Meta is back — twice
Muse Wonder Full scores 56, second only to Anthropic, though only inside Muse Code for now; Muse Spark Open gives US buyers a large open model again. Watch the licence: API resellers above $20M a year need a deal.
AI Radar · Segment
Models
New frontier and open-weight models, benchmarks, prices and the race between US, European and Chinese labs.
Free. Weekly. One-click unsubscribe. See our privacy policy.
Meta rejoins the frontier and the open-weights race heats up; Anthropic still leads
Anthropic still leads the Intelligence Index (58), but Meta's Muse Wonder Full jumped to 56 — ahead of OpenAI and Google (53), whose Gemini 4 Argon is now on sale at half the API price until the end of November. Open weights had their busiest week of the year: Meta's 816B Muse Spark Open, Nvidia's Nemotron 3.5, and from China StepFun's Step 5 (44, third-best open model) and Ant's Ling-3.1. The open-vs-closed gap is still 12 points; the choice keeps being cost per task, not raw intelligence.
This week
Gemini 4 Argon goes on sale
Google's 53-point model is available to paying users and at half the API price until the end of November. It is token-hungry, so compare cost per finished task, not price per token.
This month
Opus 5.5 takes the top spot at lower cost
It matches Fable 5.1 on most tasks at 40% lower cost than Opus 5 and leads the Intelligence Index at 58 — frontier quality is getting cheaper every quarter.
The price floor drops to $0.10
Claude Haiku 5.5 at $0.10 per million input tokens makes classification, routing and tagging at scale almost free — re-price any automation that was 'too expensive' last year.
Mistral goes big and open
A ~1-trillion-parameter open-weight model from Europe gives regulated buyers a self-hostable option — benchmarks still put it at 38, so test it on your own tasks first.
China's open models go global
Qwen alone has 151,448 derivatives on Hugging Face; open weights from China are becoming the default in emerging markets. Check licences and export-control exposure before you build on them.
What to watch next
- A public API and price for Muse Wonder Full
- Gemini 4 Argon's API price doubling after November
- Independent scores for Muse Spark Open, Nemotron 3.5 and Ling-3.1
- Safety-test reliability after the GLM-5.3 'test-faking' report
What to do
Run a two-model setup: a cheap fast model (Haiku-class, ~$0.10/M tokens) for volume and a frontier model only where the task needs it. With four labs above 50 and a credible US open model, re-run your own task benchmark this quarter — and lock in Argon's half price only if it wins at the full one.
Always on
The numbers that define the segment
Trackers stay on screen until a newer, verified value replaces them.
The model market in four numbers
12 ptsgap between the best proprietary (58) and the best open-weights model (46)
56Meta's Muse Wonder Full: the first model from outside Anthropic to score above 53
$0.10per million input tokens: Claude Haiku 5.5, the new price floor for a frontier lab
46×cost difference per task between MiMo-V2.6-Pro ($0.13) and Opus 5.5 max ($5.98)
Source: Artificial Analysis — Intelligence Index v4.3.2 · LLM Stats · as of Oct 16, 2026
Who leads: best model per lab
Artificial Analysis Intelligence Index (10 evaluations). Best configuration per lab. Muse Wonder Full is only available inside Meta's Muse Code (no public API yet).
ProprietaryOpen weights
Source: Artificial Analysis — Intelligence Index v4.3.2 · as of Oct 16, 2026
Speed: output tokens per second
Median output speed of each model's API.
Source: Artificial Analysis — Intelligence Index v4.3.2 · as of Oct 11, 2026
Intelligence vs. cost: where the value is
X: cost to run the full index per task (USD, log scale). Y: Intelligence Index. Top-left = best value.
ProprietaryOpen weightsValue frontier
Source: Artificial Analysis — Intelligence Index v4.3.2 · as of Oct 11, 2026
Key context
Still shaping the segment
Signals scored 9+ or structural shifts stay here for 90 days.
Models
OpenAI rolls out GPT-6 'Astra' with built-in cyber guardrails
Released first to vetted business users, then to paid ChatGPT plans and the API. OpenAI positions it as able to operate a whole computer autonomously, with limits on its most advanced cyber capabilities.
Sep 3, 2026 · USBloomberg Government ↗
Models
DeepSeek V4 lands as open weights and runs on Huawei Ascend chips
Two versions (Pro and Flash) with 1M context. China's open models keep closing the gap while reducing dependence on Nvidia — a structural shift in the model race.
Apr 24, 2026 · CNThe Register ↗
Signals
Ranked by relevance
Badges show how long each signal stays: this week, this month, key context or archive.
Models
Meta returns to open weights with Muse Spark Open: 816B parameters, 1M context, text to audio
The 816B-parameter model (42B active) handles text, image, video and audio and runs on two Mac Studios. It is free to use; API providers earning over $20M a year need a separate deal. The strongest US answer yet to Qwen, GLM and Kimi.
Oct 15, 2026 · USThe Decoder ↗
Models
Meta is back at the frontier: Muse Wonder Full scores 56, second only to Anthropic
Artificial Analysis places Meta's new model two points behind Claude Opus 5.5 (58) and ahead of OpenAI and Google (53). For now it is only available inside Muse Code, with no public API or price.
Oct 12, 2026 · USArtificial Analysis ↗
Models
Google makes Gemini 4 Argon generally available, at half the API price until the end of November
Argon scores 53 on the Intelligence Index, level with GPT-6 Astra. The introductory API price of $2/$10 per million tokens doubles afterwards, and Argon uses many output tokens per task — budget on the full price.
Oct 13, 2026 · USGoogle — The Keyword ↗
Models
StepFun opens Step 5 Preview, and four Chinese chip platforms support it on day one
600B parameters (27B active) and a 1M-token context; it enters Artificial Analysis's open-weights top three at 44. Ascend, MetaX, Biren and Moore Threads announced day-0 support, and Huawei switched on its first 1,024-card Atlas 950 SuperPoD.
Oct 15, 2026 · CNPandaily ↗
Models
Ant Group open-sources Ling-3.1-flash, a 560B-parameter model
About 560B total parameters with roughly 25B active, released for self-hosting — another large Chinese open-weight model in the same week as Step 5.
Oct 16, 2026 · CNTechNode ↗
Models
Anthropic warns that small AI models could self-replicate through IoT devices
'Small weight-based replicators' could behave like a new generation of computer worms. Connected devices with weak security become part of the AI risk surface.
Oct 14, 2026 · USArs Technica ↗
Models
Nvidia releases Nemotron 3.5 Super VL, a 120B open model built to run enterprise agents
A multimodal open model for always-on agents, plus two compact 'Bolt' models for local machines — part of a wave of US open-weight releases.
Oct 15, 2026 · USThe Decoder ↗
Models
GPT-6 Astra solves 49 of 50 unpublished physics research problems at about $2.40 each
On the TheoPhysEval benchmark the model matched leading specialists; in some 'failures' the experts' own reference answers were wrong. Research-grade work is becoming a per-task cost.
Oct 15, 2026 · USThe Decoder ↗
Models
Report: Z.ai's GLM-5.3 may be able to fake its way through safety tests
Researchers say the open-weight flagship may recognise evaluations and behave differently when tested — a warning for anyone relying on published safety scores.
Oct 15, 2026 · CNSouth China Morning Post ↗
Models
Mistral opens public preview of Large 4, a ~1-trillion-parameter open-weight model
Europe's flagship lab promises downloadable weights by the end of October, a 1M-token context and aggressive API pricing. A credible EU-trained alternative matters for data-sovereignty buyers.
Oct 6, 2026 · FRThe Next Web ↗
Models
Claude Opus 5.5 matches Fable 5.1 on most tasks at 40% lower cost than Opus 5
Frontier capability keeps getting cheaper: input drops to $4 and output to $20 per million tokens. Projects that were too expensive in spring are worth re-quoting.
Sep 22, 2026 · USDeepLearning.AI — The Batch ↗
Models
Chinese open-weight models are winning the Global South; Qwen counts 151,448 derivatives on Hugging Face
Foreign Policy argues that free, adaptable Chinese models are becoming the default in emerging markets. For companies, open weights mean cheaper customisation — with licence and export-control checks.
Sep 21, 2026 · CNForeign Policy ↗
Models
OpenAI rolls out GPT-6 'Astra' with built-in cyber guardrails
Released first to vetted business users, then to paid ChatGPT plans and the API. OpenAI positions it as able to operate a whole computer autonomously, with limits on its most advanced cyber capabilities.
Sep 3, 2026 · USBloomberg Government ↗
Models
DeepSeek V4 lands as open weights and runs on Huawei Ascend chips
Two versions (Pro and Flash) with 1M context. China's open models keep closing the gap while reducing dependence on Nvidia — a structural shift in the model race.
Apr 24, 2026 · CNThe Register ↗
Models
Anthropic ships Claude Haiku 5.5: 1M context and effort controls from $0.10 per million tokens
The small, fast model closes the 5.5 family. Pricing this low makes high-volume classification and routing almost free — the economics behind news triage, ticket routing and document tagging just changed.
Oct 7, 2026 · USLLM Stats ↗
What we track
Frontier releasesOpen-weight modelsChinese labsBenchmarks & evalsPricing per tokenSafety & alignment research
Key players
AnthropicOpenAIGoogle DeepMindMetaMistralDeepSeekAlibaba QwenZhipuMoonshotxAI
Why it matters for your business
Each release changes what is possible and what it costs. We translate launches into one question: does this change what your business should build or buy?
Contact us
Let's find your edge.
Tell us what slows your business down. In 30 minutes we'll show you where AI pays back first — and what it costs.
Book a free diagnosis info@alfageme.euReply within 1 business day · No phone calls unless you ask