Last issue was about capability finally getting load-bearing infrastructure underneath it. This week the infrastructure question moves to money and rules, and to how much of what is announced can actually be checked.
Smaller models keep getting cheaper and better while the people building them argue about what that does and does not mean for general intelligence. AI for science arrives with hundreds of mathematical claims that cannot yet be verified and a competing argument that real science needs experiments. And the economics and politics show how much rides on opaque numbers: flat-rate subscriptions that quietly dominate compute, and a diplomatic scramble over safety rules that mostly ignores what China actually does.
Below the three focus stories, Also This Week covers agents and the engineers using them, model failure modes and industry signals to hold loosely.
Anthropic shipped Claude Haiku 5.5, its first Haiku-tier update in about a year, and priced it to match OpenAI's GPT-6 Luna. The company calls it its cheapest, fastest and most capable small model, costing about 75% less to run than Haiku 4.5 on average; Artificial Analysis scored it 43 on its Intelligence Index, up 26 points in a year. Pricing is tiered by prompt length: $0.10 per million input tokens and $0.50 output under 100K tokens, rising to $0.50 and $2.50 above that, with cache reads at $0.01 and $0.05. Anthropic positions it as the high-volume subagent next to Opus 5.5 or Sonnet 5.5 — one engineer's summary was 'Opus does the heavy thinking, Haiku does the high-volume work' — and the same day it halved Sonnet 5.5's cache-read price from $0.20 to $0.10 per million tokens (about 20% cheaper on long agentic work) and folded computer-use and browser-use loops into its Python and TypeScript SDKs.
At the other end of the cost curve, Reflection launched Beam, a text-only 501-billion-parameter mixture-of-experts model (23 billion active) trained from scratch in the US, with Apache-2.0 weights promised this month. The team cites 23.8 trillion pretraining tokens, about 10,500 GB300 chips and four weeks each of pretraining and reinforcement learning, and claims 80.9 on SWE-bench Verified. Independent readers place it around the level of GLM 5.2 and behind GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash; one analyst estimated only about 12% hardware efficiency in pretraining. The reason it matters, in the AINews write-up, is that a segment of the market has been waiting for a credible from-scratch American open model, and Axios reported that other US labs will ship open models this month.
Against that backdrop, Nathan Lambert's Interconnects essay tries to explain why researchers inside the labs say AI will do their jobs better within a few years. His answer is that they are watching a real, massive acceleration in infrastructure and engineering capability: he agrees models will be superhuman distributed GPU engineers within a few years, which makes experimentation and tinkering far easier. But he argues that speeds up what he calls 'parallelized, AI-assisted language modeling' rather than making the models different in nature, and that it moves value from good execution to good ideas. His title gives the position: rapid progress, but not towards general superintelligence.
The three pieces describe one trend from three angles: capability per dollar is climbing fast at the small end (Haiku), the open end is being rebuilt by new US labs (Beam), and the people closest to the work are betting that the next gains come from engineering throughput. That is progress with a clear mechanism, which is a more useful claim than 'superintelligence is near' or 'it has plateaued'. Keep in mind that Beam's numbers are mostly the lab's own claims and the Haiku score is one index.
OpenAI published 722 mathematics manuscripts from an unreleased internal model in a public repository, grouped into 372 families and drawn from an evaluation of roughly 4,000 research problems, at an average of about three hours of ChatGPT Pro compute per result. It says it consulted the Institute for Advanced Study's advisory group on mathematics and AI about the release. The AINews write-up reports that most experts agree many of the 'top 500 open problems' were solved; the headline result, a 'Quasi-Riemann Hypothesis', is described as somewhere between a Fields Medal result and the biggest result in number theory in 200 years, and a rival researcher said 'it is obviously the most significant moment in mathematical history'. Specific claims, such as integer multiplication faster than n log n and a uniqueness result for the elastic inverse problem open in 3D since 1994, are reported by individual commentators and have not been independently verified. The model is not released.
Mathematics has a cheap verifier — a proof can be checked — but most of science does not. That is the argument of Periodic Labs, whose founders Liam Fedus and Ekin Dogus Cubuk appeared on Latent Space a year after launch. They say intelligence alone is not enough for scientific discovery and describe 'synthesis superintelligence': reinforcement learning grounded in physical experiments, AI-powered characterization and simulation (including density functional theory), high-throughput labs, and models that learn from the whole process of doing science, including failed experiments, instead of only its published results. Giving each instrument a bit of intelligence, in their phrasing '140 IQ', is meant to let autonomous labs compress decades of trial and error into months.
A third voice complicates the story from biology. In a Latent Space panel, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido argue that AlphaFold's breakthrough was 'only the beginning': scaling compute and data alone will not solve biology, protein structure prediction is far from solved, finding the right scaling law matters more than blindly adding size, and protein language models may already contain scientific knowledge nobody has unlocked. They add that building a virtual cell requires fundamentally different datasets, and that Biohub's goal of curing disease needs 10x, not incremental, breakthroughs. In the biomedical issue this week, BioByte 174 reported a concrete case of the same idea: a design pipeline combining a diffusion model, LigandMPNN and AlphaFold3 made enzymes for reactions where natural enzymes struggle, though with weaker activity than natural ones on one test.
AI for science is moving on two fronts at once: fast, cheaply checkable domains like mathematics, where claims arrive in the hundreds and verification becomes the bottleneck, and slow, experiment-bound domains like materials and biology, where labs think the missing ingredient is data from the physical world. Treat the 722-manuscripts story as a claim to wait on, and the experiments-first argument as the more concrete bet.
SemiAnalysis' report on subscription value shows how much of the AI business sits inside flat-rate plans. By its rough numbers, subscriptions are only about 10% of Anthropic's revenue but over 40% of its inference compute, lowering blended revenue per megawatt by about $36 million, so modelling a lab's finances requires understanding its usage limits. A plan is not worth a fixed dollar amount: credits are consumed at different ratios for each model and token type, so the same $200 plan has a different API-equivalent value depending on model and workload, and labs change the limits (OpenAI cut its $200 plan's usage in half and removed relative-usage claims). The report's headline is that Anthropic subscriptions offer 5x or more value than OpenAI's, a claim backed by a dashboard available only to the firm's subscribers.
On the policy side, SemiAnalysis' China piece walks through a fast diplomatic sequence: Dario Amodei's 12 September essay arguing that frontier labs should deliberately slow capability growth, supported within a day by Sam Altman and Elon Musk; Donald Trump calling fears of AI destroying humanity 'a HOAX' on 14 September; the first US-China AI dialogue on 20 September; Xi Jinping's state visit from 23 September; and on 29 September a one-page voluntary White House Accord on Super Intelligence signed by Trump and six executives. The report's point is that nobody asked what China is doing. Its answer: Beijing's published stance is the most safety-forward of any major AI power outside Europe, and the 14 September AI Safety Governance Framework 3.0 warns of 'self-accelerating' autonomous learning and adds a model-risk category for behaviour deviating from expectations, including models that deceive evaluators or disable shutdown scripts.
Nathan Lambert argues that the cyber-risk discourse is 'broken' in a way that makes bad policy more likely. He sorts the debate into three camps (open weights as an unacceptable societal risk, open weights as necessary for defence, and Chinese labs continuing to release open weights with strong cyber capability) and says the loudest camp, including Anthropic's report on GLM-5.3 as an offensive cyber tool, fails to ask what happens if open models are banned or why Chinese companies judge them safe to release. He points out that publicly documented cyber attacks to date have mostly involved closed models, suggests 'open dangerous, closed safe' may really be 'open unsafe, closed unsafe', and notes that in some settings, such as air-gapped government networks, open-weight models are the only deployable tools. He is frank that this is an argument about trade-offs, not a conclusion.
Flat-rate subscriptions, hardware supply and policy are the same question seen from three sides: who bears the cost of making intelligence cheap, and who writes the rules while it happens. The consistent lesson is that single-number comparisons (a plan's 'value', 'open is dangerous', 'China doesn't care about safety') hide the trade-offs, and the SemiAnalysis claims in particular rest on data and dashboards the reader cannot inspect.
Craig McLuckie and Joe Beda, who helped create Kubernetes, are betting that coding-agent harnesses belong in the cloud with control loops that recover from failures, the way Kubernetes manages applications. Their company Stacklok's open-source 'Mecatl' started in June; the pitch is to move the locus of value from the desktop to the cloud, where reliability, tool isolation and context preservation can be managed. Gergely Orosz's LDX3 keynote summary, 'The state of the tech industry in 2026', lists what changed in a year (practically nobody writes code by hand, 5 to 10 parallel agents are common, the IDE is fading, everyone builds their own harness), what broke (code review has become theatrical, quality and reliability are down, assumptions about code volume failed) and what comes next, naming cloud coding agents and harnesses. Read together they are the same thing seen from the vendor and the practitioner.
Two ByteByteGo explainers cover failure modes that matter to anyone building on a model. Sycophancy is the habit of dropping a correct answer when a user pushes back, because training rewards approval as well as accuracy; the apology can even make the wrong answer sound more reliable. And the 'lost in the middle' blind spot means facts placed in the middle of a long prompt are used less reliably than those at the start or end (a coding assistant told in the middle that logs must be kept 37 days wrote cleanup code for 30), with larger context windows no cure.
The AINews issue on TypeSafe and Jev reports that the company crossed $100 million in annual recurring revenue three weeks after launch, a figure Sequoia reportedly leaked, and that several vendors shipped 'decision' models the same day that return typed answers in a single pass instead of free text. A quiet-day AINews leads with OpenAI reportedly dismissing three safety researchers tied to the METR audit of its summer incident; the researchers say it was for prioritising safety, and OpenAI reportedly says they mishandled confidential information. Both are second-hand reports from social media summaries and should be read that way.