AI Radar
A weekly overview of the AI models available: which one to use for
which job, what it costs, and what is new.
This week, developments in the artificial intelligence space highlight new approaches to reasoning and multi-user agent access control. Operators should note that LLMs are increasingly being applied to scale formal software verification, while confidence scoring using language models continues to face skepticism. Keep an eye on these workflow integrations and agent frameworks as you evaluate your local deployment and security strategies.
issue 2026-W43 · generated 2026-10-23
· 4 new models
The short answer #quick
One pick per job: the best value model we found that is also popular
enough to be worth trusting.
| Job | Model | Price | Why |
|---|
| Cheap & small (task routing) |
Xiaomi: MiMo-V2.6-Flash
xiaomi/mimo-v2.6-flash
|
$0.168/1M
|
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blen…
|
| Medium (real work & coding) |
Qwen: Qwen3.7 Plus
qwen/qwen3.7-plus
|
$0.512/1M
|
qwen/qwen3.7-plus is a cost-effective option in the medium tier, priced at $0.512 per 1M blended tokens with an arena rating of 1455. It is a solid ch…
|
| Expensive (architecture) |
OpenAI: GPT-5.4
openai/gpt-5.4
|
$5.00/1M
|
OpenAI: GPT-5.4 is a top-tier frontier architecture that unifies Codex and GPT systems, delivering a massive 1.1M context window and an arena rating o…
|
| Coding & SWE agents |
Xiaomi: MiMo-V2.6-Flash
xiaomi/mimo-v2.6-flash
|
$0.168/1M
|
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blen…
|
| Out of the box (native tools) |
OpenAI: GPT-5.6 Luna
openai/gpt-5.6-luna
|
$0.400/1M
|
openai/gpt-5.6-luna is a fast and cost-efficient option at $0.400/1M blended tokens, making it well-suited for high-volume, latency-sensitive homelab …
|
| Voice & audio |
Google: Lyria 3 Pro Preview
google/lyria-3-pro-preview
|
$0.08 per song
|
google/lyria-3-pro-preview is part of Google's Lyria 3 family of music generation models available through the Gemini API and is well suited for gener…
|
| Open weights (self-hostable) |
Xiaomi: MiMo-V2.6-Flash
xiaomi/mimo-v2.6-flash
|
$0.168/1M
|
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blen…
|
| Image & video |
Google: Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image)
google/gemini-3.1-flash-lite-image
|
$0.500/1M
|
google/gemini-3.1-flash-lite-image is Google's fastest and most cost-efficient image model, priced at $0.500 per 1M blended tokens, making it well-sui…
|
Cheap & small (task routing) #router
Cheap enough to call on every step. Routing, classification, extraction, summarising.
#1
$0.168/1M · $0.00196 per task
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blended tokens. With an arena rating of 1452, it delivers solid performance for general self-hosted workloads, though its lower popularity score and specialized architecture mean it might not be worth the operational overhead if you do not strictly require its extensive context.
cheapopen weightslong context10 providers
#2
$0.220/1M · $0.00250 per task
z-ai/glm-5.3-flash stands out as an efficient open-weights choice for coding and long-horizon agent tasks, backed by a high arena rating of 1472 and a 1.0M context window. Priced reasonably at $0.220/1M blended tokens and supported by 36 providers, it represents great value for homelab automation unless your specific use case avoids long-context agentic workflows.
cheapopen weightslong context36 providers
#3
$0.240/1M · $0.00280 per task
google/gemma-4-31b-it offers a dense multimodal open-weights option with a configurable thinking mode and native function calling for $0.240/1M blended tokens. While its 262K context window and 1425 arena rating make it a competent self-hosted assistant for image and text tasks, it is not worth the money if you need the million-token scale provided by other candidates.
cheapopen weights16 providers
Medium (real work & coding) #workhorse
The models that do the actual work: writing, refactoring, debugging, agentic loops.
#1
$0.512/1M · $0.00576 per task
qwen/qwen3.7-plus is a cost-effective option in the medium tier, priced at $0.512 per 1M blended tokens with an arena rating of 1455. It is a solid choice for handling routine coding and real work tasks with its large 1M context window. However, with a popularity score of 40, it may not be worth the money if you require the absolute highest reasoning capabilities available in this category.
mid-pricelong context
#2
$0.522/1M · $0.00609 per task
xiaomi/mimo-v2.6-pro stands out as a strong performer in this category, offering a high arena rating of 1491 for $0.522 per 1M blended tokens. It is well-suited for demanding real work and coding tasks backed by an expansive 1.1M context window across five providers. Despite a modest popularity score of 37, its strong benchmark performance makes it a compelling value for homelab operators needing advanced capabilities.
mid-priceopen weightslong context5 providers
#3
$0.500/1M · $0.00550 per task
google/gemini-3.1-flash-lite-image is Google's fastest and most cost-efficient image model, priced at $0.500 per 1M blended tokens, making it well-suited for high-velocity developer pipelines and rapid-fire visual exploration. With an arena rating of 1313, it is not worth the money if you require top-tier rendering quality for complex, high-fidelity visual production.
mid-priceimage out
Expensive (architecture) #architect
When the answer has to be right and the reasoning has to hold up. Use sparingly.
#1
$5.00/1M · $0.0550 per task
OpenAI: GPT-5.4 is a top-tier frontier architecture that unifies Codex and GPT systems, delivering a massive 1.1M context window and an arena rating of 1469 for $5.00/1M blended tokens. It is an exceptional choice for homelab operators needing deep context handling and complex multi-domain reasoning. However, it may not be worth the cost for casual tasks where simpler models suffice.
premiumopen weightslong context7 providersapply patchweb search
#2
$5.40/1M · $0.0600 per task
Anthropic: Claude Sonnet 4.6 is a high-capability model boasting an arena rating of 1458 and a 1M token context for $5.40/1M blended tokens. It excels at iterative software development, agentic workflows, and navigating complex codebases in demanding homelab environments. It is not worth the money if you are running lightweight queries that do not leverage its advanced coding and project management strengths.
premiumopen weightslong context9 providersadvisorbashweb search
#3
$5.40/1M · $0.0600 per task
Anthropic: Claude Sonnet 4.5 delivers state-of-the-art coding benchmark performance and real-world agent optimization for $5.40/1M blended tokens with a 1M context. It shines in specialized programming and end-to-end development tasks for your local infrastructure. Given that newer iterations exist at similar price points, it may not be worth the investment compared to its direct successor.
premiumopen weightslong context7 providersadvisorbashweb search
Models that came with an agent harness bolted on. Point them at a repo and walk away.
#1
$0.168/1M · $0.00196 per task
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blended tokens. With an arena rating of 1452, it delivers solid performance for general self-hosted workloads, though its lower popularity score and specialized architecture mean it might not be worth the operational overhead if you do not strictly require its extensive context.
cheapopen weightslong context10 providers
#2
$0.220/1M · $0.00250 per task
z-ai/glm-5.3-flash stands out as an efficient open-weights choice for coding and long-horizon agent tasks, backed by a high arena rating of 1472 and a 1.0M context window. Priced reasonably at $0.220/1M blended tokens and supported by 36 providers, it represents great value for homelab automation unless your specific use case avoids long-context agentic workflows.
cheapopen weightslong context36 providers
#3
$0.343/1M · $0.00359 per task
Z.ai: GLM 5.3 achieves the highest arena rating in this lineup at 1474 and is specifically engineered for complex software engineering and long-horizon agent tasks. At $0.343 per 1M blended tokens with a 1.0M context window, it is a capable choice for demanding development workloads, though it may be an unnecessary expense for simple scripting or light automation.
cheapopen weightslong context48 providers
Speech in, speech out, plus the music and clip generators that bill per song.
#1
$0.08 per song
google/lyria-3-pro-preview is part of Google's Lyria 3 family of music generation models available through the Gemini API and is well suited for generating full-length songs at a cost of $0.08 per song. Since it is currently unrated in the arena, homelab operators should test it carefully before relying on it for critical audio pipelines. It is not worth the money if your workflow only requires short musical snippets rather than complete tracks.
cheaplong contextaudio out
#2
$0.04 per clip
google/lyria-3-clip-preview is designed for generating shorter audio segments through the Gemini API, priced at $0.04 per 30-second duration clip. This model is good for quickly prototyping short musical cues or background loops without paying for full-length song generation. However, it is not worth the money if you need complete, continuous musical compositions instead of brief 30-second clips.
cheaplong contextaudio out
#3
$4.00/1M · $0.0450 per task
openai/gpt-audio is OpenAI's first generally available audio model, featuring an upgraded decoder for more natural-sounding voices and improved voice consistency across a 128K context window. It is good for applications requiring realistic speech synthesis and conversational audio interactions. At $4.00 per 1M blended tokens, it is not worth the money if your homelab tasks only require basic text processing without audio capabilities.
mid-priceopen weightsaudio out
Open weights (self-hostable) #open-weights
Weights you can download and run on your own box. Costs you GPU time, not tokens.
#1
$0.168/1M · $0.00196 per task
xiaomi/mimo-v2.6-flash is a strong open-weights option for homelab operators seeking a massive context window of 1.1M at a low price of $0.168/1M blended tokens. With an arena rating of 1452, it delivers solid performance for general self-hosted workloads, though its lower popularity score and specialized architecture mean it might not be worth the operational overhead if you do not strictly require its extensive context.
cheapopen weightslong context10 providers
#2
$0.220/1M · $0.00250 per task
z-ai/glm-5.3-flash stands out as an efficient open-weights choice for coding and long-horizon agent tasks, backed by a high arena rating of 1472 and a 1.0M context window. Priced reasonably at $0.220/1M blended tokens and supported by 36 providers, it represents great value for homelab automation unless your specific use case avoids long-context agentic workflows.
cheapopen weightslong context36 providers
#3
$0.240/1M · $0.00280 per task
google/gemma-4-31b-it offers a dense multimodal open-weights option with a configurable thinking mode and native function calling for $0.240/1M blended tokens. While its 262K context window and 1425 arena rating make it a competent self-hosted assistant for image and text tasks, it is not worth the money if you need the million-token scale provided by other candidates.
cheapopen weights16 providers
Priced per generation, not per token. A different kind of budget.
#1
$0.500/1M · $0.00550 per task
google/gemini-3.1-flash-lite-image is Google's fastest and most cost-efficient image model, priced at $0.500 per 1M blended tokens, making it well-suited for high-velocity developer pipelines and rapid-fire visual exploration. With an arena rating of 1313, it is not worth the money if you require top-tier rendering quality for complex, high-fidelity visual production.
mid-priceimage out
#2
$2.70/1M · $0.0300 per task
google/gemini-nano-banana-2.1 provides improved product recontextualization on the Flash tier for $2.70 per 1M blended tokens, backed by an arena rating of 1426. It is a solid choice for mid-tier image generation and editing tasks, but it may not be worth the money if you prefer the lowest cost pipeline or the absolute highest multimodal reasoning performance.
mid-priceweb searchimage out
#3
$9.40/1M · $0.1100 per task
openai/gpt-5.4-image-2 combines GPT-5.4 reasoning with state-of-the-art image generation from GPT Image 2, achieving the highest arena rating of 1461 in this lineup. At $9.40 per 1M blended tokens, it is ideal for rich multimodal workflows involving reasoning, coding, and image generation, but it is definitely not worth the money for simple, low-budget generation tasks.
premiumopen weightsweb searchimage out
New this week #new
Models that appeared on OpenRouter since the last issue.
Price changes since the last issue, biggest move first.
No price moved more than 5%.
What people are talking about #news
Community and vendor chatter from the last seven days.
Today's papers from the HuggingFace daily feed.
-
Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization
Effective test-time personalization requires LLM agents to adapt to query-dependent user preferences from lightweight interaction.
-
IF-MAP: Influence-Function-Inspired Agent Prioritization in Multi-Agent Systems
Improving every agent in an LLM-based multi-agent system can be prohibitively costly, making it important to identify which agent should be prioritized to achieve the greatest system-level gain with limited resources.
-
Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems
Modern LLM agents can retrieve information from accessible documents, and then disclose a sensitive conclusion assembled from them.
-
HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second
We present HyperDrone, an extremely fast quadrotor simulator focused on high-performance batch rendering that is more than 20x faster than existing state-of-the-art batch quadrotor simulators.
-
BrickWright: Agentic Creation of Physically Feasible Brick Builds
Frontier coding agents are becoming increasingly adept at 3D modeling.
-
Who Takes the Lead? Model Ability and Task-Conditioned Collaboration with Humans
As language-model agents grow more capable, they can carry more of a task without asking anyone for help.
-
FilmFlow: Agentic System for 3D Virtual Film Cinematography
Complex camera motion is central to visual storytelling, yet translating cinematographic intent into controllable, executable camera trajectories remains challenging.
-
RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention
Autonomous research agents are increasingly used to automate large portions of scientific research, yet they are prone to research integrity violations (RIVs): behaviors that compromise the validity, reproducibility, or faithfulness of the research process.
-
Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective
Event sequences are chains of random events, each described by heterogeneous (numerical and categorical) features.
-
MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models
Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference cost remains a major bottleneck for real-time control.
How this was made #method
Sources
- OpenRouter
- LMArena
- Artificial Analysis
- HuggingFace
- Hacker News
- Reddit
- Vendor blogs
Price tiers. cheap under $0.50/1M blended tokens, medium $0.50-$5.00, expensive above. Blended is 80% input, 20% output, priced on the cheapest provider.
Value. value = quality / cost of a 10000 token in + 2000 token out task, with a $0.00 cost floor so free models score high without dividing by zero.
Popularity. A blend of arena votes, how many providers serve the model,
HuggingFace trending and how often the model is named in the week's chatter.
The prose is written by google/gemini-3.5-flash-lite, running on the numbers above it. It is told to describe only the models and prices in its input; blocks that mention an unknown model are dropped before publishing. Without an API key the report falls back to templated text and every number still stands.