Meta put 816 billion parameters on Hugging Face with the weights open — 42 billion active per token, a million-token context, and $1.25 per million tokens in on OpenRouter — and Muse Spark Open still sits behind at least six open-weight models on Artificial Analysis's aggregate index — among them GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40, against its 39. That is the honest shape of the open-weight race: America's best downloadable model is now genuinely good, leads the Chinese field on several tool and orchestration benchmarks while trailing it on others, and is not the best downloadable model. The same day, Anthropic published the three-tier access ladder it uses for outside evaluators and disclosed that six METR staff now work inside the company on employee-grade hardware and internal tooling, with publication rights it says it cannot veto. Both moves hand over something a 2023 lab would have hoarded — weights in one case, internal visibility in the other — because in both cases distribution and credibility are now worth more than the secret. The day-zero read is unchanged and slightly sharper: the inputs keep getting cheaper and more inspectable, so the scarce asset is a well-chosen old problem, not access to the model.

The Tide

Status. No shift. Two confirmations. cost-collapse gets its open-weight datapoint: a frontier-adjacent model you can download, serve on two Mac Studios, and rent at $1.25/$4.25 per million tokens. governance-as-market-structure gets the cleanest evidence yet that the access terms are being written voluntarily, by a vendor, ahead of any regulator requiring them. Neither bends the curve; both sit squarely on it. The tide holds.

Cost-collapse, the open-weight side: 816 billion parameters, downloadable, and not the best download

Meta released Muse Spark Open on Thursday: 816B total parameters with 42B active per token across 256 experts (8 active), 62 layers, a 1M-token context window, text/image/video/audio in and text out, and a reasoning_effort dial defaulting to xhigh. Weights are on Hugging Face in BF16 (~1.5TB) and NVFP4 (~530GB); the model card suggests at least 700GB of VRAM for NVFP4, which lands it on four Blackwell Ultra GPUs, eight H200s, or, per the model card's reference configuration, two 512GB M5 Ultra Mac Studios via MLX. Hosted on Baseten, Fireworks, Modal, Nebius, OpenRouter and Together, with OpenRouter listing $1.25 per million input tokens and $4.25 output. The licence permits fine-tuning, distillation and redistribution, requires a "Built with Muse Spark Open" notice, and only bites on Model-as-a-Service providers above $20M of trailing-twelve-month revenue — internal use, research and powering your own product features are exempt. Now the number that matters: Artificial Analysis scores it 39 on its Intelligence Index, behind at least six open-weight models — among them MiMo-V2.6-Pro at 46, GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40. Meta's claim is "most capable US open-weight model," and that claim holds. The broader one does not. The split underneath is more interesting than the aggregate: Muse Spark Open leads on MCP-Atlas (89.2 against Qwen3.8's 84.5 and GLM 5.3's 84.2), on JobBench (61.8 to GLM 5.3's 61.4), on HLE-Diamond without tools (26.1 to Kimi K3's 22.2) and on ZeroBench with tools (50.0 to 41.0); it loses Toolathlon Verified (72.8 to Kimi K3's 76.5 and GLM 5.3's 73.0) and OSWorld Verified (80.4 to 84.8), and trails badly on Terminal-Bench 2.1 at 76.4 against Kimi K3's 85.0, GLM 5.3's 83.9 and Qwen3.8's 82.0, and on DeepSWE 1.1 at 60 against 67.5 and 66.9. Most of these are Meta's own published comparisons, so read them as vendor benchmarks.

So what. Stop treating "open weights" as one procurement decision. On MCP-style orchestration and multimodal reasoning the best American open model is now the one to beat; on terminal work, computer use and software engineering the Chinese models are still ahead. If you are standing up an agent that calls tools and reads documents, Muse Spark Open at 42B active is the new default to benchmark against — and at $1.25 in, the economics now favour running your own evals rather than arguing about the index.

MarkTechPost — Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
Artificial Analysis on X — "Meta has released Muse Spark Open, an 816B (42B active) open-weights model that scores 39 on the Artificial Analysis Intelligence Index, making it the #1 US open-weights model"
Hugging Face — meta-models/Muse-Spark-Open-816B-A42B

Governance-as-market-structure, confirmed: the lab wrote the access terms before anyone made it

Anthropic published "An update on independent evaluators" on Thursday afternoon, and the substance is a published access ladder rather than a statement of principle. Tier 1: narrowly scoped investigations, Anthropic controls access and must approve publication. Tier 2: evaluators publish without approval and receive more information; Anthropic may require removal of specific predefined categories, and the evaluator may say publicly whether it thinks the redactions were justified. Tier 3: employee-like access, with publication rights similar to Tier 2. Six METR employees have now been onboarded at Tier 3 with company hardware, office space, and access to internal tools and systems roughly on par with Anthropic's own risk-assessment teams; METR has reviewed several of Anthropic's Risk Reports over the past year and will publish its own analysis of the company's risk levels, incidents and practices. Anthropic states plainly that it "can't veto publication," while reserving redaction over narrow categories — customer data, attorney-client privileged material, model architectures — and says that where architecture detail is highly relevant it will generally seek to share it, subject to explicit internal approvals. SecureBio is reviewing the chemical and biological chapter of the August 2026 Risk Report; Accenture, with Faculty, is working toward evaluation, red-teaming, alignment and safeguard testing; Hugging Face is among several in early discussions. The stated criteria for an evaluator: frontier expertise, independence with no equity or financial dependence, sufficient access, and freedom to publish. Anthropic says it is doing this voluntarily and that every frontier company should. Note the timing — METR published its control-assessment and risk-assessment proposals on Wednesday, covered here, asking for roughly this. The answer came in about a day.

So what. This is the template, and it is now public, which means it is quotable in a procurement conversation. If you are buying frontier capability into a regulated process, the question to put to a vendor is no longer "do you work with external evaluators" but "which tier, who, and can they publish without your approval." Anthropic has just made the ladder legible enough that answering "Tier 1" out loud is costly.

Anthropic — An update on independent evaluators
METR — Our approach to working with AI companies
METR — How an assessor could investigate control and monitoring of internal AI agents

The Waves

Workday went to investors and asked for a share of the labour budget, not the software budget

At its annual investor conference in Las Vegas, Workday's chief AI officer Joel Hellermark told analysts that automating HR tasks could take the company "beyond the software budgets into the labor and services spend," noting customers are "spending tens of millions of dollars on outsourcing" and that self-service AI agents had raised resolution rates for common tasks by 25%. Co-founder and CEO Aneel Bhusri, back in the job since February, said Workday expects consumption-based billing for agents rather than per-seat licensing — "a far better model than a people-based subscription-based model" — and compared it to what Workday itself pays Google Cloud, AWS and Azure. He also said the company uses AI internally to grow without adding headcount and might "just not hire in areas we were hiring in the past." The same week Workday opened "headless" access to its HR and finance systems, letting agents read data and perform tasks without the UI, through Model Context Protocol tools that respect existing permissions, approvals and audit trails. The mirror image showed up on the buy side the same day: The Information reported that Cognition, which makes the Devin coding agent, is now measuring how efficiently individual engineers use AI — estimating the human-hours a session replaced, dividing by its cost, and automatically raising a user's token limit when their grade is "healthy" or "satisfactory." CTO Calvin Giroux's analogy was a credit limit. Cognition is also shipping Devin Coach, which flags the twenty to thirty most common wasteful patterns, including ambiguous requests and picking an unnecessarily expensive model.

Roadmap implication. Two sides of the same repricing, and both land in the 2027 budget. Vendors are moving from per-seat to per-outcome, which means the line item you negotiate stops being headcount-shaped and starts being usage-shaped — so model your agent spend as a variable cost with a ceiling, not a licence. And the vendor-supplied efficiency grade is coming to your org whether you ask for it or not: decide now whether you want that telemetry, because the first team to measure human-hours-replaced per dollar inside your own company gets to set the benchmark rather than receive it.

The Register — Workday wants a slice of customers' labour budget that its AI agents could replace

The bill for AI-found vulnerabilities landed on a bank's income statement

Bank of America disclosed on Wednesday's third-quarter earnings call that it had put on roughly 1,000 contractors during the second quarter to patch software flaws surfaced by Anthropic's Mythos model, alongside about 3,000 existing internal staff, with CEO Brian Moynihan putting the expense in the hundreds of millions across that remediation and AML compliance work — the clearest dollar-denominated reading yet of a gap that opened in the spring, when Mythos-class scanning first sent US banks hunting through their own code. The asymmetry is the story: the model found the flaws at the cost of inference, and closing them took a four-figure headcount and a visible hit to the expense line. Alongside it, a Semperis survey of ransomware victims reported that more than 84% of surveyed organisations hit by ransomware said attackers had used AI somewhere in the attack (Semperis's own release puts the figure at 87% — reports differ), and 44% said criminals had exploited the organisation's own AI systems — a vendor survey of victims, so treat the precision accordingly, but the direction matches everything else on this thread.

Roadmap implication. Budget for the remediation, not the scan. The cheap half of security just got cheaper and the expensive half did not move, so any board paper that treats AI-assisted vulnerability discovery as a cost saving has the sign wrong on the second line. The operational question for the next quarter is throughput: how many findings per week can your team actually close, because that number — not the scanner's output — is now your real security posture. And 44% of victims reporting their own AI systems were exploited says the agent you deployed this year is already in scope.

CNBC — Bank of America hired 1,000 contractors to patch software flaws exposed by Anthropic's Mythos
Cybersecurity Insiders — Study finds AI involved in more than 84% of Ransomware Attacks

TSMC is adding a 3nm fab it would not normally build, because the AI orders will not clear

TSMC said on Thursday it will build an additional 3-nanometer fab in Tainan, with production set for the first half of 2028, adding to its existing 3nm capacity in Arizona and at Kumamoto in Japan. The notable part is the node. TSMC's leading-edge attention and capital are pointed at 2nm in Hsinchu and Kaohsiung and at A14 in Taichung; adding capacity at N3 in 2026 for 2028 output is a bet that demand at a node two generations back from the frontier will still be unsatisfied two years from now. That is not how the foundry normally allocates — mature nodes usually get harvested, not expanded — and it is a cleaner signal about the durability of AI order books than anything in the public equity tape, which spent the same week repricing the companies that finance the buildings.

Roadmap implication. Read this against the write-down arithmetic that ran here on Wednesday. The company with the best view of the order book, and the most to lose from building into a bubble, just committed capital to 2028 capacity at a node whose customers are mostly accelerator and networking silicon. If you are planning compute procurement past 2027, the supply picture at N3 is loosening, not tightening — which argues for negotiating longer terms now rather than locking scarcity pricing into a three-year contract.

Techmeme — "TSMC says it plans to build an additional 3nm fab in Tainan, Taiwan, set for production in H1 2028, alongside its 3nm capacity in Arizona and Japan's Kumamoto"
The Information — TSMC to Add a New 3nm Fab as AI Orders Pile up

The Ripples

Cursor's own engineers asked whether customer data had already reached SpaceX

The Information reported on Thursday that in late July — weeks before SpaceX's AI unit introduced Grok Bot in August, and while SpaceX was still in the process of acquiring Cursor — Cursor engineers raised concerns internally that data from their own heavy testing of the tool, potentially including confidential information about customers, suppliers and business partners, may have been sent to SpaceX's AI unit. One of Cursor's lawyers warned in internal communications viewed by The Information that if it became known the company was using that data for training, it could cause a "huge trust problem."

Do this now. Do this now: check whether your coding agent's vendor has changed corporate parent since you signed, and what your contract says about training on telemetry from internal dogfooding. An acquisition does not renegotiate your data terms for you, and the people who noticed the problem here were the vendor's own engineers — which is to say, nobody on the customer side was positioned to catch it.

The Information — Cursor Staff Investigated Whether Confidential Data Reached SpaceX for AI Training

ASML looks like it is building a hybrid bonder, and the tool runs on a track

SemiAnalysis published a deep dive on Thursday arguing from 27 bonding patent families filed between December 2022 and November 2024, a March report from The Elec, and ASML's own 2024 investor-day roadmap that the company is building a die-to-wafer hybrid bonding tool — the harder of the two bonding problems, and the one where ASML's alignment and particle-control expertise should transfer. ASML has announced nothing. The design in the anchor filing is unusual: instead of a bond head picking dies off a diced wafer, each die rides its own holder around a SuperTrak-style track past inspection, cleaning, plasma activation, metrology and alignment stations, with the stations reaching in to the die rather than the reverse. SemiAnalysis flags unresolved problems — feed-forward metrology across a shuttle rated to 4G of acceleration, contact and bond-wave control ASML has never done, track-side station modules nobody currently sells, and a SuperTrak platform rated ISO 6 against the ISO 3 a bond interface needs — and puts volume production no earlier than 2030, with ASML management guiding revenue "well beyond 2030."

Do this now. Nothing to act on this quarter, and that is the point: treat this as a 2030 competitive read, not a 2027 one. The near-term consequence is for Besi, whose hybrid-bonding business is roughly 70% logic bonders concentrated at TSMC and Intel — both of them ASML customers of twenty-plus years — while its HBM growth case slips as adoption moves to 2029 or beyond and Korean challengers price at about half.

SemiAnalysis — ASML Is Building a Hybrid Bonding Tool!

A new benchmark gave agents the ambiguous version of the task, and the scores fell off

FDE-Gym, posted Thursday, evaluates agents on realistic customer-delivery work: requirements that are ambiguous and need clarifying, and constraints that are never stated and have to be discovered. The benchmark is 95 tasks drawn from 20 customer scenarios across 18 sectors, run against 13 model configurations from seven providers, and the headline number is what withholding the unstated constraints costs: 0.28 off the score for leading models. The best configuration averaged 0.56 and fully resolved 14 of the 95 tasks; on the 23 tasks where it had obtained every key fact, it still cleared 0.8 on only 10. It is the measurement that has been missing from the agent-benchmark crop, which mostly grades models on problems a human has already reduced to a clean specification.

Do this now. Add one row to your own agent evals this quarter: the same task, briefed badly. The gap between those two numbers is the honest size of the human scoping layer you still need to staff — and it is a better capital-allocation input than any leaderboard, because it tells you whether to hire a solutions engineer or buy more tokens.

Hugging Face Papers — FDE-Gym: The Last Mile of AI Delivery
arXiv — FDE-Gym: The Last Mile of AI Delivery

A team of agents post-trained a base model up to its official instruction-tuned version

Athanor, also posted Thursday, reports that a team of autonomous agents using an open research system independently post-trained a small base model into a cross-domain generalist that matched or exceeded the official instruction-tuned release of the same base model. The work is agent-run research — the agents chose the data actions and the training recipe — and the result is parity with a human-built artefact rather than a new frontier.

Do this now. Parity, not superiority, is the interesting claim: it says the post-training recipe for a small model is now a solved-enough problem to be delegated. If your team is spending senior researcher time fine-tuning small models for internal tasks, that is the work to point an agent harness at first — the ceiling is known, the evaluation is cheap, and a matched result is a real result.

Hugging Face Papers — Athanor: An Agent-Native System for Autonomous Post-Training Research
arXiv — Athanor: An Agent-Native System for Autonomous Post-Training Research

Read this and every past edition at excelsiorgroup.ai/insights/signal.