See which model wins at each price.
Benchmark scores plotted against cost and tokens per task, plus a local collector for your own agents' token use. Compare models on the work your agents actually do.
Capability and cost
Higher score, lower cost
108 options
Method & data
What the axes mean. Output tokens are answer plus reasoning generated per Intelligence Index task; they exclude input and cache traffic. Cost is the publisher’s estimated total task cost. The horizontal axes are logarithmic, so equal spacing represents equal proportional change.
Cohort and frontier. The source has 697 records. 113 are current configurations with a non-estimated score, output-token count, and cost breakdown, and 108 of those report a task cost above zero. The curve connects the configurations with the highest score at each resource budget; aicharts draws it from this cohort. No benchmark families are blended.
Index construction. The publisher’s 10-evaluation index weights agents 30% · coding 20% · scientific 20% · general 30%. These model-level results are separate from the coding-agent chart and its total-token counts.
Source. Publisher methodology · Source terms. Measurements come from the publisher’s public models leaderboard; the snapshot date is a retrieval date, not a model’s evaluation date.
About aicharts
Picking a model means guessing at a tradeoff between quality and price, because benchmark scores and prices live in different places. aicharts plots published benchmark scores against cost and tokens per task on one chart, so the strongest option at each budget is visible instead of implied. A local collector measures your own agents' token use, so the cost question covers your work, not only the models. aicharts is free and open source.
Other places to compare models: Artificial Analysis runs the Intelligence Index and Coding Agent Index charted here, and its own leaderboards cover more than 250 models with speed and latency. Arena (formerly LMArena) ranks models by people's votes in blind side-by-side chats. Epoch AI runs its own benchmarks and tracks how capabilities change over time. OpenRouter lists live API prices and which models its users send the most tokens to. aicharts puts published scores against cost per task, draws the best score at each budget, and keeps each benchmark on its own scale. Checked Sep 28, 2026.
Measure your own agents
The aicharts collector adds up the tokens, cost and time your coding agents use, per model and day. Prompts and transcripts stay on your machine. In development. How it works
Recent Intelligence Index listings
Intelligence Index v4.3.2. Card dates come from the Artificial Analysis listing and are not verified first-party release dates.
Anthropic Claude Haiku 5.5
Artificial Analysis Intelligence Index v4.3.2
Index43.4
Mistral Mistral Large 4 Preview
Artificial Analysis Intelligence Index v4.3.2
Index38.4
InclusionAI Ling 3.1 Flash
Artificial Analysis Intelligence Index v4.3.2
Index41.1
Upstage Solar Mini 4
Artificial Analysis Intelligence Index v4.3.2
Index24.1
First-party release radar
New releases found at first-party sources
24 labs · 30 first-party sourcesThese links come from each lab's own release pages, so a model can appear here before aggregators such as OpenRouter list it. The date shows when aicharts first found the page, which can differ from the official release date. Official dates and benchmark scores come from their own sources.
Release radar
New, awaiting complete benchmark coverage
27 incomplete · 5 with early DeepSWE · OpenRouter checked- Muse WonderMeta · first observed
- GLM 5.3 FlashZ.ai · first observed Early DeepSWE 63.4% pass@1 · max · 4 runs · OpenRouter match
Models listed here do not yet have a complete result on the Artificial Analysis coding-agent chart. Where shown, an early DeepSWE v1.1 pass@1 comes straight from DataCurve's mini-swe-agent leaderboard, matched to the model through OpenRouter's catalog, or through Artificial Analysis when OpenRouter has no match. That result depends on DataCurve's harness, so it stays off the Artificial Analysis chart and model cards. Models with partial Artificial Analysis results can appear there, with missing metrics marked.
Recent models and notes
Compact Index listings and published analysis. Model scores come from Artificial Analysis Intelligence Index v4.3.2.
- NoteHaiku 5.5 scores 43.4 at $0.213, one row off the cost frontierClaude Haiku 5.5 (max) scores 43.4 on the Intelligence Index at $0.213 per task, 32nd of 108 configurations. One cheaper row keeps the max row off the cost frontier; the xhigh row is a vertex of it.
- NoteGemini 4 Argon at 63.8: every higher row costs at least 2.2xAntigravity CLI · Gemini 4 Argon (default) scores 63.8 on AA Index at $5.84 per task, third of 31 configurations. Every row above it costs at least 2.2x as much.
- NoteGPT-6.1 Sol’s best Codex row is 62.9 at $1.04 a task, not maxCodex · GPT-6.1 Sol (xhigh) scores 62.9 on AA Index at $1.04 per task, fourth of 31 configurations. Max scores 60.1 at $1.55.
- ModelClaude Haiku 5.5Anthropic · Index 43.4 · $0.21 per task
- ModelMistral Large 4 PreviewMistral · Index 38.4 · $1.1 per task
- ModelLing 3.1 FlashInclusionAI · Index 41.1 · $1 per task
- ModelGemini 4 ArgonGoogle · Index 52.6 · $2 per task
- ModelGPT-6.1 SolOpenAI · Index 51.8 · $0.72 per task
Subscription vs API vs GPUs
One maxed ChatGPT Pro seat implies a monthly token volume. The calculator prices it five ways: the subscription sticker, OpenAI and DeepSeek API rates, GPUs you buy, and GPUs you rent.
