Claude Opus 5.5
Long-running coding agents and knowledge work.
Claude API · standard processing
COMPARE MODELS
Filter by task and budget. Check versions, serving conditions and sources before you choose.
Language model comparisons use the uncached input price in USD per 1 million tokens. Output is billed separately and shown on each card. Excludes tools, cache writes, storage, and tax.
TASK LEVEL × INPUT PRICE
Vertical positions use published Artificial Analysis Intelligence Index scores. Task bands are our browsing guide; test models on your own workload before choosing.
Dot height is the exact score; the right-side regions show its task tier. Labels are offset for readability. Select a point or model name for details.
Not plotted because a verified index score or a positive input price is unavailable: Kimi K3.1 Pro, Kimi K3.1 Pro Highspeed, MiniMax M2.7 Highspeed, Qwen3.8-Max.
The vertical axis uses published Artificial Analysis Intelligence Index scores without editorial score adjustments. Our browsing bands group scores above 50 under complex planning, 39–50 under everyday coding, and below 39 under simple tasks. These categories are not task-specific test results. Task buttons highlight a band without changing the published score or model placement.
Each model’s details show the recorded index version, reasoning setting, source and source-check date. Records are dated snapshots, not a live leaderboard or a test of SaveMyToken’s API service. Models without a reviewed score are omitted. Artificial Analysis methodology ↗
“Lower cost” describes token prices, not proven value per successful task. For real value, compare correctness, retries, tools, and total billed tokens on the same tasks.
The horizontal axis uses a logarithmic scale: equal spacing means equal price ratios. Prices are USD per one million native tokens, using uncached input under each entry’s stated API conditions. Price filters, sorting, and chart positions use the same input rate. Capability evidence and pricing sources are linked in each detail page.
Model detail pages show a (fast) badge only when a published median output speed exceeds 150 tokens/s in the tested settings. That speed does not include time to first token or total task time. Speed methodology ↗
Long-running coding agents and knowledge work.
Claude API · standard processing
General coding, visual tasks, and document workflows.
Claude API · standard processing
Multimodal reasoning and tool workflows with peak and off-peak API rates.
DeepSeek API · peak hours
Software engineering, autonomous agents, and enterprise workflows.
Gemini Developer API · Standard · introductory
Complex coding, scientific reasoning, and multi-step agents with multimodal input.
Gemini Developer API · Standard · introductory through November 30, 2026
Text-only reasoning, complex coding, and long-running agents.
Z.ai API · listed USD rates
Use-case tags are editorial groupings based on official capabilities. Language models earn (fast) only with a published median output speed above 150 tokens/s in the linked test settings. Video speed labels remain provider-reported. Unknown values are excluded from price bands and sorted last.
Original research and practical experience, with short notes to help you choose.
A comparison of model–harness combinations that puts task success and cost next to each other.
How domain-specific benchmark slices and weights change the meaning of a model capability score.
A dated evaluation of GPT-6 Astra across capability, reasoning settings, token use, and task cost.
| # | Model / reasoning mode | Compare | ||||
|---|---|---|---|---|---|---|
| 01 | ✦Gemini 3.1 ProGoogle · High | 94.3 | 80.6 | 91.7 | 68.5 | |
| 02 | ◎GPT-5.4OpenAI · xHigh | 93.0 | Not reported | Not reported | 75.1 | |
| 03 | ✳Claude Opus 4.6Anthropic · Max | 91.3 | 80.8 | 88.8 | 65.4 | |
| 04 | KKimi K2.6Moonshot AI · ThinkingOpen weights | 90.5 | 80.2 | 89.6 | 66.7 | |
| 05 | DDeepSeek V4 FlashDeepSeek · MaxOpen weights | 88.1 | 79.0 | 91.6 | 56.9 | |
| 06 | ZGLM-5.1Z.ai · ThinkingOpen weights | 86.2 | Not reported | Not reported | 63.5 |
Percentage of real GitHub issues resolved. Results depend on the agent framework, tools, and sampling setup.
Graduate-level science questions, measured with Pass@1.
February 2026 HMMT math problems, measured with Pass@1. Do not combine with AIME or other years.
Multidisciplinary knowledge and reasoning using exact match (EM), distinct from the original MMLU.
Code generation using source-reported Pass@1. Missing public results are marked as not reported.
Accuracy on multi-step terminal tasks. Results depend on the agent setup and tool permissions.
Software engineering tasks resolved, using a different task set from SWE-bench Verified.
This table uses HLE Pass@1 without tools. Do not mix it with tool-enabled results.
Complex questions requiring browsing and multi-step research, measured with Pass@1.
Multi-round retrieval over 1M-token context (MMR). Missing values mean not reported, not zero.
MODEL NOTEBOOK
8 source snapshots from DS-V4-HUUB. Unreviewed specifications are not used as current benchmark evidence.
DeepSeek
Source snapshot · Review required →Anthropic
Source snapshot · Review required →OpenAI
Source snapshot · Review required →MiniMax
Source snapshot · Review required →xAI
Source snapshot · Review required →Alibaba
Source snapshot · Review required →Zhipu AI
Source snapshot · Review required →