Skip to content

Model radar

Choose a model based on the work, not the logo.

We supplement verified vendor data with editorial recommendations: which tasks to entrust to the model, when to choose a cheaper option, and when to combine multiple models.

There is no single best model. A more capable model may not be faster or more cost-effective. Compare use cases, pricing and availability. Check speed on your own task.

8 current profiles 4 combinations 4 methodologies

Choose by task ↓ Compare two models →

Start with your task

What do you need to do?

Choose the type of work. We will show models suited to it based on documented capabilities.

Practical selection

Currently tracked models

Starting with the latest documented release. Entries with the same date are sorted by name; unknown dates come last. Newer does not mean better.

Open comparison →
Data for the current selection checked The selection is neither a popularity ranking nor a complete catalog.

Online model prices are standard API rates per million tokens (pieces of text), excluding discounts for repeated input and batch processing. App subscriptions and tool usage are billed separately.

Anthropic More complex analysis and development

Claude Opus 5.5

Published

The current version of the Opus series for working with long prompts, text and code.

Usage and availability
Best used for
  • Programming and multi-step tasks
  • Analysis of long documents and images

Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry

Costs
$4 input / $20 output per million tokens in standard API mode.
Strengths and supporting evidence →
OpenAI Simple tasks at low cost

GPT-6 Luna

Published

A cost-effective option for clearly defined, repetitive tasks.

Usage and availability
Best used for
  • Text classification and brief summaries
  • Processing a large number of simple requests

OpenAI API

Costs
$0.10 input / $0.50 output per million tokens. Above 272 thousand input tokens: 2× input and 1.5× output.
Strengths and supporting evidence →
OpenAI Everyday work and programming

GPT-6 Sol

Published

A versatile choice for working with text, code and images.

Usage and availability
Best used for
  • Writing and fixing code
  • Processing documents and multi-step tasks

OpenAI API

Costs
$2 input / $10 output per million tokens. Above 272 thousand input tokens: 2× input and 1.5× output.
Strengths and supporting evidence →
xAI Code and working with information

Grok 4.7

Published

A model for coding and multistep tasks involving text and images.

Usage and availability
Best used for
  • Code development and fixes
  • Analysis of text and image materials

xAI API, Cursor and Grok Build

Costs
$2 input / $6 output per million tokens below 200 thousand input tokens; above this threshold, $4 / $12.
Strengths and supporting evidence →
OpenAI Complex tasks from brief to result

GPT-6 Astra

Published

A model designed for complex analysis, development and tool use.

Usage and availability
Best used for
  • Analysis of complex problems
  • Research, document creation and larger programming tasks

OpenAI API

Costs
$10 input / $50 output per million tokens. Above 272 thousand input tokens: 2× input and 1.5× output.
Strengths and supporting evidence →
Google Documents, media and automation

Gemini 3.8 Flash

Published

Processes text, images, audio and video; the output is text.

Usage and availability
Best used for
  • Analysis of documents, audio and video
  • Programming and multi-step business tasks

Gemini API and Google AI Studio

Costs
$0.75 input / $3.75 output per million tokens through December 31, 2026. From January 1, 2027: $1.50 / $7.50.
Strengths and supporting evidence →
Anthropic Long and demanding tasks

Claude Fable 5.1

Published

A variant for demanding work where longer processing times and higher costs are acceptable.

Usage and availability
Best used for
  • Complex tasks with multiple sequential steps
  • Working with extensive context and code

Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry

Costs
$10 input / $50 output per million tokens in standard API mode.
Strengths and supporting evidence →
BottleCap AI Downloadable model subject to license terms

ThinkingCap-Qwen3.8-27B

Release date unverified

A modification of Qwen3.8-27B focused on shorter internal task processing.

Usage and availability
Best used for
  • Testing the model on your own hardware
  • Token usage comparison with the default model

Hugging Face after accepting the access terms. PolyForm Small Business 1.0.0 license and additional permission for personal use.

Costs
Self-hosting costs. Commercial licensing under the terms of BottleCap AI; no uniform public price is listed.
Strengths and supporting evidence →

Other models tracked

News outside the editorial selection

Individual announcements from the last 30 days. Verified release dates come first; reports without a verified date follow. The report date is not the release date.

Google

Gemini 4 Hydrogen

Published

Price: Not substantiated by the source.

Availability: Generally available in GitHub Copilot for users of the Pro, Pro+, Max, Business and Enterprise plans; deployment is taking place through a gradual rollout, and administrators of Business and Enterprise plans can control access to the model via model policy in Copilot settings.

Specifications, measurements, and limitations
According to the sources, it is suitable for
  • agentic coding
  • tool use
  • multi-step workflows

According to Google, the model outperforms competing models from OpenAI and Anthropic in a range of benchmarks.

OpenAI

GPT-6.1 Terra

Published

Price: The price is $1 per million input tokens and $5 per million output tokens. Cached input costs $0.05 per million tokens.

Availability: The model is available through Responses API and Chat Completions API, as well as in ChatGPT Work and Codex for users on Plus, Pro, Enterprise and Edu plans.

Specifications, measurements, and limitations
Inputs
The model processes visual inputs and generates text.
Documented measurements
  • Coding Agent Index v2.0 48 bodů According to Artificial Analysis, the model scored 48 points in an evaluation of coding agents.
  • AA-Omniscience 38 % In this private benchmark, the model guessed the answer in 38 % of cases when it did not know it.
  • Terminal-Bench 4.0 50.3 % At the maximum reasoning level, the model solved 50.3 % of tasks.
  • Obrana proti nepřímým útokům prompt injection 98.8 % According to the system card, the model resisted 98.8 % of tested attacks involving instructions hidden in third-party content.
According to the sources, it is suitable for
  • Drafting documents.
  • Interpreting technical drawings.
  • Automated software operation and form filling.
Documented limits
  • For the most demanding programming tasks, OpenAI continues to recommend GPT-6.1 Sol and GPT-6 Astra.

The model has half the API input and output rates of GPT-6.1 Sol. According to Ted Sanders of OpenAI, it is probably better at technical communication than GPT-6 Astra and GPT-6.1 Sol and offers similar capabilities in visual tasks and computer use at higher speed.

Thinking Machines Lab

Inkling 2 Mini

open weights

Release date unverified · First news item November 20, 2026

Price: Not substantiated by the source.

Availability: The model is available for download on the Hugging Face platform (thinkingmachines/Inkling-2-Mini) with support in the transformers, llama.cpp, and MLX libraries for local deployment, and in SGLang for production deployment on data center hardware.

Specifications, measurements, and limitations
Specifications
approximately 26 billion parameters in total (Mixture-of-Experts), approximately 3 billion active per token
Context
up to 1 million tokens
Inputs
text and images as input, text as output
Documented measurements
According to the sources, it is suitable for
  • local deployment thanks to its compact size and context of up to 1 million tokens
  • coding tasks, where the model leads in the FrontierCode v2.0 benchmark
  • working with both text and images thanks to multimodal inputs
Documented limits
  • in the AutomationBench-AA v1.0.6 benchmark it lags behind the Qwen3.8-27B model
  • the research variant with a variable number of loop passes has been tested only to a limited extent

The company compares the model with the Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL models and, based on its own informal tests, considers it at least comparable to the Qwen 3.8 27B model.

RadixArk

DaisyRose-27B

open weights

Release date unverified · First news item November 18, 2026

Price: Not substantiated by the source.

Availability: The weights are available for download on Hugging Face; deployment is recommended using the SGLang tool with DSpark speculative decoding.

Specifications, measurements, and limitations
Context
256K tokens
Inputs
text and images as input, text as output
Licence
Apache 2.0
Documented measurements
  • Terminal Bench 4.1 pass@1 12,0 % (oproti 5,6 % u modelu Qwen3.8-27B) The model improves the success rate of completing terminal agentic tasks compared to the base model.
  • Terminal Bench 2.1 75,5 % (oproti 73,0 % u modelu Qwen3.8-27B) Shows improvement on an older version of the terminal agentic benchmark.
  • DeepSWE v1.1 46,2 % (oproti 42,2 % u modelu Qwen3.8-27B) Shows improvement in software engineering capabilities.
  • SWE Bench Pro v2 94,7 % (oproti 91,3 % u modelu Qwen3.8-27B) Shows improvement in solving software tasks on this benchmark.
  • HLE (text-only) 34,3 % (oproti 33,9 % u modelu Qwen3.8-27B) Shows a small improvement on a test of difficult knowledge questions.
According to the sources, it is suitable for
  • Long agentic tasks with sequences of up to 1 million tokens
  • Terminal and coding agentic tasks requiring multiple sequential steps
  • Tasks requiring precise instruction-following

According to the source, the DaisyRose-27B model achieves a higher score than the base model Qwen3.8-27B on all the listed benchmarks.

NVIDIA

Nemotron 3.5 Bolt E4B 4GB

Release date unverified · First news item November 17, 2026

Price: Not substantiated by the source.

Specifications, measurements, and limitations
Documented measurements
According to the sources, it is suitable for
  • Local voice and device assistants
  • Interactive virtual characters
  • Embedded and edge interfaces on GeForce RTX, Jetson, and RTX/DGX Spark hardware

The model is compared with the Gemma-4-E4B-it and Qwen3.5-4B models, against which, according to the source, it achieves the highest average accuracy on 21 benchmarks while generating the smallest number of output tokens, both in reasoning mode and without it.

Moonshot AI

Kimi K3.1 Pro Highspeed

open weights

Release date unverified · First news item November 11, 2026

Price: Not substantiated by the source.

Availability: Available on the Amazon Bedrock platform via the Chat Completions, Responses, Converse, and Invoke API as well as via Playground, with profiles for the global, US, and Indian regions, and via integration with the OpenCode tool.

Specifications, measurements, and limitations
Specifications
3T-class architecture, the same as in the Kimi K3 model.
Context
1 million tokens
Inputs
Text and image input, text output; supports tool calling, structured output, and streaming.
Documented measurements
According to the sources, it is suitable for
  • Long-running agentic tasks
  • Financial modeling and work with office documents
  • Programming for game engines, 3D scenes, and CAD scripts

A faster follow-up version of Moonshot AI's Kimi K3.1 Pro model, focused on long-running agentic tasks and reducing factual errors.

Thinking Machines

Inkling 2 Small

Release date unverified · First news item November 6, 2026

Price: $1.18 per task (according to Artificial Analysis)

Availability: Offered as a proprietary model via the Tinker platform; public weights under the Apache 2.0 license are to be released later, after external safety testing is completed

Specifications, measurements, and limitations
Specifications
594 billion parameters total, 17 billion active
Inputs
text and images as input
Documented measurements
  • Artificial Analysis Intelligence Index 31 The score is 18 points higher than its predecessor, Inkling Small, and corresponds roughly to the level of Google's Gemini 3.8 Flash model.
  • AutomationBench-AA 46 % The score on agentic tasks rose from 5% for the predecessor to 46%.
  • SWEBench Pro v2 (Hard) 86,3 % The figure shows the model's performance on the hard version of the SWE-bench coding test.
  • FrontierCode 2.0 25,4 % On this coding test, according to the source, DeepSeek-V4.1-Flash remains in the lead with 28.5%.
  • GDP.pdf 25,2 % This is the highest score among the compared models on the test of reading and understanding complex PDF documents.
According to the sources, it is suitable for
  • agentic tasks with multi-step workflow completion
  • long-horizon coding tasks trained using reinforcement learning
  • visual tasks such as reading PDF documents and designing websites
Documented limits
  • according to the company, the model sometimes communicates inefficiently and does not choose the best tools for large codebases
  • lacks audio support, and the model is less optimized for chat
  • the model may overreact to health-related topics and produce cluttered output

The source compares the model with its predecessor, Inkling Small, with Google's Gemini 3.8 Flash in terms of overall score, and on price with Gemini 3.8 Flash and Claude Sonnet 5.5; according to the source, DeepSeek-V4.1-Flash remains in the lead on most coding tests.

Mistral AI

Mistral Large 4

open weights

Release date unverified · First news item October 29, 2026

Price: Not substantiated by the source.

Availability: The model weights have been released for download, and the model can be deployed independently (self-hosted), or through the Mistral Studio platform in the cloud, a private VPC, or on-premises.

Specifications, measurements, and limitations
Specifications
1 trillion parameters
Inputs
natively multimodal
Documented measurements
  • FrontierCode 2.0 19,7 % Cognition's benchmark assesses whether a repository maintainer would actually merge a change created by the agent.
  • Terminal-Bench 4.0 41 % This represents a 13 percentage point improvement over the public preview version of the model on terminal tasks, putting it, according to the source, in a leading position among open-weight models.
  • Artificial Analysis Coding Agent Index 48 % According to the source, the result is comparable to the GLM-5.3 model among open-weight models, but with 47 percent fewer tokens per task.
  • Artificial Analysis Cyber Index 51 % The index measures the model's ability to find, reproduce, and fix vulnerabilities in code without breaking existing functionality.
  • Průměrná míra odmítnutí u škodlivých kybernetických požadavků 96,6 % According to the source, this is the highest measured refusal rate for malicious requests among the open-weight models the company measures.
According to the sources, it is suitable for
  • agentic and terminal coding tasks with lower token consumption
  • defensive cybersecurity, i.e., finding, reproducing, and fixing vulnerabilities
  • enterprise tasks involving creating and editing spreadsheets and financial workflows
Documented limits
  • the source cites the risk of capabilities being potentially misused for real-world attacks, which is why the company continues to adjust its guardrails
  • the specialized variant for penetration testing and red-teaming is available only to vetted, selected partners, not publicly

The source compares the model with other open-weight models, specifically GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, in coding and in Surge Code's blind pairwise evaluation, where, according to the source, the model ranked behind Claude Opus 5.5.

We do not display missing or corrupted data. Older announcements without verified identity remain in the news archive. Measurements in announcements are not original tests conducted by AI Radar.

Older profiles · 10

Historical overview from 31. 7. 2026. Prices, availability and recommendations here have not been verified as current. This is not a list of discontinued models.

How to combine multiple models

One model does not have to do everything.

The following workflows are editorial suggestions to try. A low-cost model can sort simple tasks, a stronger one can solve difficult cases, and fixed rules can check the result.

01

Many tasks without unnecessary costs

Sorting and processing hundreds to thousands of similar inputs.

A low-cost model handles routine work, and the more expensive one gets only the truly difficult cases.

  1. 1. Sorting

    It identifies the task type, finds the necessary information and estimates its confidence.

  2. 2. Difficult cases

    When the first model is not confident enough, a stronger model takes over.

  3. 3. Checking

    Fixed rules in the program and a human check both the format and a sample of the results.

Caution: Use your own data to verify when a stronger model should take over the work. A model cannot reliably assess its own answer.

02

Quick code edits with a second check

Modifying a program in use where an error could cause harm.

One system makes the change, another looks for blind spots, and tests decide.

  1. 1. Editing

    Reviews the code, proposes the smallest possible change and runs the necessary tests.

  2. 2. Independent review

    It receives the task, the changes made and the test results, but not the original rationale.

  3. 3. Decision

    Automated checks, tests and human approval take precedence over model opinions.

Caution: A second model will not help if both receive the same incorrect assumption without evidence.

03

Documents → analysis → verification

Research, strategy or decision support based on a large number of documents.

A fast model prepares the material, and a stronger model focuses on analyzing it.

  1. 1. Preparing materials

    Finds important claims in documents and media and adds links to their sources.

  2. 2. Analysis

    It builds arguments, identifies contradictions and formulates recommendations from traceable source material.

  3. 3. Verification

    It looks for contradictions, but facts are ultimately checked against the original documents.

Caution: Agreement between two models is not proof; the cited inputs are what matters.

04

Sensitive data under your own control

Sensitive materials cannot be sent in full to an external online service.

The sensitive part stays on your own server, and the online service receives only the necessary data.

  1. 1. Processing on my own server

    A local model can help find facts. Verify the removal of sensitive data with a separate check before sending.

  2. 2. Optional analysis

    The online model receives only text without sensitive data and only the information it needs.

  3. 3. Return and review

    The result is combined with internal data only on your own server.

Caution: First, check the license for your intended use. Being able to run a model yourself does not in itself guarantee privacy. Server security and what data is stored also matter.

How to read the supporting material

Manufacturer data and firsthand experience carry different weight.

The current profiles are based on the linked documentation from the providers. We do not yet have our own comparative measurements of these versions. The following websites are additional sources for verification, not evidence for the ratings of all the models above. Some tests also evaluate the model together with the tools it uses.

How AI Radar works →
What we are not measuring yet

Public tests will not tell you how a model will handle your own texts, data and workflows. Before deciding, try it on a few real tasks. Prices are often quoted per token, meaning small pieces of text.