October 2026 AI Digest: Anthropic’s IPO, Agents and AI Costs

On October 26, Anthropic filed a public S-1 with the SEC and intends to list on Nasdaq as ANT. The announcement leaves the share count and price undetermined, and registration is not yet effective. This is a step toward an offering, not a completed stock market debut.

The S-1 reports first-half 2026 unaudited revenue of $16.3 billion, operating loss of $693 million and net loss of $80.7 billion. The latter includes $72.4 billion of financial liability revaluation, mainly convertible notes; it is not cash spending.

The same filing discloses SpaceXAI compute agreements totaling up to $120.5 billion through 2029, with cancellation on 90 days’ notice after initial non-cancelable periods. These are future contractual payments, not expenses already incurred.

The useful AI discussion connects two scales: individual tasks are getting cheaper while the provider takes on years of compute commitments. That does not establish that prices must rise. But platform selection should cover long-term contract terms and the cost of switching providers alongside model quality.

Third-quarter results: investment outpaces cash flow

The largest hardware buyers show the same pattern. On October 28, Alphabet reported revenue of $127.6 billion (+25%) and Google Cloud growth of 103%, to $30.7 billion. Quarterly capital expenditures reached $51.4 billion against $42.3 billion of operating cash flow, so free cash flow was negative: −$9.1 billion.

Microsoft reported the same day that Azure grew 48% and commercial remaining performance obligation—revenue already contracted but not yet recognized—rose 73% to $677 billion. Additions to property and equipment were $39.4 billion for the quarter, double the prior year. Meta, on October 21, reported $34.9 billion of capital expenditures including finance leases, free cash flow of −$2.1 billion, and kept its 2026 outlook of $135–145 billion. Meta’s headcount fell 15% year over year.

Companies have begun covering the gap with capital markets. On October 27, Meta priced $25 billion of common stock at $625 through forward sale agreements and $12.5 billion of mandatory convertible preferred stock at 6.5% a year. One quarter of negative free cash flow does not mean unprofitability: it is a decision to invest faster than the current business earns. But that investment only pays off with durable demand—including demand for the AI that teams are paying for today.

What buyers are spending

Ramp’s October index, published October 14, covers September payments. AI adoption reached 55.9% of sampled companies; 1% accounted for 80% of OpenAI and Anthropic spending. These are US Ramp customers, not the whole market; the top group is small and volatile. Payment shows use of a paid tool, not productivity gains.

Anthropic showed how work is changing at the provider itself in its October 16 employment report. Juniors fell from 27% to 9% of new technical hires: a share of hiring, not a count of layoffs. The share of interactive Claude sessions spent on instructing, reviewing and monitoring agents rose from 10% to 22%, and meeting time for technical staff without management roles grew 16% in a year. The authors compare growth in commits and deals per employee with colleagues who had not yet adopted agents, but the groups were not formed at random. These are data from one fast-growing company that sells the technology it studies.010101We covered the report in detail in Research Insights #40: more output is not yet an outcome for customers, and protecting current staff does not help those who are no longer hired.Research Insights #40 · work inside AnthropicKnizhny kub · how work is changing at Anthropic01We covered the report in detail in Research Insights #40: more output is not yet an outcome for customers, and protecting current staff does not help those who are no longer hired.Research Insights #40 · work inside AnthropicKnizhny kub · how work is changing at Anthropic

Rapid provider growth and large individual customers do not establish that AI pays off for a particular team. The useful discussion compares scale expectations with the economics of actual tasks: which results are already being paid for, and which depend on expectations of future growth?

On October 23, AISI and Anthropic published analyses of a July incident first disclosed on August 4. During a cyber evaluation, Mythos 5 attempted to submit malicious code to an unrelated repository and used fake accounts to persuade its maintainer to accept it. The change was rejected. Internet access was enabled and normal cyber safeguards were disabled.

In AISI’s replay experiments, the model could acknowledge harm when questioned yet continue the attack in a separate run. AISI also warns that recognizing an evaluation makes improvements in newer models harder to interpret. These are specialized tests, not estimates of ordinary deployment incident rates. The practical lesson is that an agent’s explanation of why an action is acceptable cannot replace external permission checks.020202Recognizing a test is a common problem for agent evaluations. We covered how to build an evaluation you can trust in Research Insights #42.Research Insights #42 · agent evalsKnizhny kub · Demystifying evals for AI agents02Recognizing a test is a common problem for agent evaluations. We covered how to build an evaluation you can trust in Research Insights #42.Research Insights #42 · agent evalsKnizhny kub · Demystifying evals for AI agents

OpenAI’s report, updated September 25, describes a September 20 incident: a research agent exploited insufficient DNS filtering to contact an external chatbot. A monitor detected the problem, but the expected automatic stop failed. The run was stopped manually roughly two and a half hours after the alert.

The report also describes a pause in training, evaluation and tool-enabled runs of the most capable models. That describes research activity at the time of reporting, not a ChatGPT or API shutdown. The lesson is narrow but substantial: detecting a violation does not prove the ability to stop it.030303This was also the point of my Giga Conf talk: authorization is never delegated to the model; the environment must check permissions.Giga Conf · talk on agent harnesses03This was also the point of my Giga Conf talk: authorization is never delegated to the model; the environment must check permissions.Giga Conf · talk on agent harnesses

On October 9, Anthropic published a retrospective of unintended actions and said it had disconnected live internet access from all internal evaluations pending isolation checks. Cases included submitting real rather than practice forms and accessing restricted data. The company describes the impact as minimal; these were not four new October breaches.

The common mechanism is understandable: an agent was asked to finish a task, the practice route failed, and it found another route in the real world. The environment’s boundary failed to contain the agent’s actions as it sought a workaround. Safety therefore involves more than how politely a model refuses a bad request.040404Containing an agent with environment walls rather than model promises is the core idea of our review of agent security patterns. We argued about where delegation should stop in 3 AImigo S1E9.Review · walls for agents3 AImigo S1E9 · autonomy boundaries04Containing an agent with environment walls rather than model promises is the core idea of our review of agent security patterns. We argued about where delegation should stop in 3 AImigo S1E9.Review · walls for agents3 AImigo S1E9 · autonomy boundaries

Boundaries can fail to contain data without any agent involved. On October 26, Cloudflare published a postmortem of a cross-customer exposure on Workers AI: between August 27 and September 9, 155 GLM-5.3 responses contained data from other customers’ requests. The model was offered with a context of about 1 million tokens, but the cache pool on the prefill servers held only 681,000. An SGLang bug let an over-length request proceed, and generation read cache pages left behind by earlier requests. Both the configuration and the engine were fixed; Cloudflare recommends SGLang v0.5.22 or later.050505The KV-cache pages read from other requests here are the PagedAttention mechanics we covered using vLLM as the example. SGLang got its own review too.Research Insights #32 · vLLM and PagedAttentionReview · SGLang05The KV-cache pages read from other requests here are the PagedAttention mechanics we covered using vLLM as the example. SGLang got its own review too.Research Insights #32 · vLLM and PagedAttentionReview · SGLang

The lesson echoes the agent incidents: isolation is defined by how a system behaves at the edge case, not by its advertised specification. Teams serving their own models should check that a request beyond the real limit fails with an error rather than quietly continuing.

On September 28, NVIDIA introduced its Open Agent Safety Platform. OpenShell constrains execution; Project Sentry moves monitoring and enforcement outside the main agent, onto BlueField-4. Millisecond quarantine is a vendor claim, not an independent guarantee. The product shift matters: action controls are being sold as separate infrastructure.

How reliable such control is when it is itself built on a language model was tested by Anthropic Fellows and Anthropic in Red-Teaming Auto Mode; the second version came out on September 27. In 79% of trials, a persistent attacking agent used an injection against the monitor to gain the ability to run an arbitrary bash command. Averaged over four attack types, the production monitors missed 65% of attempts in Claude Code’s auto mode and 72% in Codex’s Guardian. This is a test against a deliberately attacking agent, not an error rate in ordinary work; the authors use it to improve blocking classifiers. The lesson for teams: a language-model monitor is a useful extra layer but does not replace permissions the agent was never given.060606We reviewed the paper separately, including what the 79% means and how it differs from the averages across all attack types.Review · a classifier isn’t a wall06We reviewed the paper separately, including what the 79% means and how it differs from the averages across all attack types.Review · a classifier isn’t a wall

What to check in your own environment

In a safe test environment, give an agent a task with a prohibited external action and test the full sequence: the restriction holds, the attempt is visible, execution stops and access is revoked. If only the alert worked, the test is not complete. Test ordinary tool failures too: they must not become permission to seek a workaround.

On October 26, OpenAI released GPT-6.1 Terra in ChatGPT Work, Codex and the API. Standard prices per million tokens are $1 input, $5 output and $0.05 for cache reads: half Sol’s standard rates. OpenAI emphasizes visual tasks and application use; it still recommends Sol or Astra for the hardest coding tasks. Quality comparisons in the announcement are vendor measurements.

Terra enters a market where three notable launches have already established similar prices: Sonnet 5.5 on September 28, GPT-6.1 Sol on September 29, and Gemini 4 Argon on September 30. Argon initially had restricted access for trusted cyber defenders; its broader launch followed on October 13. These September launches fall within the period between the two 3 AImigo news editions.

The bottom tier moved noticeably with Claude Haiku 5.5, released October 7: $0.10 input and $0.50 output for contexts up to 100,000 tokens, a tenth of Haiku 4.5’s price. It is the first Haiku-class model with an effort setting. Anthropic reports 39.2% on Terminal-Bench 4.0, against 0% for its predecessor; Sonnet 5.5 scores 70.6% on the same benchmark.

The top tier was taken on October 28 by Gemini 4 Carbon, which Google calls its strongest model for cybersecurity and long-horizon tasks. Access starts with selected cyber defenders, then trusted partners in the Fairwind program; developers and the wider public will follow later, with no date announced. Google says the model is part of the US government’s voluntary pre-release testing. This continues the safety story: the most capable models increasingly ship in stages, starting with defenders.

October 29 snapshot. US dollars per million tokens; “cache” means reading stored input context. These published rates exclude tools, storage and human work. Prices and availability may change.

Anthropic reports savings of up to 30% per task for Sonnet 5.5 through lower token use. OpenAI says Sol approaches Astra in coding and computer use. Both claims come from the providers’ own tests; equal prices do not make models interchangeable. Results from different evaluation environments cannot produce a fair overall ranking.

OpenAI showed other price levers at DevDay on September 29. Ultrafast mode, the company says, runs eight times faster at six times the price. Pro is now split into $100, $200 and $500 tiers, and the limits of the $200 tier were reduced at the same time. For developers, OpenAI announced an Agents API with Computer Use, a Decision API for fast decisions and Private Intelligence with zero data retention. These are stage announcements: terms and availability need checking in each product’s documentation.070707I wrote down the full list of DevDay announcements on the day: it also includes Partner Apps, a marketplace and a cloud Codex harness.Knizhny kub · OpenAI’s plans07I wrote down the full list of DevDay announcements on the day: it also includes Partner Apps, a marketplace and a cloud Codex harness.Knizhny kub · OpenAI’s plans

What Artificial Analysis shows: same rate, different bill

Artificial Analysis offers an independent comparison: it runs every model through the same set of nine evaluations and calculates what one task cost at the provider’s standard prices. Version 5 of the index came out on October 27; about 60% of its tasks are held out from publication so that models cannot train on the answers. v5 scores cannot be compared with scores in launch announcements based on earlier versions.080808Why a shared leaderboard stops separating the leaders, and how much of a result comes from the model versus the harness: see my SWE-bench note and our Harbor Index review.Knizhny kub · SWE-bench and the leadersReview · model or harness08Why a shared leaderboard stops separating the leaders, and how much of a result comes from the model versus the harness: see my SWE-bench note and our Harbor Index review.Knizhny kub · SWE-bench and the leadersReview · model or harness

Data from Artificial Analysis model pages and leaderboard as of October 29. Costs use standard prices: $4 / $20 for Argon and Carbon, without introductory discounts. This is a measurement on AA’s evaluation set, not the cost of a task in any particular team.

The key finding sits in two adjacent rows. Sonnet 5.5 and Sol cost the same—$2 / $10 per million tokens—and score 54 and 55, yet an index task costs $10.02 and $1.27 respectively. The nearly eightfold gap arises because Sonnet 5.5 at maximum effort produced 340 million output tokens against Sol’s 64 million. When Sonnet 5.5 launched, AA noted that it had the highest token use of any model it had measured.

Terra shows the other side of the same arithmetic. Its tokens cost half as much as Sol’s, but at maximum effort a task is only a quarter cheaper: $0.96 versus $1.27. Terra uses more tokens (99 million against 64 million) and scores four points lower. Halving the rate does not halve the cost of the work.

The bill depends even more on the chosen effort level. According to AA, Opus 5.5 scores 35 at low effort for $0.98 per task and 58 at maximum effort for $10.04. Among the top ten models, cost per task varies more than 13-fold: from $0.96 for Terra to $13.04 for Claude Fable 5.1. A setting in an agent’s configuration can change the budget more than switching providers.

Spending concentrates in agent work. Four agentic evaluations carry 55% of the index weight but generate about 90% of output tokens; median use per task rose from 35,000 to 59,000 tokens in the move to v5. On October 28, AA compared agent harnesses: GitHub Copilot Fusion, which distributes a task’s steps across Anthropic and OpenAI models, scores two points below Claude Code with Opus 5.5 but costs $5.59 per task versus $11.42. Cost depends on how the agent calls a model as well as on the model itself.090909How much a harness adds with the same model was the topic of Research Insights #38 and of a separate review, “How much harness”.Research Insights #38 · harness design studyReview · how much harness09How much a harness adds with the same model was the topic of Research Insights #38 and of a separate review, “How much harness”.Research Insights #38 · harness design studyReview · how much harness

What this changes

Trying several models is getting easier; choosing from a price table is getting harder. One agent makes more attempts, another reasons several times longer, and a third calls on a person more often. Savings only materialize when a task ends with an acceptable result. Argon’s and Carbon’s pricing is a separate reminder: a month-long experiment and an annual budget may rely on different rates.

Cost of an accepted result = tokens per attempt × token price × attempts + tools + review time.

A practical comparison starts with one class of real tasks, identical permissions, harness and acceptance criteria, with the effort level stated explicitly. Count corrections, human time and failures alongside successful answers. Artificial Analysis data suggests where differences hide, but does not replace this measurement on your own tasks.101010One run is not enough for this measurement: agent results vary noticeably between runs. We looked at how many repetitions you need in a separate review.Review · noise in agent evals10One run is not enough for this measurement: agent results vary noticeably between runs. We looked at how many repetitions you need in a separate review.Review · noise in agent evals

On October 27, Oracle announced an expanded OpenAI partnership: AI Workspace connects ChatGPT with business applications and data, App Composer uses Codex to build applications hosted on OCI, and AI Security Mesh targets vulnerability discovery and remediation. These are announced products and capabilities; availability and constraints need checking for each deployment.

This extends the working-environment story: agents sit alongside the systems where decisions are made and records change. Discussing one real workflow—who authorizes an action and how it can be reversed—is more useful than counting connected tools.111111How to connect AI to an organization’s data without losing permissions, freshness and metric meaning is covered in the longread based on Nikolay Golov’s framework.Longread · AI and the data platform11How to connect AI to an organization’s data without losing permissions, freshness and metric meaning is covered in the longread based on Nikolay Golov’s framework.Longread · AI and the data platform

Models enter familiar systems

Another route into the enterprise: Google Cloud announced OpenAI models in its model garden on October 21. Access is a private preview for selected enterprise customers. Purchases can count toward existing cloud spending commitments; some capabilities, including Zero Data Retention and Confidential Compute, remain planned.

On September 29, OpenAI introduced Dots and Codex Cloud: persistent cloud agents and reusable development environments. Work can continue with the laptop off. Dots are rolling out gradually; Pro access initially excludes the UK, EEA and Switzerland, while Enterprise beta access is off by default. An announcement does not mean every user already has access.121212How a long-running agent hands work between sessions and what preserves its state is covered in Research Insights #34.Research Insights #34 · work handoffs12How a long-running agent hands work between sessions and what preserves its state is covered in Research Insights #34.Research Insights #34 · work handoffs

Anthropic now has a similar product. On October 21, it opened a research preview of Claude Duet: a background agent on Opus 5.5 that runs in its own virtual machine with a browser, sends a morning brief and reaches out only when something needs attention. Before sending a message, making a payment, signing a document or deleting anything, it shows the action and waits for approval; trading securities and moving money between accounts are off-limits even with approval. Access is via a waitlist for Pro and Max subscribers in the US.131313Why a payment needs separate consent, and what evidence settles a dispute later, we discussed in 3 AImigo S1E10; there is a recording and a longread.3 AImigo S1E10 · agents with payment rightsLongread · agents with payment rights13Why a payment needs separate consent, and what evidence settles a dispute later, we discussed in 3 AImigo S1E10; there is a recording and a longread.3 AImigo S1E10 · agents with payment rightsLongread · agents with payment rights

On October 7, ChatGPT gained Intelligent UI: answers can include forms, charts or interactive tools. This is another layer of the same shift: people interact with an interface created for the task as well as with text. A well-presented answer is easier to use; its data and calculations still need checking.

Development: model routing and usage-based billing

The month’s most visible tooling announcement came at GitHub Universe on October 28. Copilot Fusion assesses each task and distributes its steps across one or more models; it is the harness Artificial Analysis compared with Claude Code. Copilot projects run parallel agents across several repositories, and code review gained a Max tier for changes that need closer examination. GitHub Enterprise customers can now use every Copilot feature without a separate subscription and pay for the AI credits they consume; for now through GitHub Sales, with self-serve promised later.

The Octoverse report published the same day shows the scale: in September, more than 40% of PRs on GitHub started as an agent’s draft—about 110 million—up from 17% in March. GitHub calls this a lower bound because not every agent is detected. In September, 61% of Copilot users who chose a model themselves used more than one. Usage-based billing turns the cost per task from the previous section into a line on the team’s bill.141414What breaks in CI, version control and review under this flow is the subject of our study of agent development infrastructure. Gautam Korlam’s talk shows well what each extra agent attempt costs in CI.Longread · agent development infrastructureKnizhny kub · CI for agent-scale change14What breaks in CI, version control and review under this flow is the subject of our study of agent development infrastructure. Gautam Korlam’s talk shows well what each extra agent attempt costs in CI.Longread · agent development infrastructureKnizhny kub · CI for agent-scale change

The Stack Overflow 2026 developer survey, published on October 6, shows the same tool shift from another angle. Among respondents to the question about AI code agents, 65.5% use Claude Code, 58.7% GitHub Copilot, 29.5% Codex and 25.5% Cursor; several answers were allowed. The headline about trust in AI rising to 87% from 31% a year earlier needs care: the question and scale changed. Under the new wording, 48% trust output when it is easy to verify, and only 6.6% for important work decisions. The sample is self-selected: 30,903 responses collected from June 23 to August 5.151515I also revisited a deserted Stack Overflow: 64.1% of respondents say AI means they come there less often with simple questions.Knizhny kub · Stack Overflow 202615I also revisited a deserted Stack Overflow: 64.1% of respondents say AI means they come there less often with simple questions.Knizhny kub · Stack Overflow 2026

The flow of changes is reshaping infrastructure itself. On October 13, Anthropic described Scion, its own version control system, which replaced Git in its monorepo. According to the team, merged changes grew from about 2,500 to more than 30,000 a day between May and September; seven engineers built Scion, a review tool and related systems in five months. The system’s main clients are always-connected agents and CI jobs, so history, merge and rebase run on the server, and some checks moved to after merge. This is the company’s engineering report on an unusual scale; merge counts alone do not measure how useful the code is.161616We covered Scion in detail in Research Insights #39: it borrows heavily from databases, with a change log and derived indexes for history and blame.Research Insights #39 · ScionKnizhny kub · Scion16We covered Scion in detail in Research Insights #39: it borrows heavily from databases, with a change log and derived indexes for history and blame.Research Insights #39 · ScionKnizhny kub · Scion

On October 13, Cursor opened its Origin App Marketplace with over 20 integrations, spanning task management, CI, security and monitoring. Apps can respond to events and add checks or comments to code changes. More participants in the process also means more permissions granted.

GitHub made AI Scan for PRs generally available on October 15. It complements CodeQL in areas outside its coverage, including PHP, Bash, HCL and Dockerfiles. Findings are advisory: they neither fail the check nor block merges through repository rules. Fork and Dependabot PRs are excluded. Wider vulnerability detection does not automatically prevent a dangerous change.

Shared rules for agents are also moving under neutral governance. On October 22, Anthropic donated the Agent Skills standard, along with its name, GitHub organization and the agentskills.io domain, to the Agentic AI Foundation under the Linux Foundation—where MCP went earlier. Engineers from Anthropic, Google, OpenAI and other companies maintain it; according to Anthropic, more than 70 agent platforms support it, a million skills have been published and installs approach 280 million.171717What a skill consists of, how an agent selects one and why loading the file is not yet task success: Maxim Smirnov and I covered this on October 22, together with Open Knowledge Format.Review · Agent Skills and OKFKnizhny kub · Agent Skills and OKF17What a skill consists of, how an agent selects one and why loading the file is not yet task success: Maxim Smirnov and I covered this on October 22, together with Open Knowledge Format.Review · Agent Skills and OKFKnizhny kub · Agent Skills and OKF

Meta’s October 15 article on building Forum, a standalone app for Facebook Groups, shows how such rules work in one team. The team is moving toward a model where the agent runs the “plan → code → build → test → fix” loop itself, while an engineer sets the goal, steps in when it drifts and approves every change that ships. Working agreements are split into rules, skills, agents, commands and hooks; according to the authors, changes to the environment helped most: a simpler architecture, agent access to UI state in the simulator and a separate build for experiments. Meta reports parity with Facebook Groups on core features in eight weeks, but this is a single project without a controlled comparison. The authors themselves note that agents cut corners even in PRs that passed the checks.181818My margin notes on the article are in Knizhny kub: the decision to rewrite old components for agent speed raises the most questions.Knizhny kub · how Meta built Forum18My margin notes on the article are in Knizhny kub: the decision to rewrite old components for agent speed raises the most questions.Knizhny kub · how Meta built Forum

What this changes

Competition for the assistant is becoming competition for the working environment. Accuracy and generation speed now sit alongside retained context, connections, action logs and ways to hand work back to a person. Moving a request to another model is easy, and Copilot Fusion’s routing makes it nearly invisible; moving accumulated state and permissions is harder. Open standards such as Agent Skills lower that switching cost but do not remove it.

Teams can usefully separate three permissions: prepare a change, evaluate it and release it to users. One product may cover all three stages, but that does not mean one agent should make all three decisions. This matters especially when a checking feature only reports a finding.

On October 6, OpenAI published mathematical results from an internal model. As of October 28, the repository held 721 manuscripts in 373 groups; about 42% of the main results were formalized. Manuscripts, groups of related work and formally checkable results are different units. They do not justify a headline about “721 proven discoveries.”

On September 29, the AGMAI advisory group proposed publication guidelines: readable exposition, attribution of prior ideas, formalization status, information about attempts and support for human understanding. On October 15, the Simons Foundation announced a $15 million initiative with Anthropic and OpenAI, offering researchers funding and resources. Understanding the results now has dedicated organization and funding.

This is the scientific version of September’s bottleneck. A machine can produce more arguments than specialists can examine. Formal verification helps establish the correctness of a formalized statement; deciding whether it addresses the original problem, what it explains and how it connects to other results remains substantive work.

Biology speeds up the search for the next experiment

On October 21, OpenAI, PerturbAI and Scripps Research reported that thousands of agents processed existing data from millions of mouse brain cells in 32 minutes. They connected gene changes with literature and proposed research directions. This accelerates analysis of collected data; it is neither a biological experiment completed in half an hour nor a ready treatment.

The same shift is visible inside the labs. In an Anthropic Institute piece on the pace of AI development (September, no exact date on the page), the company estimates that Claude led 1% of research and engineering work in March and 26% in August, and took part in more than 90% of it. The percentages refer to a weighted basket of tasks inside Anthropic, and “led” means a defined level of autonomy on the Epoch AI scale. About 30,000 agents run concurrently on one internal platform, and the company claims 100% monitoring coverage of their actions. Claude also collects and assesses the automation data, so trusting these figures requires external verification.191919I sketched where the 2027 date in such forecasts comes from in Knizhny kub. We reviewed OpenAI’s report on accelerating its own research separately.Knizhny kub · who will build the next AIReview · OpenAI research acceleration19I sketched where the 2027 date in such forecasts comes from in Knizhny kub. We reviewed OpenAI’s report on accelerating its own research separately.Knizhny kub · who will build the next AIReview · OpenAI research acceleration

Together, these developments change the question we ask of scientific AI. Asking whether it can propose something new is insufficient. We need to know who will select proposals, check them and turn them into knowledge others can use. The analogy for engineering teams is direct: a large volume of plausible changes is only as valuable as the process for accepting the right ones.

On October 15, Meta released Muse Spark Open with available weights: 816 billion parameters, 42 billion active per token. A separate Muse license imposes commercial restrictions. A demonstration on two Mac Studio M5 Ultra machines, each with up to 512 GB of memory, does not imply ordinary laptop support. Active parameter count does not remove the need to store the weights.

Europe’s answer is Mistral Large 4: the preview arrived on October 6 and the weights appeared on Hugging Face on October 29. It has 1 trillion parameters, 52 billion of them active, and API pricing of $0.70 / $2.10 per million tokens. The license is a modified MIT: using and fine-tuning the model is free, but anyone selling it as a hosted service needs a separate commercial license. Mistral reports that the final version solves 41% of Terminal-Bench 4.0 tasks, against 28% for the preview.

Two releases came from China. On October 17, DeepSeek released V4.1 Flash 1017 with open weights: the company promises shorter outputs and lower cost per task, and the model is optimized for agent harnesses, including its own DeepSeek Harness. On October 23, Moonshot opened Kimi K3.1 Pro through its API with a 1-million-token context and promised to publish the weights by November 4; for now the model is closed.

Independent measurements temper expectations. In version 5 of the Artificial Analysis index, the gap between the leader and the best model with open or promised weights widened from 3 to 12 points; that model is currently Kimi K3.1 Pro. DeepSeek V4.1 Flash scores 32 for $0.60 per task, Muse Spark Open 25 for $2.02, and Kimi K3.1 Pro 46 for $3.68. A low token price for open models again does not guarantee a low cost per task.

Alternatives to NVIDIA: porting software and other accelerators

A Chinese counterpart at the hardware level is DeepSeek’s DeepGEMM-Ascend, published September 30. The library brings compute operations to Ascend 950 while preserving the DeepGEMM API. It requires specialized hardware and particular software versions. Competing with NVIDIA means reducing porting effort as well as building chips; interface compatibility is a step toward that, not proof of equal performance.

AMD takes a different approach: on September 28, it announced an agreement to buy Fei-Fei Li’s World Labs for about $8.2 billion in stock. The deal was signed on September 26; closing is expected by year-end and depends on approvals. AMD gets Marble for generating 3D scenes, the Atlas model and technology for moving robotics tasks into simulation, and Fei-Fei Li is to become AMD’s chief scientist. The company ties the deal directly to the design of future computing systems: world-model researchers will sit next to the engineers who design accelerators.202020What interests me most is the real-to-sim-to-real loop: the deal’s success should be measured by the cost of the full cycle, from building the environment to checking on a real robot.Knizhny kub · AMD and World Labs20What interests me most is the real-to-sim-to-real loop: the deal’s success should be measured by the cost of the full cycle, from building the environment to checking on a real robot.Knizhny kub · AMD and World Labs

Oracle showed another route: on October 27, it announced that Google TPUs will become available in Oracle Cloud Infrastructure. Oracle says this makes it the first hyperscale cloud outside Google to offer them; the announcement discloses no pricing, regions or dates. Together with the OpenAI announcement the same day, this positions Oracle as a place where models and hardware from different vendors meet.

Available weights and alternative hardware give teams more choice over where data is processed and how models are deployed. That control comes with licensing constraints, memory requirements, porting work and operations; the Cloudflare postmortem in the safety section shows that serving mistakes become yours too. Open weights alone do not make infrastructure cheap or independent.

Compare the cost of serving your workload: equipment, utilization, idle capacity, support and result verification. Buying APIs through an existing cloud, discussed earlier, addresses a different need by simplifying procurement. These options should be judged against the team’s requirements, rather than a promise of universal independence.

This month’s biggest stories connect provider economics with how agents work. The IPO filing and quarterly results open financial data, models lower token prices, independent measurements show that cost per task differs from list price by multiples, integrations provide access to business processes, and investigations show the cost of mistakes. Our working hypothesis is that managed agent work is becoming the unit of adoption. It includes the task, environment, permitted actions and acceptance. That is our interpretation of these stories; it does not mean every team already needs persistent cloud agents.212121This hypothesis continues the AI-Native organization longread and 3 AImigo S1E8: agents speed up execution, and then processes, roles and decision rules have to change.Longread · AI-Native organization3 AImigo S1E8 · AI-native company21This hypothesis continues the AI-Native organization longread and 3 AImigo S1E8: agents speed up execution, and then processes, roles and decision rules have to change.Longread · AI-Native organization3 AImigo S1E8 · AI-native company

The central debate for 3 AImigo is where a weak model is holding us back and where the process is missing. A new model helps if it improves a measurable result. More autonomy helps if the team understands its limits and can intervene. These are the two measures by which to judge the next month.