a price day. Anthropic shipped the first model in its Claude 5.5 family and cut the cost of running it, then OpenAI answered within hours by halving the list price of its mid-tier GPT-6 models. Neither company framed the move as a discount war, and both are right that cheaper inference changes what people are willing to automate. The rest of the week's news sits downstream of that same arithmetic: a trillion-parameter open-weight model from Xiaomi, a memory maker redirecting its factories toward the part of the market that is short, and a Japanese conglomerate borrowing more than $11 billion to keep paying for its stake in OpenAI.
Below: seven stories from September 20 to 23, 2026.
1. Anthropic ships Claude Opus 5.5 at 40% lower running cost, its first release since calling for a slowdown
Anthropic released Claude Opus 5.5 on September 22, the first model in a new Claude 5.5 family and the company's first launch since chief executive Dario Amodei published an essay asking the industry to pace itself. Anthropic says the model matches Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, with list pricing at $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5's $5 and $25. Cache reads dropped 60% to $0.20 per million tokens, and output arrives more than 30% faster.
The safety story is the part Anthropic spent the most words on, and it comes with outside verification attached. Opus 5.5 was tested before launch by METR and Frontier Design, and Anthropic says it posted the strongest score to date on its automated behavioural audit, the company's most comprehensive alignment test. In a dedicated containment evaluation, Anthropic reports the model was about 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass containment boundaries. Because its cyber capabilities are strong enough to matter, most cybersecurity tasks get re-routed to Opus 4.8, and the company is extending its Cyber Verification Program to cover Opus 5.5 across three tiers of increasingly permissive trusted access. Biology work carries the same safeguards as Fable 5.1, with a new Life Sciences Verification Program for vetted labs, startups and pharmaceutical companies. Opus 5.5 also ships with preserved thinking, the anti-distillation measure introduced with Fable 5.1 that stops API users from editing Claude's prior context to extract reasoning, applied to accounts created on or after August 31.
The benchmark Anthropic chose to lead with is CursorBench, where it says Opus 5.5 scores 11 points above GPT-5.6 Sol at roughly one-third the cost per call. In an internal test that asked three models to write a report on a company's quarterly performance using a copy of the web where the earnings release was hard to find, with an automated grader checking every figure against sources, 16 of Opus 5.5's 18 reports cleared the quality bar and neither Fable 5.1 nor Opus 5 managed one. Walleye Capital, an investment firm that tested early, reported that the model largely solved its evaluation suite at the lowest effort setting and at higher settings found an error in the evaluation instructions themselves. Claude Sonnet 5.5 and Claude Haiku 5.5 follow in the coming weeks, and Anthropic raised five-hour usage limits on Pro, Max and Team plans and added a rate-limit reset subscribers can bank and spend when they need it. The model is live on AWS, Google Cloud and Microsoft Azure.
— Anthropic · Reuters
🔗 Anthropic · Reuters · CNBC TV18
2. OpenAI answers within hours with GPT-6 Sol and GPT-6 Luna, priced 50% below its own promotional tier
OpenAI expanded the GPT-6 family on September 22 with GPT-6 Sol and GPT-6 Luna, less than three weeks after launching GPT-6 Astra as the flagship. Sol is priced at $2 per million input tokens and $10 per million output tokens, half of GPT-5.6 Sol's promotional rate. Luna goes further at $0.10 input and $0.50 output per million tokens. OpenAI attributes the cut to better caching and inference efficiency and says it is passing the savings through rather than buying share with a temporary discount.
The product framing is about reach, not a new ceiling. Both models are described as built on the technical progress behind Astra, aimed at professional work, coding, automation and computer use, and OpenAI says GPT-6 Sol roughly halves the error rate of the model it replaces. Access runs through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, plus the API, while free and Go tier users get GPT-6 Luna in the desktop app. Neither model is available in the plain ChatGPT chat interface yet, which keeps the cheap capacity pointed at agentic and work surfaces where token counts are high.
Read the two announcements together and the pattern is hard to miss. Anthropic cut 20% off list and 60% off cache reads on the same day OpenAI halved its mid-tier pricing, and both companies are competing for the same long-horizon coding and computer-use workloads where a small per-token difference compounds across thousands of tool calls. Cache pricing is the number to watch, because it is what decides whether an agent can afford to keep a large context resident across a session. For teams that standardized on one vendor, Tuesday was a reminder that the price of the same capability can fall by half between quarters, and that the switching cost is mostly integration work rather than model quality.
— OpenAI · 新浪科技
3. Xiaomi open-sources a 1.02T-parameter model under MIT, and streams the training run that built it
Xiaomi released and open-sourced the MiMo-V2.6 series on September 22. The flagship MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion activated per token over a million-token context, and MiMo-V2.6-Flash cuts that to 309 billion total with 15 billion active. Both take text, image, video and audio as input and return text. Weights for both reinforcement-learned checkpoints plus a MiMo-V2.6-Distill-Qwen-9B checkpoint are on Hugging Face under an MIT license, alongside a technical report, more than 7,000 RL task environments and a set of composable cybersecurity harnesses.
Artificial Analysis scored Pro at 46.32 on version 4.3 of its Intelligence Index, which the company says made it the highest-scoring open-weight model at release, ahead of Z AI's GLM-5.3 at 45 and level with xAI's proprietary Grok 4.7, while still trailing Claude Fable 5.1 and GPT-6 Astra. The benchmarker put the operating cost at $0.13 per index task and spent $206.66 running the full evaluation. List pricing is unchanged from the previous generation at $0.435 per million uncached input tokens and $0.87 per million output, with cached input at $0.0036. A separately priced UltraSpeed mode claims up to 20 times faster output at ten times the standard rates.
The training disclosures are the unusual part. Xiaomi live-streamed a six-day run that it says cost about $2.62 million for Pro and $850,000 for Flash, with each model completing 30 reinforcement-learning updates and roughly 750,000 trajectories. Each update processed 1,568 prompts with 16 rollouts apiece, around 25,000 trajectories and 2.7 to 3.7 billion tokens, drawn from a shared pool of coding, general-agent, visual and cybersecurity environments rather than separate runs per skill. Graders were given more compute so that comparisons among successful trajectories rewarded answers that were more accurate, shorter and more token-efficient. On the held-out DeepSWE v1.1 software engineering benchmark, Pro climbed from 58.4 to 72.57 during the run and Flash from 48.8 to 65.68. Xiaomi also reports that Pro helped researchers formalize the main theorem of the paper Period Three Implies Chaos in Lean 4, producing more than 6,000 lines that the kernel accepted with no unfinished proof placeholders, despite no Lean-specific post-training.
Two numbers in the release deserve skepticism. MiMo-V2.6-Flash scores 95.1 on CyberGym against the flagship's 94.0, and a smaller model beating a larger one on a security benchmark usually means the benchmark has stopped discriminating. CyberGym's top tier hands the agent a source diff and a patched codebase and asks it to reproduce a crash, which is guided reproduction rather than discovery, and the ordering snaps back on evaluations that require turning a bug into a working exploit: Pro scores 47.9 on ExploitBench against Flash's 25.3, and 17.8 against 6.0 on ExploitGym. The open weights are still the headline for anyone who wants a trillion-parameter model they can inspect, but the capability that security teams actually worry about remains behind closed doors at every lab.
— Xiaomi MiMo · Artificial Analysis
🔗 Xiaomi MiMo on Hugging Face · Silicon Report · ThreatFrontier
4. JetBrains puts its agent tooling under one name and bets on a vendor-neutral control layer
JetBrains announced Air on September 22, a system of products for agentic software development that pulls together roughly six months of public experiments. Three products sit inside it. Air in JetBrains IDEs is the in-editor experience for directing agents and checking their output. Air Teams coordinates delivery workflows that mix people and autonomous agents. Air Governance, previously called JetBrains Central, covers organizational policy, visibility, auditability, cost management and accountability. The company's own agent, Junie, is supported across every Air surface.
The interesting design decision is that Air does not require JetBrains' agent. It works with Claude Agent, Codex, Copilot, OpenCode and any agent that speaks the Agent Client Protocol, an open standard JetBrains developed with Zed, across IDEs, web and CLI, with mobile marked as coming soon. Inside the IDE plugin, developers can keep several projects open with their own sessions and track unread updates, changed files and outgoing commits across all of them, let the tool find compatible agents already installed for one-click connection, open diffs straight from a session and leave comments on specific lines for the agent to act on. Model access is bring-your-own: an existing provider subscription, your own API keys, or JetBrains AI credits billed at public API rates, with switching between subscriptions allowed mid-session. Air Teams adds shared projects in dedicated cloud environments, automations for repeatable work like code review and release notes, per-project controls over VM size, internet access and secrets, and MCP servers configured once and shared across the team.
JetBrains states the multi-vendor stance as a principle rather than a hedge: models differ in what they are good at, rankings move every few months, and teams inside one company already make different choices. The cost of that openness today is real, which the company names directly. Context does not carry between tools, spending cannot be attributed, and policies have to be rebuilt for each new service. Air is the attempt to sit above all of it. The diagnosis underneath the launch is that individuals adopted agents faster than organizations built anything around them, and that the expensive problems now show up in review, rework, security and spend rather than in whether the code compiles. Obvious mistakes get caught quickly; the hard case is output that passes a surface check while carrying a wrong assumption that surfaces weeks later. Parts of Air have not shipped yet, and a neutral control layer is only worth as much as the third-party agents that agree to support it.
— JetBrains · Unite.AI
🔗 JetBrains Blog · Unite.AI · Agent Client Protocol
5. Samsung plans a 40% HBM ramp and pushes its newest memory to 80% of shipments
Samsung is preparing to lift average monthly HBM wafer starts from roughly 180,000 this year to about 250,000 next year, an increase near 40%, with the HBM4 family going from around 40% of shipments to roughly 80%, according to industry sources cited by the Seoul Economic Daily. A second indicator moved faster: outsourced cleaning of the glass carriers used during wafer thinning is set to rise from about 20,000 sheets a month this year to 50,000 next year, which is 2.5 times, up from 10,000 in 2025. Carriers are reused, so that figure is a signal about intent rather than a direct conversion into sellable memory, and the wafer number is the better denominator.
The reason Samsung can push this hard is that it builds more of the stack itself. HBM4 went into mass production in February 2026 on sixth-generation 10-nanometer-class 1c DRAM paired with a 4-nanometer logic base die made in-house, where SK Hynix uses TSMC's 12-nanometer process for its base die. Twelve-layer HBM4E samples went to customers including Nvidia in May. Yields reportedly climbed from below 60% at the start of mass production to around 80%, which is what turns extra wafer starts into extra qualified product. The base die is also competing for capacity: HBM4 base dies are said to consume between 50% and 60% of Samsung's available 4-nanometer lines, which are running near full utilization, and the company has raised prices on new 4-nanometer orders. The demand does not stop at memory customers. Nvidia's Groq 3 LPU also draws on that node.
The competitive picture has already shifted. Counterpoint data shows Samsung's share of HBM revenue rising from 21% in the first quarter to 33% in the second, while SK Hynix fell from 58% to 50% and Micron held around 18%. Samsung says HBM4 revenue will more than triple quarter over quarter and that HBM4 will exceed 60% of its HBM revenue in the second half. Nobody is standing still: SK Hynix is taking monthly wafer input at its M15X fab in Cheongju from 10,000 to 80,000 and accelerating its first Yongin fab for 2027, while Micron plans to add up to 60,000 wafers a month of HBM capacity by year end, reaching roughly 100,000 against an estimated 150,000 to 200,000 each at the two Korean makers. Citi expects global HBM bit demand to grow 62% next year, and inventories at Samsung and SK Hynix have reportedly fallen below 10 days. Micron reports guiding fourth-quarter gross margin to about 86%, which tells you what a shortage does to memory economics, and also why the buyers now sign multi-year agreements with price ceilings.
— Samsung · Seoul Economic Daily
🔗 Samsung Newsroom · TrendForce · CTOL
6. Unitree puts a 22-degree-of-freedom hand on sale for 39,900 yuan
Unitree released the Dex5-S dexterous hand on September 21, priced from 39,900 yuan, roughly $5,600, with 22 degrees of freedom per hand at one-to-one human hand scale. The dimensions are 102 by 187 by 26 millimeters and the hand weighs about 620 grams, down from the 1,000 grams of the Dex5-1 listed in the company's IPO prospectus. The degree count breaks down as five in the thumb, four each in the index, middle and ring fingers, and five in the pinky, which gains a roll axis over the previous generation. The shell is aluminum alloy.
Every joint is driven by its own motor with direct drive, backdrivability and dual-encoder closed-loop control, so an external force can move a powered joint instead of meeting a locked mechanism, and each joint carries peak impact torque protection to shield its gears. Unitree rates the hand at 2 kilograms peak load and 1 kilogram sustained, and calls out thermal management as a design goal because overheating alarms interrupt training runs. Power input spans 15 to 65 volts, communications run over gigabit Ethernet, USB or RS-485, and the gigabit path supports control frequencies up to 1,000 hertz. A palm camera can connect directly to the hand without external cabling, which simplifies integration on a robot. Tactile sensing is the Pro version only; the standard model reports position, torque and IMU data. The hand fits Unitree's H1 and G1 humanoids and can also be used as a standalone manipulation system.
The demos in Unitree's video cover opening bottle caps, pulling can tabs, solving a Rubik's cube, cutting paper with scissors, tossing a coin and playing piano keys. What matters for research teams is the price. A pair of Dex5-S hands starts near 80,000 yuan before controllers, mounts, tax and integration, well below most research-grade quotes, and Unitree already sells bodies to attach them to. The previous Dex5, released in April 2025, offered 20 degrees of freedom with 16 active and 94 tactile points and a stated maximum load of 4.5 kilograms, so the new model trades some payload for a smaller, lighter, more finely articulated hand. What the launch notes do not include is a full tactile map for the S variant, ROS and SDK packaging, or published mounting guides for the G1, H1 and H2, which are the documents integrators will wait for. The hand pairs with UnifoLM-WLA-1.0, the 6-billion-parameter embodied model Unitree open-sourced earlier in September after training on roughly 2,500 hours of real-robot data.
— Unitree · 证券时报
7. SoftBank borrows $11.15 billion in the largest APAC non-financial bond deal on record
SoftBank Group launched the sale of $10 billion of dollar-denominated and €1 billion, about $1.15 billion, of euro-denominated senior unsecured notes on September 21, according to a term sheet seen by Reuters. At the planned size it would be the largest non-financial corporate bond offering ever from Asia-Pacific and Japan, beating 7-Eleven's $10.93 billion sale in January 2021, and it would rank among the twenty biggest corporate bond deals globally this year. The dollar notes split into 3.5, 5.5 and 7.5-year maturities and the euro notes into four and six-year maturities, with pricing expected on September 24 and settlement on September 29.
The proceeds go to a specific line item. SoftBank needs $10 billion for the third tranche of its follow-on investment in OpenAI, scheduled to close on October 1, and the bonds cancel the $10 billion bridge loan it arranged for that payment. Completing the full investment takes SoftBank's cumulative stake in OpenAI to $64.6 billion and an ownership interest of about 13%. The financing is increasingly load-bearing for the group. SoftBank took a $40 billion bridge loan in March for earlier OpenAI commitments, raised an $11.87 billion two-year loan from roughly 20 banks the previous week, lifted a margin loan secured by Arm shares by $5 billion to $25 billion, and has issued close to $15 billion of bonds in various currencies this year, making it the largest junk-grade borrower in the 2026 bond market. Apollo is reportedly in talks to raise its loan to SoftBank to $9 billion. Fitch assigned the proposed notes BB+, below investment grade, and said it expects SoftBank's debt to rise as it funds committed investments while the group retains adequate liquidity and capital market access. SoftBank reported a loan-to-value ratio of 17% as of March 31 against a policy of staying below 25%.
The structure explains the timing. Bridge loans are short-term money meant to carry a borrower until permanent financing lands, and SoftBank is now moving a large piece of its OpenAI exposure onto its own balance sheet as long-dated public debt. Bondholders are being asked to finance an unlisted position whose value depends on OpenAI's valuation and eventual liquidity, and expectations for an OpenAI listing have been drifting toward 2027. That leaves the coupon as the honest read on how much compensation investors demand for lending against private AI equity, and the September 24 pricing will be the first hard number. The bet has already moved SoftBank's own shares: they fell as much as 13% in the week before the bond launch.
— SoftBank Group · Reuters
🔗 Reuters · FinanceFeeds · Tech Startups
What to watch next
Two of these stories are the same story told twice. Anthropic and OpenAI both cut the price of frontier-adjacent capability on the same Tuesday, and the interesting question is not who undercut whom but whether cheaper tokens change what teams are willing to leave running unattended. Watch cache pricing in particular, because it decides how much context an agent can keep resident, and that number moved 60% in Anthropic's favor. Xiaomi's release is worth reading in full for the training diary even if you never deploy the model, since a live-streamed six-day RL run with a published cost is still rare enough to be useful evidence. The dates to mark are September 24, when SoftBank's bond prices and tells you what credit investors think private AI equity is worth as collateral, and the weeks after, when Claude Sonnet 5.5 and Haiku 5.5 arrive and the price floor moves again.
KD Agentic publishes this digest daily. Previous editions cover model releases, agent frameworks, robotics and AI infrastructure.
Top comments (0)