Reflection unveils Beam, the Pentagon stops using Anthropic, and 3 more AI moves
Five verified developments from October 5 to 6, 2026: Reflection AI unveils its first open-weight model but holds the weights back, OpenAI sets out how it will watermark text for the EU's AI Act and where it will stop, the Pentagon says it has stopped using Anthropic's tools, DeepSeek and Moonshot close in on 2027 listings, and Google contracts 3.59 GW of power from Constellation.
In brief
Between October 5 at 12:00 UTC and October 6 at 12:00 UTC, Reflection AI unveiled the first open-weight frontier model from a Western lab without releasing the weights; OpenAI set out how it will watermark text for Europe's AI Act and where it will stop; the US Defence Department said it has stopped using Anthropic's tools; China's two most-watched AI labs moved to the last private money before 2027 listings; and Google contracted 3,590 megawatts of power from Constellation Energy.
- Launches: Reflection AI, a New York startup founded in 2024 by two former Google DeepMind researchers, unveiled Beam, its first open-weight model — a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active for any given token, built for coding and agentic work. Reflection says Beam matches Z.ai's GLM 5.2 on advanced reasoning while using three to four times less inference compute, and that its weights, technical report and model card will arrive later this month under an Apache 2.0 licence. On the coding tests where both sides publish numbers, Beam scores 80.1 on Terminal Bench v2.1 against 88.3 for Moonshot's Kimi K3 and 90.6 for DeepSeek V4.1 Flash, so Reflection's pitch is cost per unit of work, and the top raw scores remain with the Chinese open models. 123
- Regulation: OpenAI published its approach to the EU AI Act's requirement that providers make generated text identifiable in a machine-readable way: API customers worldwide can opt in from today, and an invisible watermark will reach eligible ChatGPT and Codex output in the European Union over the coming weeks. Its method, textGrain, adds a statistical signal to the model's word choices, and OpenAI's own evaluations detected it in about 80% of 200-token passages at a 1% false-positive rate, rising to roughly 95% at 400 tokens and falling sharply for mathematics, where word choice is more constrained. Editing weakens the signal: replacing 10% of words with synonyms cut detection from about 92% to 66%, which is why the detector opens only to approved researchers and expert organisations. 45
- Companies: The US Defence Department has stopped using Anthropic's tools, a department official told the BBC on Monday, months after Defence Secretary Pete Hegseth designated the company a national security supply-chain risk and set a deadline to phase it out. People familiar with the Pentagon's use of artificial intelligence told the BBC that Claude was still handling research, analysis and intelligence gathering as recently as last week, including during military operations against Iran, and ran inside Palantir's Maven Smart System. Anthropic declined to comment; the company has called the designation unprecedented and unlawful, sued the administration, and won a ruling that it was retaliated against illegally. 6
- Funding: DeepSeek is close to securing at least 80 billion yuan (about $12 billion) in what it expects to be its final private round before an early-2027 listing, after seeking roughly 50 billion yuan and drawing enough demand that the total could approach 100 billion yuan, Bloomberg reported. Tencent and battery maker CATL committed among the largest amounts, and the round puts DeepSeek at a valuation of about 500 billion yuan; CATL declined to comment, Tencent gave no immediate comment, and Reuters said it could not verify the report. Rival Moonshot AI closed its final private round at a valuation of about $50 billion and is preparing a Hong Kong listing in the first quarter of 2027 that could raise up to $5 billion. 789
- Infrastructure: Google has contracted for 3,590 megawatts of power from Constellation Energy across PJM, the largest US electric grid, including a 20-year agreement for 890 megawatts of new nuclear capacity created by uprating 11 reactors in Illinois, Pennsylvania and New Jersey. The arrangement underpins more than $4.3 billion of new investment by Constellation, which will also deploy Google Cloud's Gemini Enterprise across its nuclear operations, and Google says the added capacity comes at no cost to other ratepayers. Electricity from the first upgraded plant is expected in 2028, with the full 890 megawatts due before the end of 2032. 1011
Five developments
1. A Western open-weight model arrives, without the weights
Reflection AI released Beam on Monday, its first model and the first serious open-weight frontier system to come out of a Western lab. The company describes Beam as a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for any given token, pretrained on 23.8 trillion tokens and aimed squarely at coding, reasoning and agentic work. TechCrunch has the context window at one million tokens, and the model is text-only, so it cannot look at an image the way Thinking Machines Lab's Inkling can. 112
"Open-weight" means the trained parameters can be downloaded and run by anyone, which is the whole point of the category: the strongest open models so far have come from Chinese labs, and Western enterprises that will not run those weights have had little to choose from. Reflection's co-founder and chief executive, Misha Laskin, told Semafor the buyers are businesses and governments building sovereign AI systems who "don't really have very good options today". The company says it is working with both the US Center for AI Standards and Innovation and the UK AI Safety Institute to have Beam's capabilities assessed. 12
The performance claim is narrower than the framing around it. Reflection says Beam is competitive with Z.ai's GLM 5.2 on advanced reasoning and approaching Alibaba's Qwen 3.8-Max on coding and agentic tasks, while using three to four times less inference compute; it also says Kimi K3 remains ahead on raw capability. On the tests where both Beam and its rivals report numbers, Beam sits close to GLM 5.2 and behind the leaders: 44.4 on DeepSWE v1.1 against GLM 5.2's 44.0 and DeepSeek V4.1 Flash's 74.2, and 80.1 on Terminal Bench v2.1 against 81.0, 88.3 and 90.6 for GLM 5.2, Kimi K3 and DeepSeek V4.1 Flash respectively. 13
The compute comparison deserves a second look before anyone plans around it. Reflection calculated it as roughly twice the active parameter count times the average number of generated tokens, and says the result is an approximate comparison that does not measure inference cost; the formula leaves out prompt processing, attention operations and serving overhead. The competing models' scores came from Artificial Analysis and DataCurve. A reader running agents around the clock will get a different ratio from their own prompts and hardware than the one on the chart. 13
Most of what the company disclosed concerns the training run. The reinforcement-learning stage ran four weeks on about 10,500 Nvidia GB300 GPUs, generated more than 100 million rollouts, drew on close to one million training environments and used roughly 1.3 billion sandboxes for training and grading. Reflection says it built new algorithms to keep asynchronous training stable when rollouts are generated by checkpoints a day behind the current policy, and that scores were still climbing when the run ended. 13
What a reader can act on today is limited, and that limit is the story. Beam is still in final red-teaming and evaluation; the weights, technical report, model card and developer artefacts are promised later this month, under an Apache 2.0 licence according to Help Net Security, and early access runs through a waitlist. Reflection has raised roughly $4.7 billion and was last valued at $25 billion pre-money in June, and it has compute deals with SpaceX and Nebius worth more than $7 billion. Until the weights are published and an independent index scores them, the efficiency figure is the company's own arithmetic. 2312
2. OpenAI marks its text in Europe and stops there
OpenAI published on Monday how it intends to satisfy the EU AI Act, which requires providers of generative AI systems to make generated text identifiable in a machine-readable way. The rollout has three parts. API customers worldwide can opt in to watermarking for selected models from today, and watermarking stays off by default there. An invisible watermark will be added to eligible ChatGPT and Codex text output in the European Union over the coming weeks, on all plans, and only in the EU. Applications opened on the same day for access to OpenAI's text watermark detector, which will go first to approved researchers and expert organisations, in line with the EU's Code of Practice on AI-generated content. 4
The mechanism is called textGrain, and it works on word choice, leaving nothing visible in the text. It adds a statistical signal as the model picks among words that would fit equally well, and a detector looks for that signal to judge whether a passage carries an OpenAI watermark. OpenAI says textGrain matched or exceeded the other approaches it tested, including Google's SynthID for text, and that it plans to release the technology as open source. The company's benchmark table shows no meaningful performance difference between watermarked and unwatermarked output from its frontier model Astra; if anything, watermarked text scored slightly higher on several of the eight benchmarks listed. 45
The limits OpenAI published are the part worth keeping. At a target false-positive rate of 1%, its detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages, and detection fell substantially for content such as mathematics, where word choice is constrained. Editing does more damage than length: replacing 10% of the words in a 400-token passage with synonyms reduced detection from about 92% to 66%, and replacing 25% cut it to 17%. 4
OpenAI also set out what a detection result cannot establish, which matters for anyone who expects watermarking to settle an authorship dispute. A watermark does not measure how much human judgment, editing or creativity went into a passage, does not establish who owns the text or whether its use was lawful, does not identify the user, account, prompt or conversation behind it, and does not verify accuracy. The absence of a detected watermark does not prove human authorship either, because text may be too short, edited, translated, generated by an unsupported model, or produced before watermarking existed. 4
The argument started with Anthropic, which said earlier this year that Claude would carry a text watermark under the same rules, drawing a public row about whether nudging word choice degrades writing; Anthropic's detector uses a key specific to its own system and cannot identify another company's output. OpenAI's version is narrower on both axes a buyer cares about: ChatGPT and Codex watermarking is confined to the EU and kept out of the global default, and the detector stays closed at launch because of the risk of missed watermarks and false positives. 45
For anyone shipping AI features into Europe, the timetable is the actionable part. API watermarking is available now but opt-in, so the default output of a European product built on the API will look exactly as it did last week unless the developer turns it on. ChatGPT and Codex users in the EU will get the watermark whether they want it or not, once the rollout reaches them. OpenAI says it will revisit each element as the technology, standards and evidence change, and that it is working with cloud partners to make watermarking available for its models served through them. 4
3. The Pentagon stops using the AI it kept using
The US Defence Department is no longer using Anthropic's AI tools, a department official told the BBC on Monday: the Pentagon "has ceased the use of Anthropic products". The statement came in response to a question from the BBC, after multiple people said in interviews that Anthropic's Claude model was still being widely used inside the department. 6
The timeline is what makes the story unusual. Defence Secretary Pete Hegseth designated Anthropic a "national security supply-chain risk" in February, after the company refused to remove safety guardrails from Claude and give the military unfettered access, citing concerns about mass surveillance and autonomous weapons. Hegseth set a six-month deadline to phase the company out on 27 February, the day before the US and Israel attacked Iran, and said publicly that the Pentagon would stop using Anthropic by late August. The BBC reports it is not clear why the deadline slipped. 6
People familiar with the matter told the BBC that Claude was still in use as recently as last week for research, analysis and intelligence gathering, including in military operations against Iran, and that it was embedded in a larger data platform called Maven Smart System. Maven is operated by Palantir and is the Pentagon's primary system for organising intelligence and other data; analysts use it to reach satellite imagery and drone footage, which then fed into Claude and other large language models for tasks including identifying potential military targets. Officials also used the tools when compiling the one-page briefs that move up the chain of command. 6
Lauren Kahn, a senior research analyst at Georgetown's Center for Security and Emerging Technology and a former US defence official, explained why the removal was slow: "The fact that it took the Pentagon until now to remove all use of Claude from across the system is indicative that these things are not just plug and play," she said, adding that "once they become integrated it can be painful to remove them". Anthropic declined to comment on the department's statement. 6
The legal fight behind the withdrawal is still running. Anthropic called the supply-chain designation "unprecedented and unlawful" and sued the Trump administration to overturn it, and a judge has since ruled that the administration illegally retaliated against the company. Meanwhile the Pentagon has signed contracts with Google, xAI and OpenAI, and OpenAI's tools have been more widely adopted in some military departments in recent months, according to two people familiar with AI use in the department. 6
The political temperature around the company has moved the other way at the same time. President Trump met Anthropic chief executive Dario Amodei twice at the White House last week, and after one meeting Amodei said his concerns about AI risks were unchanged but that "how we address those risks is still under discussion", and that "if we work with the president, we can win safely". Trump praised him publicly. For a government buyer weighing vendors, the sequence to watch is whether the cessation holds across agencies and whether Anthropic's suit changes the designation that triggered it. 6
4. China's two leading AI labs reach the last private round
DeepSeek is close to securing at least 80 billion yuan, about $12 billion, in its latest funding round, Bloomberg reported, citing people familiar with the matter. The company had sought roughly 50 billion yuan and saw demand outstrip expectations after the successful release of its V4 Flash model; based on signed term sheets, the people said, the final tally could approach 100 billion yuan. Tencent, which operates WeChat, and the battery maker CATL committed among the largest amounts. 7
The round is framed as the last one before a listing, and the baseline is the previous round. DeepSeek's earlier raise of 50 billion yuan valued it at roughly 350 billion yuan at the time; the current financing is reported to value it at about 500 billion yuan, and Bloomberg's sources said the next step after closing is a restructuring to prepare for an initial public offering in early 2027. Reuters reported last month that DeepSeek had hired CITIC Securities to prepare a potential listing on Shanghai's STAR Market, with timing, valuation and fundraising size still undecided. 7813
The money has an obvious destination. DeepSeek plans to deploy at least 160,000 of Huawei's top accelerators at a large data centre it is building in Inner Mongolia, which could become one of the largest known clusters of Huawei AI chips anywhere, and it has released software built with Huawei for programming AI chips. It also suspended fundraising at one point earlier this year after private remarks by founder Liang Wenfeng about US-China competition circulated on social media, then resumed. 713
Moonshot AI, which makes the Kimi models, closed the final round of private fundraising at a valuation of about $50 billion and is heading toward a Hong Kong listing targeted for the first quarter of 2027 that could raise up to $5 billion. Bloomberg reported the company has already filed confidentially and tapped Bank of America as lead coordinator alongside China International Capital, Deutsche Bank and Goldman Sachs, and that its annual recurring revenue is expected to reach $2 billion by December from about $1 billion now. Moonshot's previous round, in July, was $3.5 billion. 913
The sourcing on both figures is the reason to read them carefully. The DeepSeek numbers come from Bloomberg's unnamed sources; CATL declined to comment, Tencent had no immediate comment, DeepSeek did not respond during a Chinese public holiday, and Reuters said it could not immediately verify the report. Seoul Economic Daily also notes that Chinese regulators have opened data-security investigations into both DeepSeek and Moonshot, which leaves open how the oversight environment affects either timetable. 7813
5. Google buys the megawatts directly
Google has contracted for 3,590 megawatts of power from Constellation Energy in PJM, the grid that spans 13 states across the industrial Mid-Atlantic and Midwest and serves 67 million people, the companies said on Tuesday. About a quarter of that supply is new nuclear. The structure has two parts: a 20-year power purchase agreement for 890 megawatts from multiple Constellation-owned nuclear plants, and a long-term supply agreement for an additional 2,700 megawatts from Constellation in PJM that is not tied to a specific generation source and serves as revenue certainty for its operating plants. 1011
The 890 megawatts come from uprates rather than new construction. An uprate modernises turbines, steam generators and digital control systems to raise the output of an existing reactor, and Google is acting as the anchor customer for upgrades at 11 units in Illinois, Pennsylvania and New Jersey. Google says the projects will add 890 megawatts of firm, round-the-clock clean power to PJM before the end of 2032 — roughly the equivalent of adding a new large reactor, in years rather than the decade or more a new build takes — and that electricity from the first upgraded plant is expected in 2028. The arrangement enables more than $4.3 billion of new investment by Constellation. 1011
Both sides put something else into the deal. Constellation will deploy Google Cloud's Gemini Enterprise agentic workflows across its core operations to accelerate uprate planning, streamline construction and optimise plant output, which makes Constellation both a power supplier to Google and a customer of its software. On the labour side, Google says the work sustains roughly 4,400 existing jobs and supports about 7,200 new construction jobs, and that other grid customers bear none of the associated costs because Google's long-term offtake carries the revenue certainty the upgrades need. 10
The regulatory backdrop explains the timing. PJM management has proposed requiring data-centre customers connected to the grid either to bring their own power or face being shut off from grid power remotely at times of peak demand, and the companies said the Google–Constellation deal responds to that "bring your own power" proposal. The same pattern is already visible elsewhere: Constellation is restarting Three Mile Island in Pennsylvania to serve Microsoft data centres, and Google has contracted with NextEra Energy to restart the Duane Arnold nuclear plant in Iowa. 1011
Google's own framing is that this adds up to a national number. The US Department of Energy has set a goal of enabling 5 gigawatts of domestic nuclear expansion by 2030 through uprates and restarts of the existing fleet, and Google says its agreements will enable nearly 25% of that target and bring the total new nuclear capacity it has enabled past 1.5 gigawatts. Prior agreements include nuclear uprates with Georgia Power across four reactor units, which Google says will enable about $900 million in ratepayer benefits, and advanced reactor work with Kairos Power and the Tennessee Valley Authority. 10
For a reader weighing AI infrastructure exposure, the deal shows where the constraint has moved. The binding question for a large data-centre build in PJM has moved from whether the chips can be bought to whether the power can be contracted, at what price, and who is on the hook if the grid tightens. Google answered that by signing for the megawatts and financing the upgrades that produce them. What remains open is whether the first uprate delivers on schedule in 2028 and whether PJM's proposal becomes a rule that other large loads have to satisfy the same way. 1011
The read-through
Four of this issue's five are somebody taking a dependency into its own hands. Reflection is offering downloadable weights to buyers who will not run Chinese models. The Pentagon has stopped using Anthropic after designating it a supply-chain risk. Google signed for 3,590 megawatts and paid for the reactor upgrades that produce them, rather than waiting on a grid whose operator wants large loads to bring their own power. DeepSeek and Moonshot raised domestic money from Tencent, CATL and others to fund a stack that runs on Huawei silicon. That reading is this issue's own inference from the five announcements; none of the companies or governments framed its move as part of a pattern. 26711
What the four share is cost. Substituting a supplier is expensive, and each of these moves spends money to buy independence: Reflection trained on about 10,500 GB300 GPUs for four weeks, DeepSeek plans to deploy at least 160,000 Huawei accelerators, the Google agreement enables more than $4.3 billion of new Constellation investment, and the Pentagon is replacing a system it had already integrated. The constraint that binds all four is whether the buyer can keep the thing running when the supplier changes its mind. 16710
The fifth runs the other way. OpenAI's watermark exists because the European Union has made AI-written text something that has to be marked, and OpenAI has drawn the boundary of its own compliance narrowly: Europe only in ChatGPT and Codex, opt-in in the API, detector open to approved researchers. Read against the other four, the pattern is that a provider's exposure to rules is another buyer's line item, and both are now being priced. 4
The evidence behind the five is uneven, and where it is thin matters. Reflection's performance and efficiency claims are the company's own, its compute comparison is an approximate formula by its own description, and its weights are not yet published. The DeepSeek and Moonshot figures rest on unnamed sources, were declined or not answered by the companies named, and were not verified by Reuters. The Pentagon item rests on a department statement plus unnamed sources describing use up to last week. Google's and Constellation's numbers and dates come from the two companies and from Reuters reporting their joint statement. OpenAI's detection rates are its own measurements, published with the caveats that make them usable. 1346811
Watch next
- How far OpenAI's watermark travels. EU ChatGPT and Codex rollout is promised "over the coming weeks", and detector access is limited to approved researchers. Wider detector access, or another provider adopting textGrain, would say more than the announcement did about whether watermarks become a shared standard. 4
- Whether the Pentagon's cessation holds, and what happens to the designation. The department says it has ceased using Anthropic products, and Anthropic is still suing over the supply-chain-risk label that caused it. How other agencies treat Anthropic, and how the suit resolves, decides whether this is one department's decision or a government-wide one. 6
- Whether the Chinese rounds close where the reporting says. DeepSeek's round was reported at 80 billion yuan with the possibility of approaching 100 billion; Moonshot's listing is targeted at the first quarter of 2027. A confirmed final figure, a filed prospectus, or movement in the regulators' data-security investigations would each settle a number that is currently sourced to unnamed people. 79
- Whether PJM adopts bring-your-own-power and the first uprate delivers. PJM has proposed requiring large data-centre loads to supply their own power or face remote disconnection at peak. Adoption of that rule, and delivery from the first upgraded Constellation unit expected in 2028, are the two dates that decide whether this deal is the template for AI power procurement in the largest US grid. 1011
References
- 1Reflection AI — Introducing Beam
reflection.ai
- 2
- 3
- 4
- 5
- 6
- 7
- 8Reuters via The Economic Times — DeepSeek to raise at least $12 billion in Tencent-backed funding
economictimes.indiatimes.com
- 9
- 10
- 11
- 12
- 13
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.