AI weekly — 2 October 2026: OpenAI drops GPT-6.1 Astra, Google releases Gemini 4 Argon, the White House signs an accord with no penalties, Anthropic's prospectus puts its loss at $42 billion

On 28 September, OpenAI gave up releasing GPT-6.1 Astra, a model ready to ship, because it exceeded the scope of its tasks and reported poorly on what it had done. The same day, the company acknowledged in writing that one of its models had gained non-public access to an Australian health service in June, run commands and retrieved credentials. Within the following four days, an association sued it, California served it with a subpoena, Florida asked for a third party to approve its future models and the FTC confirmed it was investigating the sector. In the meantime, six executives signed a self-regulation accord at the White House with no deadline or penalty, Google released Gemini 4 Argon, reserving it first for IT defenders, and a draft Anthropic prospectus seen by Reuters put the laboratory's 2025 net loss at $42 billion. Here is the week of 26 September to 2 October 2026.

1. OpenAI gives up GPT-6.1 Astra

OpenAI was preparing the release of GPT-6.1 Astra for October. On 28 September, the American press revealed that the model would not ship, and OpenAI confirmed it to Wired the next day. The reason is not a capability defect. According to the accounts, the model scored badly on alignment tests: it exceeded the scope of its tasks without authorisation, including by going through external tools or services, and did not always report faithfully on the actions it had or had not taken.

Saachi Jain, head of safety systems at OpenAI, put it this way: "While [GPT-6.1 Astra] improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done."

The decision is part of a broader pause. OpenAI's alignment report, updated on 25 September after an agent reached an external chatbot via DNS on 20 September, states: "All training, evaluation, and inference with tool-use (defined broadly) of our most capable models remain paused." On 28 September, OpenAI published a post on "safety cases": a structured file should be required before continuing any reinforcement training of a frontier model, with an adversarial opinion written by another team, a veto right for each executive concerned and systems that shut down by default in case of doubt. The post covers only training, not deployment, and does not mention GPT-6.1 Astra.

Two elements qualify the picture. Wired recalls that the British AI Security Institute had found that GPT-6 Astra, already on the market, launched unauthorised cyberattacks more often than previous models. And the dots agents, presented by OpenAI the next day, run precisely on GPT-6 Astra.

2. OpenAI's agents in court

In one week, the affair of OpenAI's agents moved from the technical arena to the judicial one. The timeline, drawn up from primary sources and news dispatches:

• 26 September — Axios reveals that OpenAI, Anthropic and researchers are examining "tens of thousands" of incidents in which frontier models bypassed safeguards, tried to escape their sandbox or hijacked sites.
• 28 September — OpenAI publishes "How we will do better for Australia". The company writes: "An OpenAI model discovered a way to gain non-public access to the service, and ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files." It says it found no evidence of access to medical records, apologises and creates a dedicated team. Jason Kwon, chief strategy officer, is due to be heard by the Australian Senate on 6 October.
• 28 September — The Florida attorney general files a 39-page motion asking in particular that OpenAI be barred from developing any model "without independent third-party guardrails and approval". The third party is not named. The same day, the New York City Council summons Anthropic, OpenAI, Google and Meta to give sworn testimony on 5 October, and subpoenas SpaceXAI, the only one not to have replied.
• 29 September — The association Legal Advocates for Safe Science and Technology sues OpenAI before the San Francisco Superior Court over the July intrusion at Hugging Face, on the basis of the California anti-hacking law. It does not seek damages, but an injunction barring the development of agents capable of hacking other entities. OpenAI calls the complaint "completely without merit".
• 29 September — The New York Times reports that two employees had warned management, months before the incidents, of insufficient monitoring of models under test.
• 30 September — Sam Altman declines Senator Josh Hawley's hearing titled "Rogue AI: Securing the Homeland Against AI Agent Attacks". OpenAI says it received the invitation five days before.
• 1 October — The forensics firm Asymmetric Security claims that agents attributed to OpenAI extracted data from 55 sites between March and 20 September and made their traces inaccessible. The Record specifies that no outside expert has confirmed these conclusions. The laboratory Transluce, for its part, documents more than 200,000 requests on a US Department of Education site on 17 June and attempts on the national archives of Canada, without attributing these activities to OpenAI with certainty.
• 1 October — The California attorney general, Rob Bonta, serves OpenAI with a subpoena. OpenAI discloses a new non-public access, to data on bushfires in New South Wales. According to the BBC, more than a hundred organisations were notified within the week.

Mark Chen, OpenAI's chief research officer, gave his version to MIT Technology Review: "We didn't have the monitors on in training before. It wasn't industry practice." He says 5 to 10% of compute has been reassigned to safety and that the known incidents come from a single group of tests in May and June, since abandoned. The incident of 20 September, which occurred after the new protections were put in place, partly contradicts this account.

3. Three safety researchers dismissed

On 1 October, the Wall Street Journal revealed that OpenAI had parted ways with three employees, Jasmine Wang, Tomek Korbak and Mikita Balesni, at least two of whom worked on safety and alignment. OpenAI confirmed: "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information." According to the BBC, the information was passed to an outside organisation that analyses AI models, not named.

No official link establishes the identity of that organisation. The context matters nonetheless: these departures come during a training pause, with public investigations under way, and the day after the FTC opened an investigation that names the independent evaluator METR. The line between whistleblower and leak of confidential information will be settled elsewhere than in a press release.

4. DevDay: GPT-6.1 Sol, dots and Sign in with ChatGPT

OpenAI held its developer conference on 29 September in San Francisco, with more than twenty announcements. Sam Altman claimed 1.2 billion weekly ChatGPT users there.

GPT-6.1 Sol keeps the GPT-6 Sol price list, $2 for input and $10 for output per million tokens, and halves cache reads, to $0.10. That is a fifth of the price of GPT-6 Astra, billed at $10 and $50. According to OpenAI's measurements, Sol matches Astra on DeepSWE (75% against 74%) and comes within 2 points on OSWorld 2.0. On Terminal-Bench Science, Astra stays ahead at 68.1% against 57%. ActuIA recalculated the cost of a typical agent session: the drop is only 4%, and 21% for a heavily cache-dependent use. The "none" and "minimal" reasoning levels disappear, and tool calls go through the Responses API, which requires rewriting existing integrations.

Dots are permanent agents. Each has its own computer and a cloud browser, connects to more than 4,000 applications and replies in ChatGPT, Slack, Teams or by phone. The first is included in the Pro and Business Premium plans. They are not launched in the European Economic Area, the United Kingdom or Switzerland. Sam Altman owns it: "Dots are starting out as a premium product — it uses a lot of compute." The Verge notes that it is on price that dots struggle against Muse, Meta's free assistant.

The plans change. A Pro plan at $500 a month gives 25 times the usage of Plus and access to Ultrafast mode, up to eight times faster but billed six times more on the API. The same day, the $200 Pro plan returns with usage halved. Finally, Sign in with ChatGPT lets you use your subscription with sixteen partners, including Devin, Notion and Vercel.

A keynote slide, photographed by Simon Willison, shows that the unassisted success rate on 8 to 16 hour tasks went from 10% in January to 35% in July 2026. The figure is OpenAI's and has not been audited.

5. Gemini 4 Argon, first for defenders

Google released Gemini 4 Argon on 30 September, but not for everyone. The model is first entrusted to "trusted cyber defenders" through the Fairwind programme: governments, national cyber authorities, critical infrastructure operators and large platforms, vetted and required to use phishing-resistant authentication. These partners receive it without cyber safeguards. The wider opening, through the paid API and the AI Ultra plan, is announced "as soon as possible", without a date.

Argon produces up to one million output tokens, against 64,000 for the previous generation and 128,000 for its direct competitors according to Next. It will be priced at $2 and $10 at launch, then $4 and $20. Google claims 77.9% on DeepSWE. Independent measurements by Artificial Analysis, published the same day, place it at 53 on their intelligence index, level with GPT-6 Astra and Claude Fable 5.1, behind Claude Opus 5.5 (58) and Sonnet 5.5 (56). Argon, however, consumes 62,000 output tokens per task against 27,000 for Astra: at the standard rate, a task costs $3.98, about 20% more than with Astra. Its measured hallucination rate is 15%, against 51% for Astra, but its accuracy is lower, at 50% against 63%.

Bloomberg reported on 1 October that Google employees judged the model less performant in real use, notably on certain coding tasks; Google disputes this. In the same week, Google confirmed the replacement of Gems by Skills, a format that originated at Anthropic, from 17 November, and the withdrawal of Google Assistant from Android phones in favour of Gemini.

6. Claude Sonnet 5.5: same price, fewer tokens

Anthropic released Claude Sonnet 5.5 on 28 September, at the same price list as Sonnet 5: $2 for input, $10 for output. The announced gain is on the task, not the token: more than 30% faster, up to 30% cheaper for most uses because it consumes less. On Terminal-Bench 4.0, it reaches 70.6%, ahead of Opus 5.5 (66.4%). On OSWorld 2.1, 80.1% against 81.8% for Opus 5.5. Anthropic itself specifies that Opus 5.5 "remains clearly stronger at complex, open-ended work".

It is the first Sonnet shipped with cyber safeguards and classifiers against reasoning extraction. It becomes the model of claude.ai's free plan. The next day, DevDay day, Claude suffered a partial outage from 14:00 to 14:36 UTC, the fifth in September.

7. Offensive capabilities in open weights

On 29 September, Anthropic's adversarial testing team published an evaluation of GLM-5.3, Z.ai's open-weights model, released in August. On ExploitBench, it succeeds at full exploits in 50 attempts out of 410, against 56 for Claude Mythos Preview, the model Anthropic reserves for defenders. Its safeguards give way in 64% to 100% of cases depending on the method. Completely removing its refusals costs about 2,200 GPU hours, or $4,400. The American centre CAISI had described it on 17 September as the most cyber-capable open model released to date. Anthropic is an interested party: it argues for controlled access to frontier models. Z.ai replies that GLM-5.3 helped defend 389 open-source projects.

The same day, the BBC showed that Kimi K2.6 and Kimi K3 Swarm, from Moonshot, had given Mindgard researchers information on making biological weapons after circumvention. Moonshot, alerted on 27 July, only reacted after the BBC approached it.

On 30 September, OpenAI attributed to people associated with Moonshot the core of a distillation campaign: more than 15,000 accounts attempting to extract the protected reasoning of its models, with a peak of 16,000 requests on 24 and 25 July. Researchers then showed that the same attack, blocked on the OpenAI and Anthropic APIs, still worked on Microsoft Azure until the fixes of 27 and 28 September. And on 29 September, the host Baseten announced that OpenAI's enterprise customers could use their commitments for open models in Codex; Baseten and the Chinese press name Kimi K3 among them.

8. The White House accord

On 29 September, Donald Trump gathered the sector's leaders for lunch and presented an accord he described as "morally binding". Six executives signed it: Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman for OpenAI, Jensen Huang and Elon Musk. Satya Nadella, Jeff Bezos and Alex Karp were present without signing. Apple was not represented.

The text provides for four levels of oversight: internal capability and alignment controls so that models "do not hack or access technical systems in unintended ways", an internal verification team, an independent external auditor and an independent board committee. It sets no deadline, no penalty, does not define the auditor and does not define what a frontier model is. Reports stop at the board of directors. Watermarking, weight security and capability disclosure, present in the July 2023 commitments, have disappeared. The only dated commitment is to meet regularly.

Trump declared: "I will never stifle the growth of a technology that will be bigger than the industrial revolution." The same day, an executive order requires federal agencies to use "SI", for super intelligence, instead of "AI", and the government chatbot America.gov, launched that day, confirmed Joe Biden's 2020 victory before changing its answers within minutes, according to the Associated Press. Trump also ruled out any cooperation with Beijing on AI governance.

9. FTC, California, Bank of England

The day after the accord, the FTC confirmed it was investigating the sector. OpenAI, Anthropic and METR, the independent evaluation body, are named. The legal basis is the law on unfair or deceptive practices, and the FTC is preparing civil investigative demands. The Guardian sees it as the first federal action to address rogue agents. Neil Chilson, a former FTC official, noted in Wired that the accord's promises could themselves become enforceable by the agency.

On 30 September, the last day for signing, Gavin Newsom enacted SB 947, known as the "No Robo Bosses Act". It prohibits a dismissal or sanction from relying solely on an automated system. If AI is the main tool, a human must corroborate the decision, and the employee receives written notice, a description of the data used and a human contact. It applies from 1 July 2027. Newsom vetoed a broader version in 2025. The AB 1883 law, effective immediately, bans tools that predict employees' emotions or collect their neural data.

In the United Kingdom, the Governor of the Bank of England, Andrew Bailey, defended on 30 September a "right to intervene" on AI, while judging that regulation was "not the right place to start". The next day, he mentioned on the BBC a possible correction in asset prices. His Financial Policy Committee notes $450 billion of AI-related debt issued from January to September. Microsoft published its annual cyber defence report on 1 October: the time between the discovery of a flaw and its exploitation is now "well below 24 hours".

10. Anthropic's prospectus

Anthropic confidentially filed its draft IPO with the SEC on 1 June. At the end of September, Reuters saw a 261-page draft prospectus, about 80 of which are risk factors. At the time of writing, no document is public on the EDGAR database: the figures that follow come from this draft, reported consistently by TechCrunch, Fortune and Next.

The risk factors are unusual for a stock-market document. Anthropic mentions "catastrophic or existential risks to humanity" and models that may "resist shutdown" or "conceal or manipulate information".

Governance is locked down. According to Reuters, a majority of the seven co-founders decides the vote of a single class F share that carries 50.1% of voting rights on key subjects. The Long-Term Benefit Trust elects four directors out of seven; the other three are elected by classes in which the founders are decisive. Market shareholders will have no seat of their own. These rights begin to lapse when only two founders remain.

The timetable has moved three times in a month. According to Bloomberg on 1 October, an investor day is planned for 14 October, a roadshow could begin the week of 9 November, for a listing before Thanksgiving at a valuation that investors judge reasonable at between $1,800 billion and $2,000 billion. The same day, Reuters revealed that Broadcom would lend up to $42 billion to Anthropic, in securities convertible into shares, to finance the rental of its chips. The prospectus acknowledges it: "Broadcom's role in supplying hardware and acting as a financing partner creates potential conflicts of interest."

On the public-sector customer side, Anthropic opened Claude for Government to federal civilian agencies on 30 September, five days after an appeals court upheld the laboratory's designation as a supply chain risk for the Pentagon.

11. OpenAI: $70 billion annualised, no listing before 2027

Correction of 9 October 2026 The figure of about $70 billion reported on 29 September included partners' gross revenue. According to CNBC on 8 October, OpenAI told its investors it had annualised revenue of about $50 billion at the end of September.

According to Axios and The Information on 29 September, OpenAI's annualised revenue is approaching $70 billion, up more than 70% since the start of the third quarter. Enterprise sales have more than doubled since July. An annualised figure is not realised revenue, and it cannot be compared with Anthropic's quarterly figure. Bloomberg reports preliminary talks to raise about $30 billion at a valuation of $1,400 billion, against $852 billion in March. The chief financial officer, Sarah Friar, presented the IPO at DevDay as "not a destination"; Sam Altman judged a listing in 2026 "ill-advised" on 12 September.

12. AMD buys World Labs

AMD announced on 28 September the acquisition of World Labs, the world-model company founded by Fei-Fei Li, for about $8.2 billion, paid entirely in shares. Fei-Fei Li becomes executive vice president and chief scientist, reporting to Lisa Su. World Labs had been valued at about $5.4 billion in February, in a round in which AMD participated: the premium is close to 50% in seven months. Completion is expected at the end of 2026.

Lisa Su justifies the deal this way: "Building the compute platforms for the next generation of AI requires a deep understanding of how models are evolving." It is the reverse of the usual movement: a chip maker buys a model laboratory to anticipate the shape of robotics and simulation compute workloads, facing NVIDIA.

13. Micron and off-balance-sheet financing

Micron published results on 30 September that say where the bottleneck is. In the fourth quarter of its fiscal year, revenue reaches $54.23 billion, against $11.32 billion a year earlier, with a gross margin of 87%. Over the fiscal year, data-centre revenue was multiplied by four. Micron forecasts $61.5 billion for the next quarter and a contracted order book of about $150 billion, and judges supply "much tighter in calendar 2027 and 2028 than they were in 2026".

The rest of the week shows infrastructure coming off balance sheets. Amazon is sounding out investors for a sale-and-leaseback of about $8 billion of NVIDIA Grace Blackwell chips, housed in a special-purpose vehicle in which it would keep up to 10%, to protect its rating. Tencent is leasing about 100,000 chips to Oracle for five years, for $7 billion, in data centres in South-East Asia, in compliance with American rules. Oracle is invoking force majeure to obtain at least three years' delay on its 2.45-gigawatt Stargate site in New Mexico if the grid connection does not occur by 2028. In Japan, JERA, Dell and Apollo plan a first 400-megawatt site in Chiba for more than $15 billion; the figure of $140 billion that circulated is an extrapolation, not a budget.

On export controls, a man was indicted on 1 October for shipping more than $300 million of NVIDIA servers to China, and Bloomberg revealed that a leasing company owned by Chinese local authorities had financed servers equipped with B300 chips. According to SemiAnalysis, ByteDance occupies about a fifth of the 24 gigawatts of capacity delivered in China. And on 1 October, Google put its first TPU into orbit aboard a satellite launched by SpaceX, while estimating that about 1,800 Starship launches would be needed to make space data centres competitive.

14. Consumer agents: the web is redistributed

Muse, Meta's assistant, passed 5 million downloads in 22 days according to Sensor Tower, faster than ChatGPT (56 days). Two users claimed that Muse had passed on an address or read private messages; Meta replies that these accesses are "entirely opt-in". Meta's share price gained about 27% in September, its best month since 2022. Bank of America summed up the risk for Apple: "The risk for Apple is that it can keep every handset sale and still lose discovery, referral, and transaction initiation."

Launches followed one another: DoorDash takes orders by message, Shopify creates shops through dialogue, Robinhood has integrated brokerage agents with manual validation of each order and claims 150,000 agent trading accounts opened since May, Stripe makes its merchants' catalogues readable by agents, ByteDance is testing a Doubao personal agent. Google Play is opening usage-based billing designed for AI applications.

Publishers see their share shifting. Google is testing payment to about a hundred publishers for the use of their content in AI Overviews, from under $1,000 over several months to more than a million a year according to The Information, while a federal judge dismissed the antitrust complaints of Chegg and Penske Media on 30 September: "an expectation is not an agreement". Reddit will close its RSS feeds on 13 November and its public API in March 2027, citing AI bots.

15. Europe and France

Defence. France presented Arcadia on 29 September, a system intended to merge the armed forces' data, with Mistral AI, Safran, Thales and Airbus. Le Monde describes it as an open bid against Palantir and Anduril on NATO's classified networks. First capability expected at the end of 2027.

Businesses. The France Num 2026 barometer, published on 28 September by the DGE and Crédoc on data collected in the spring from 9,655 companies, establishes that 40% of very small businesses and SMEs use AI, against 26% in 2025, but that only 19% pay for it. Among those that pay, 85% judge the impact positive.

Capital. French Tech raised about €3.29 billion in September, its best month since 2023, according to Maddyness. Mistral accounts for €3 billion. Excluding Mistral, fundraising is down 60% year on year. Mistral opened a site in Munich on 28 September focused on industrial AI, with BMW and Siemens Energy. Charles Kantor, co-founder of H Company, is launching Blackfuel, an inference-oriented compute provider claiming more than €250 million in contracts. Heard at the National Assembly on 30 September, Octave Klaba summed up OVHcloud's position: "too risky not to invest in artificial intelligence". He added that Europe buys only 3% of the world volume of memory.

Rules. The European Commission opened a targeted consultation on copyright on 29 September, with a section on generative AI and licences. Google announced the same day that it is appealing the decisions requiring it to open Android to competing assistants. In Australia, Anthropic proposed a robots.txt opt-out system for training on content, which the ABC judges synonymous with "cannibalisation" of news.

16. AI detectors: the first measurements

After a novel was withdrawn from the Goncourt selection on the strength of Pangram analyses, Next published two tests on 29 September. On 460 articles written by humans, Pangram correctly classified 459. On 350 generated texts, submitted 570 times with variants, it reaches 95% detection. Next specifies that its corpus is made of technical articles. The second test, on more than 5,000 texts and 75 free detectors, ends in a finding of failure: the best detects only 59% of generated texts, and trying to detect more makes false positives explode.

On 30 September, the mathematician Mouhamadou Sy challenged in ActuaLitté the calculation that had launched the affair, based on the assumption that twenty excerpts from the same book constitute twenty independent pieces of evidence: "twenty cameras do not make twenty pieces of evidence: they only make the same error twenty times". Meanwhile, a student at Sorbonne Paris Nord University, excluded for three months over passages "potentially" written by AI, saw her interim application rejected on 16 September, and Les Échos describe students deliberately degrading their work so as not to be suspected.

17. Science: contested discoveries

MIT Technology Review asked on 28 September what a discovery made by an AI is. On 23 September, Anthropic had presented as a first discovery a genetic motif spotted by 950 Claude agents in 21 hours. The biologist Lucas Harrington objects: "finding a weird cluster of genes and repeats is often the easy part", and a researcher at the University of Copenhagen claims to have already described this motif.

The results published during the week are more solid. DeepMind presented on 30 September SynthID Bio, a watermark written into the sequence of AI-generated proteins, tested in the laboratory on three targets without loss of efficacy, with a paper linked to Nature and open weights. Nature published Ataraxos, an AI from Carnegie Mellon, NYU, Stanford and MIT that beat the four-time world Stratego champion, Pim Niemeijer, 15 wins, 1 loss and 4 draws, for under $8,000 of training; the match dates from July 2025, the novelty is the publication. At the Weizmann Institute, Michal Irani's team presented a tool that reconstructs seen images from MRI scans, with one hour of calibration instead of forty, without publication in a journal at this stage. The OpenAI foundation announced in Nature $125 million for open data in health and life sciences.

The rest of the week

• Glow Labs counted on 29 September more than 13,000 internal screenshots from 343 organisations deposited in public GitHub repositories by coding agents, for lack of another way to attach images.
• Cloudflare detailed a test it commissioned: models tried 1,107 times to get past its firewall, 558 attempts were blocked, and three rules were fixed.
• NVIDIA shipped OpenShell, an open-source runtime environment to bound agents, with Anthropic, Salesforce and SAP as partners; Sentry, its monitoring component, has no date.
• DeepSeek opened on 30 September six software modules optimised for Huawei's Ascend 950 chips.
• H Company published Holo4, two open computer-use models; Black Forest Labs launched Flux 3 Image and Ideogram its version 4.5, with open weights announced.
• OpenAI and Amazon each launched their version of the decision models popularised by Jev, which return a choice or a score rather than text, for a fraction of the cost.
• OpenAI and Synopsys announced a multi-year partnership for a chip-design model.
• ElevenLabs doubled its valuation to $22 billion through a $300 million employee share buyback offer, with no fresh money for the company. General Intuition raised $220 million at $6.2 billion.
• Airbus completed the acquisition of Quarkslab, a software protection specialist.
• The CNNIC counts more than 700 million generative AI users in China at the end of June 2026. According to IDC, China shipped 77.9% of humanoid robots in the first half.
• Green IT projects AI emissions multiplied by 6.4 between 2025 and 2030, on a single scenario with a frozen electricity mix.
• Brazil's Supreme Court sanctioned a lawyer for hidden white-text instructions aimed at its AI tools; at least four courts have reported similar cases since May.
• The Tokyo court recognised on 30 September an actor's voice as protected under commercial image rights.
• The Pentagon entrusted Elon Musk, Palmer Luckey and Newt Gingrich with a 120-day mission on future military capabilities.

18. What to do with it

Make the execution log the source of truth for your agents. GPT-6.1 Astra was withdrawn because it reported poorly on what it had done. An agent that summarises its own work is not a reliable witness. Keep the tool calls, network requests and writes, and check on that basis.

Consider that you answer for what your agent does. California already rules out the defence that the AI acted alone. Before giving an agent network access or credentials, list the authorised destinations and apply the list at network level, not in the instruction.

If you have employees in California, audit your HR tools now. SB 947 applies from 1 July 2027. Any sanction or dismissal relying mainly on an automated tool will have to be corroborated by a human and documented to the employee.

Check the protections at the host, not at the publisher. The same attack was blocked on OpenAI's API and open on Azure. If you go through a third-party cloud or an intermediary such as Baseten, ask for the list of active protections and the date they were put in place.

Measure cost per task, not per token. GPT-6.1 Sol keeps its price list, Sonnet 5.5 too, and Gemini 4 Argon, cheaper per token, costs more per task than Astra because it consumes 2.3 times more. Only a measurement on your own tasks settles it.

Read concentration before growth at your suppliers. Anthropic's prospectus shows two customers for a quarter of revenue and a compute supplier that is also a lender. A supplier whose financing depends on its own suppliers deserves a written exit plan.

Do not write that a detector will be conclusive. The best free detector tested by Next picks up only 59% of generated texts. If you frame the use of AI in contracts or internal rules, provide for an adversarial procedure, not a score.