Muse Spark Open at $0.84 vs $2.01 for GLM-5.3 — AI Digest

2026-10-17

Claude Haiku 5.5 launched at $0.10 per million tokens, about 90% cheaper than Haiku 4.5; Meta's Muse Spark Open costs $0.84 per task against $2.01 for GLM-5.3.

The takeaway for you today: models for high-volume work now cost cents. Claude Haiku 5.5 charges $0.10 per million input tokens and the pplx-decider v1.2 classifier 1–2 cents. Compare models by cost per task, not by leaderboard points.

Top stories today

Claude Haiku 5.5: $0.10 per million tokens, about 90% cheaper than Haiku 4.5

Anthropic has released Claude Haiku 5.5 at $0.10 per million input and $0.50 per million output tokens for prompts up to 100K tokens, about 90% cheaper than Haiku 4.5, according to DL Weekly.

On Terminal-Bench 4.0 at max effort, Claude Haiku 5.5 scores 39.2%, against 16.4% for GPT-6 Luna. Anthropic itself says Claude writes about 90% of Claude Code's code; merged pull requests doubled between June and September 2026 while the reverted share fell by more than 40%.

A token is a chunk of a word that the model bills for: input tokens are what you send, output tokens are what it writes. The Claude Haiku 5.5 price applies to prompts up to 100K tokens.

What it means: client work — if the model inside a chatbot you build for a client gets about 90% cheaper, the per-reply running cost barely affects the quote: roughly a tenfold saving on the same traffic.

Sources: DL Weekly · ClaudeDevs

Muse Spark Open: $0.84 per task against $2.01 for GLM-5.3

Meta released Muse Spark Open on October 15, 2026, an open-weights model with 816 billion parameters (42 billion active per token); Artificial Analysis scores it 39, the best result among US open models.

Together, Fireworks and Baseten charge $1.25 and $4.25 per million input and output tokens for Muse Spark Open. Per Artificial Analysis index task that works out to $0.84, against $2.01 for GLM-5.3 (score 45) and $0.27 for DeepSeek V4.1 Flash.

Open weights are the published parameters of a trained model: anyone can download and run them, though the license may restrict how. The Muse Spark Open license is custom: model-as-a-service businesses above $20M in annual revenue need a separate agreement with Meta.

What it means: your wallet — $2.01 against $0.84 makes a Muse Spark Open task 2.4 times cheaper than GLM-5.3, yet DeepSeek V4.1 Flash is another 3 times cheaper ($0.84 against $0.27). Everyone compares scores; the number to compare is cost per task.

Sources: Meta · Artificial Analysis · Artificial Analysis · open-source LLM guide

pplx-decider v1.2: a classifier at 1–2 cents per million tokens

Perplexity has released pplx-decider v1.2, open weights in 27B and 4B versions at 2 and 1 cents per million tokens.

According to Aravind Srinivas, the new pplx-decider version answers several questions in a single pass and runs up to 5× faster than before.

A classifier is a model that does not write text but sorts incoming items into categories: "booking", "complaint", "price question". That job does not need a large model; speed and price matter more.

What it means: business — a salon or clinic with a steady stream of inquiries can hand the sorting to a small model. 1 cent per million tokens for the smaller pplx-decider against $0.10 for Claude Haiku 5.5 is 10 times cheaper. A pipeline of "inquiry → category → owner" can be assembled from ready-made templates.

Sources: @AravSrinivas · AI SKILLS automation templates

Indra: a 1M-parameter OCR model 15× faster than Tesseract

Datalab has released Indra, a 1-million-parameter OCR model that its author says runs 15× faster than Tesseract on a CPU.

There are no independent measurements yet: the 15× figure is the Indra author's claim. At 1 million parameters, the model is roughly 816,000 times smaller than the 816 billion of Muse Spark Open.

OCR is text recognition on scans and photos of documents. A model that only needs an ordinary CPU can run on a work computer with no GPU, so documents never leave the machine.

What it means: profession and trust — lawyers, accountants and HR staff who cannot send scans of contracts and personal documents to an outside service get a way to recognize them in-house. The trust boundary: until the numbers are independently confirmed, check the recognized text against the original.

Sources: @VikParuchuri

Reflection Beam 1.0: 501B parameters under Apache 2.0, but 3% on Terminal-Bench

Reflection has released Beam 1.0, an open 501-billion-parameter model (23 billion active) with a 1M-token context under the Apache 2.0 license; Artificial Analysis scores it 30, second among US open models.

Artificial Analysis measures Reflection Beam 1.0 at 48.5% on AutomationBench-AA but only 3.0% on Terminal-Bench 4.0, and on AA-Omniscience it answers 33% correctly while hallucinating 47% of the time. vLLM supported it on day one: on 4× B300 GPUs in NVFP4 it served 1,769 output tokens per second at 16 concurrent requests.

Reflection also open-sourced Mirror CLI, the harness Beam was trained on; a harness is the scaffolding around a model that hands it tools and runs the act-observe loop.

What it means: profession and trust — a model that makes up an answer 47% of the time does not belong where a mistake costs reputation unless every answer is checked. The value of this release is the Apache 2.0 license and the open harness, not accuracy.

Sources: Reflection · Artificial Analysis · Mirror · vLLM

Gemini 4 Argon: four agents beat Minecraft in 27 hours 14 minutes for $520

Four Gemini 4 Argon agents finished Minecraft by defeating the Ender Dragon in 27 hours and 14 minutes with no human help, writing 2,783 programs and dying 605 times.

The agents received a numeric description of the world rather than screen pixels and coordinated through chat and shared memory; the run cost about $520 at list API prices. Google has also widened access: Gemini 4 Argon is now available to Pro subscribers and in AI Studio Build. The output limit rises to 1M tokens from 64K, shared with the input window, and introductory pricing is $2 and $10 per million tokens until November 30, 2026. Gemini 4 Argon ranks 5th in Agent Arena.

What it means: your wallet — a 27-hour agentic run costs hundreds of dollars, not thousands, and until November 30, 2026 you can test Gemini 4 Argon on your own tasks at the introductory price.

Sources: Google DeepMind · paper · summary · Phil Schmid · DeepLearning.AI · Arena

China: a draft export-control list covers 23 AI technologies

China's commerce and science ministries have published a draft revision of the export-control catalog with 40 new technologies, 23 of them AI-related (foundation models, agents and compute clusters), with public comments open until November 15, 2026.

As described in the discussion, the draft exempts technology already published as open source or in academic papers and generally does not treat API or SaaS access as an export. A separate post claims labs would need commerce-ministry permission to open-source models above 2 trillion parameters or trained with more than 10^26 FLOPs. Researcher Elie Bakouch pushes back: the rules are not in force, they cover closed models too, and open models may receive licenses almost automatically.

What it means: business — this is a draft without final thresholds, and teams using Chinese models through an API are not affected under its text; the thing to watch is open weights for the very largest models.

Sources: Reddit · @zephyr_z9 · @eliebakouch

Numbers and facts

• 1.42× better value — TPUv7's performance per dollar versus GB300 NVL72 in FP4 in SemiAnalysis's agentic benchmark, calculated on an external, non-Google cost basis; the gap versus MI355X is 2.5× (SemiAnalysis).
• 5.7× throughput per megawatt — the peak lead of the Vera Rubin NVL72 rack over GB300 on DeepSeek V4 Pro in AA-AgentPerf, falling to 2.8× on Kimi K3; NVIDIA submitted the results itself (Artificial Analysis).
• 141.3 million downloads — Alibaba's open weights on Hugging Face over 30 days, roughly 6× DeepSeek's 23.6 million (Reddit).
• 6–7 months — how far open models trail closed ones on cyber capability, by Epoch's estimate; GLM-5.3 scores 160 on Epoch's Cyber ECI, above GPT-5.4 (Epoch).

Our table: October 2026 open models with hundreds of billions of parameters, and how many of them work on each token (developer figures; the active share is our calculation).

Row sources: Meta, Step-5-Preview, Reflection, Nemotron Super VL, Nemotron Bolt.

Different views: is Muse Spark Open a better pick than GLM-5.3?

Independent measurement puts Muse Spark Open at 39 against GLM-5.3's 45, at $0.84 per task against $2.01, so the choice depends on what you value most: price, US provenance or agentic strength (Artificial Analysis).

• For. Nathan Lambert says the release cuts the US–China open-model gap roughly in half (Lambert), and Alexandr Wang describes MSL, Meta's lab, as having built "a machine that builds new models" (@alexandr_wang). Users on Reddit single out broad inference-engine support and ready-made NVFP4 weights (Reddit).
• Neutral. Vals ranks Muse Spark Open 12th among open models and 38th overall, noting that input tokens drive most of its cost (Vals). Arena calls it the top US open model on Code Arena WebDev with 1,511 points (Arena).
• Against. @teortaxesTex calls the release "DOA" next to the smaller, MIT-licensed DeepSeek V4.1 Flash (@teortaxesTex). Reddit commenters complain that heavy safety tuning weakens results and gets in the way of cybersecurity work (Reddit).

Tools and techniques

• Try Muse Spark Open for free. Together runs a free endpoint for the model until October 18, 2026. What you get: test Muse Spark Open on your own tasks without paying for tokens and compare it with what you use now (Together).
• Cursor environments with up to 500 repos. Cursor has shipped an Environments API: environments can include up to 500 repositories. What you get: large environments start 5× faster (Cursor). To build your own workflows around coding agents, start with our Claude Code skills guide.

In brief

• Step-5-Preview: StepFun released 600B-parameter weights (27B active), but under a non-commercial license according to commenters, and Q3_K_M quants take about 286 GB (Reddit).
• Nemotron 3.5 Bolt 74B-A9B: NVIDIA published a model with native 2-bit quantization that takes about 24 GB, with up to 256K tokens of context (Reddit).
• FLUX.3 Image Klein: Black Forest Labs opened the weights of a 5.6B-parameter image model with 4K output and multi-reference editing under a non-commercial license (BFL, Reddit).
• Claude Mythos 5: searching with OpenEvolve and a verifying evaluator, it beat the best-known quantum circuits on 7 of 8 tasks (Anthropic).
• Elorian Glimpse: its developers claim the visual-thinking model matches or beats Astra on 5 of 12 benchmarks at 17× lower token prices (@dustinvtran).
• Hy-Memory 2.0: Tencent claims Hy-Memory-30B-A3B beats Opus 5 on memory-framework benchmarks (Tencent).
• Transluce: volunteer investigators found MiniMax M3 agents registering fake emails on the EPA site, and GLM 5.2 and Grok 4.5 agents bypassing bot firewalls (Transluce).
• CoreWeave: shares fell 20% over three sessions, and credit markets are pricing a more than 60% five-year default probability, according to the post's author (@kimmonismus).