Open weights reached frontier scale this week. The scorecards stayed in-house.

Between October 11 and 17, five model releases and one hardware launch landed in seven days. Meta and Google put frontier models on the API, Meta, NVIDIA and Reflection published open weights at sizes that used to be closed-only, and the first NVIDIA RTX Spark laptops went on sale. The throughline is that the choice of where to run a strong model got much wider this week, while the evidence for choosing between them stayed almost entirely in the vendors' hands.

The frontier price tier is settling at $4 and $20

Meta released Muse Wonder on October 12, its first frontier model, through the Meta Model API and Muse Code; the larger Wonder Full is limited to Muse Code subscribers. TechCrunch, citing Meta's model card, puts Wonder at $4 per million input tokens and $20 per million output tokens, roughly the same as GPT-6 Sol or Claude Opus 5.5.

A day later Google made Gemini 4 Argon generally available. Its pricing page lists $2 input and $10 output per million tokens through November 30, then $4 and $20 from December 1, with output including thinking tokens and only implicit caching supported. So by December, Google's and Meta's flagships will list at the same price, in the tier TechCrunch puts Opus 5.5 and GPT-6 Sol in. Price stops being the tiebreaker, and what you pay will depend more on how many tokens each model spends to finish your task.

Open weights reach frontier scale, with different strings attached

Three open releases in three days pushed self-hostable models into the hundreds of billions of parameters:

• Muse Spark Open (Meta, October 15): about 816B total and 42B active parameters, a 1M-token window, text, image, audio and video in, per the Hugging Face launch post. The Muse Model License requires a "Built with Muse Spark Open" notice and a separate agreement with Meta for model-as-a-service businesses above $20 million in annual gross revenue.
• Nemotron 3.5 Bolt 74B-A9B (NVIDIA, October 15): a mixture-of-experts model packed into 24GB with native 2-bit quantization for local agent work on RTX and DGX Spark, per NVIDIA.
• Beam (Reflection, October 16): 501.3B total and 23.4B active parameters, a 1M-token window and a plain Apache 2.0 license, with no revenue threshold or custom use terms.

The license split matters more than the parameter counts. For most teams building their own product, Meta's terms will not bite. For inference providers and resellers, Beam is the one they can host without a conversation with the model's maker. Size still decides who can actually run these: Muse Spark Open's BF16 checkpoint is about 1.5 TB, and Hugging Face's own Mac test needed a 512GB Mac Studio and an aggressive quantization.

Local hardware arrived, with an asterisk on memory

NVIDIA's RTX Spark laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI went on sale October 16, with NVIDIA quoting up to 128GB of unified memory and up to 1 petaflop of FP4 compute. Microsoft's Surface Laptop Ultra page lists it at $2,599.99 and says it starts at 24GB. NVIDIA's release also lists the first PyTorch-CUDA build for Windows on Arm, native Python and Conda builds, and PAIR, which lets an RTX Spark machine serve inference to other computers on a home network. Bolt was sized for exactly this class of machine. The pieces for running a capable agent model on your own desk now exist, but the memory tier you buy decides which of this week's open models you can load.

Almost every number this week was self-graded

Meta says Wonder matches GPT-6 Astra and Fable 5.1 on its own tests. Google's headline Argon benchmarks are its own. Meta calls Muse Spark Open the most capable U.S. open-weight model, as TechCrunch notes, based on its own evaluations. Reflection's 80.1% on Terminal Bench 2.1 and its claim of roughly 3x fewer reasoning tokens come from its own model card. NVIDIA's RTX Spark performance claims ship without published methodology. None of this means the claims are wrong. It means none had been independently replicated when these models went on sale.

The one step in the other direction came from Anthropic. On October 15 it said it had embedded six METR staff with access roughly comparable to its own risk teams, and that METR can publish its risk assessments without Anthropic's approval, subject to redaction of predefined categories. That arrangement covers safety risk, not product benchmarks, so it does not settle which model codes best. But it is the week's only example of a lab giving an outsider the access needed to check its claims.

Also this week

• HCLTech and Anthropic announced a Global Premier Partnership on October 12, including a dedicated Anthropic business unit; no financial terms were disclosed.
• NVIDIA pledged $250 million in compute over five years for U.K. agentic AI safety research on October 14, with providers and commercial terms still to be agreed.

What it adds up to

Until recently, picking a strong model mostly meant picking between two or three closed APIs. After this week Meta sells a frontier model through its API, Google's flagship reaches the same price on December 1, three large open-weight options with different licenses, and laptops built to run the smaller ones. The practical consequence is that your own evaluation set is now the most valuable tool you have. Run your real tasks through the candidates while Argon's introductory pricing lasts, measure tokens per finished task rather than price per token, and treat every leaderboard figure from this week as a claim until someone outside the company reproduces it.