AI research becomes verifiable in the physical world.
Google DeepMind
Prices, context and performance of 360 language models from 52 providers, plus 178 models for images, video and audio. With a Fit Score for what you need them for.
Pick your use case and budget. Every model we can score is ranked by our Fit Score.
Use case
Writing, planning, brainstorming
Priority
Zero retention: the provider does not store prompts or responses.
Costs estimated for 1M input and 250K output tokens per month, API base rates.
Recommendations
196 matching models
The Fit Score factors in price and budget. For pure performance without price, see the leaderboard. Click a model name to see how its score is built. Method · Calculated from stored data.
Switch check
Is there a cheaper alternative to your current model with a similar performance index?
Pick a task: text, code, websites, images, video or speech. You see which models lead there right now, ranked by performance only. Prices are shown alongside but never change the rank.
Show
Text & reasoning
Building & design
Image, video & audio
Top 10: Overall
Index points, not a percentage. A gap of one or two points is hardly noticeable in everyday use.
One variant per model, free variants merged. Monthly cost for your token volume from the model finder. Data sources in the Method section.
AI models read, write and bill in tokens, small pieces of text. Here is how much that is, and how far your budget goes.
1M tokens ≈ 750,000 words ≈ 3,000 pages ≈ 10 novels
Claude Sonnet 5.5: $2.00 per 1M input tokens, $10.00 per 1M output tokens. Preset to the top recommendation from the finder.
Rule of thumb for English text: 1 token ≈ 0.8 words, 1 page = 250 words, 1 novel = 300 pages, 1 line of code ≈ 10 tokens. Actual numbers vary by model, language and content; costs are base rates without caching, tiered pricing or taxes.
The strongest model from each of 5 providers is only 11.2 points apart. Your task decides which one fits, not the ranking.
The full AgentKernel catalog, strongest models first. The Fit column also weighs price and budget for your “Everyday” profile.
Updating catalog … · Updated Oct 14, 2026, 15:00 UTC
1–15 of 360 models
Prices in USD per 1M tokens, base rate. * Tiered pricing possible. “open”: data missing, “–”: does not meet your requirements. Neither is a bad rating.
10 models are within ten points of the top of the performance index. At 1M input and 250,000 output tokens they cost between $2.313 and $22.50 per month.
Releases, research and reporting from provider blogs and independent trade media, original headlines linking to the publisher.
Fetching sources …
Google DeepMind
OpenAI
Anthropic
Sources: OpenAI, Google DeepMind, Google Research, Microsoft Research, Apple Machine Learning, Hugging Face, NVIDIA, AWS Machine Learning, MIT Technology Review, Ars Technica, TechCrunch, The Verge, Wired, Simon Willison, and The Decoder. Topics are derived from keywords in the headline. Announcements by providers are not independent tests.
115 models from 25 providers accept at least 1M tokens in a single request. That is about 3,000 pages of text or 10 novels at once, enough for whole codebases or case files.
Tokens, costs and context in plain words. The figures come from the current AgentKernel catalog.
A token is a small piece of text, usually part of a word. AI models read, write and bill in tokens. As a rule of thumb, one token is about 0.75 English words: 1,000 tokens are roughly 750 words or three pages.
One million tokens are about 750,000 words, roughly 3,000 pages or 10 novels. That covers, for example, about 1,300 detailed chat replies, 1,500 email drafts or around 260 summaries of ten-page documents.
At 1M input and 250,000 output tokens per month, the 211 rated paid models in the AgentKernel catalog cost between $0.037 and $300.00, with a median of $0.90 (API base rates, as of Oct 14, 2026). Chat subscriptions are billed differently; free variants usually come with usage limits.
It depends on the task and the budget. The strongest model from each of 5 providers is only 11.2 points apart on the AgentKernel performance index. The model finder weighs performance, cost, context and features for your use case instead of crowning a winner.
118 models in the AgentKernel catalog have a coding index. The strongest coding model from each of 5 providers is 5.4 points apart, so price and context often decide more than rank. In the model finder, the Coding profile averages the general performance index with the coding index.
The AgentKernel catalog lists 57 image models. In head-to-head user votes, GPT Image 2.5 Sunburst (OpenAI), GPT Image 2.5 Flare (OpenAI), GPT Image 2 (medium) (OpenAI) currently lead. Which one fits depends on style, text in the image and price; the leaderboard also ranks image editing and graphics & logos separately.
30 video models are in the catalog, 17 of them with generated sound, clips up to 30 seconds and prices from $0.02 per second of video. In head-to-head user votes for video from text, Gemini Omni 1.1 Flash (Google), Gemini Omni Flash (Google), FLUX 3 Video (Black Forest Labs) currently lead. The leaderboard ranks video from text, from an image and video editing separately and shows which of them you can book via API.
In blind head-to-head votes, Inworld TTS MAX (Inworld), CastleFlow v1.0 (Async), Aurora (–) currently sound the most natural. The leaderboard shows the full ranking with vote counts and uncertainty, plus which models have open weights.
260 language models in the catalog are offered by at least one provider with zero data retention, meaning prompts and responses are not stored; 57 of them also in an EU region. 150 models have open weights and can be self-hosted. The model finder can filter for both; this is provider information, not legal advice.
The context window is the largest amount of text a model can handle in a single request, measured in tokens: your question, attached documents, the conversation history and the answer together. 115 models in the catalog handle at least 1M tokens, roughly 3,000 pages at once.
Input tokens are the text you send to the model, output tokens the text it writes. Output is usually much more expensive: in the AgentKernel catalog the median output price is 4 times the input price. Long documents with short answers are therefore often cheaper than you might expect.
The AgentKernel catalog of 360 models from 52 providers is synced every hour and the news every 15 minutes. This page reflects data as of Oct 14, 2026. Prices are the listed base rates in US dollars per million tokens.
The Fit Score combines four components from 0 to 100: overall performance, budget headroom, context headroom and matching features. The weighting depends on your use case and priority; the full formula is in the Method section.
Editorial summaries by AgentKernel as of 2 October 2026. Sources link to the original publications.
OpenAI presents Sol for demanding coding and knowledge work. Anthropic aims Sonnet at well-defined everyday tasks. The common trend: more specialisation within a model family.
Topic: Models & costs
Our take: compare quality and token prices separately. Then test on your own tasks whether the seemingly cheaper candidate also needs fewer attempts.
What matters is whether a system reliably finishes longer, multi-step tasks. METR studies this with tasks whose difficulty is expressed as the time a human needs for them.
Status: Measurable progress
What to watch next
Success rates on longer tasks, error recovery and the number of human interventions needed.
A 50% time horizon means an estimated 50% chance of success. It is not a guaranteed autonomous runtime and not a measure of how automatable every job is.
METR: Task-Completion Time HorizonsHow the Fit Score (version 1.1) is built: four components from 0 to 100, weighted by use case and priority.
Data sources: catalog, prices and capabilities from the OpenRouter Models API; performance, coding and agentic index based on the Artificial Analysis indices of the same name; design & build, image and speech placements from Design Arena; knowledge cutoff as stated by the provider; zero-data-retention hosting, EU regions and availability from OpenRouter’s endpoint list; open weights when a model is published on Hugging Face; video facts from the OpenRouter video API. We check and normalise this data but do not run the benchmarks ourselves. The Fit Score is guidance, not a task-specific test. Try your favourites on your own tasks.
Rankings for images, video and speech: LMArena Leaderboard Dataset (create and edit images, video from text, from image and editing; licence CC BY 4.0, reduced to the overall ranking, names normalised); CC BY 4.0; TTS Arena V2 (text to speech).