Claude Opus
Core Facts
Claude Opus is a flagship large language model launched by Anthropic and is the version in the Claude series with the largest parameter scale and strongest reasoning capability.
The Claude model family is generally divided into three tiers: Haiku, Sonnet, and Opus.
The Opus series, together with Sonnet (balanced) and Haiku (lightweight), form Anthropic's three-tier product matrix
Timeline (last 90 days)
On the PFDelta task, Claude Opus achieved a 64.4% improvement through three changes: using the gradient-safe squared form of the evaluation metric as an auxiliary loss, adjusting the MSE weight coefficient from 0.1 to 0.3, and switching to a cosine annealing learning rate strategy with warmup.
All four agents achieved positive average relative improvements across the 25 tasks: Claude Opus 10.4%, Claude Sonnet 9.6%, Gemini 6.8%, and Codex 4.4%; the existing best results on all benchmarks were surpassed, and no academic misconduct was found in the audit stage.
There are clear differences in workflow patterns: Claude Opus basically uses up the full 12-hour budget, checking logs after each round of experiments before starting the next, making it an iterative optimizer; Codex determines a long-running script in the first hour and then runs hundreds of experiments in one go, making it a planner; Gemini oscillates between the two modes.
The experiment selected four frontier coding agents—Claude Opus, Claude Sonnet, Codex, and Gemini CLI—each allocated eight H200 GPUs and 12 hours of runtime.
On the Mars terrain segmentation task MarsBench, Claude Opus replaced RetinaNet with Faster R-CNN, raising mAP from 16.5 to 30.4.
After adding the skill plugins, Claude Opus's average relative improvement rose from 10.4% to 20.1%, and Codex's rose from 4.4% to 8.1%; Claude Opus improved on 18 of 20 benchmarks, and Codex improved on 12 of 20 benchmarks.
The distribution of optimization strategies differs across agents: for Claude Opus, about 56% was hyperparameter tuning and 44% was model architecture changes; for Codex, hyperparameter tuning accounted for 44%, with feature engineering, data augmentation, and loss design each accounting for 24%; for Gemini, hyperparameter tuning accounted for 48% and architecture changes 32%.
Haiku corresponds to speed priority, Sonnet to balance, and Opus to the strongest reasoning capability.
In Ruben Soto's early comparison tests, Terra ran noticeably faster and at lower cost, but the results may still need to be run a few more times.
Claude Opus scored highest in the test, reaching 87%.
40 more timeline events
All Facts (20)
Claude Opus is a flagship large language model launched by Anthropic and is the version in the Claude series with the largest parameter scale and strongest reasoning capability.
80%VerifiedThe Claude model family is generally divided into three tiers: Haiku, Sonnet, and Opus.
75%VerifiedThe Opus series, together with Sonnet (balanced) and Haiku (lightweight), form Anthropic's three-tier product matrix
75%VerifiedThe API pricing for Opus-tier models is typically 3–5 times that of the mid-tier Sonnet.
70%VerifiedThe correct metric for evaluating model cost is not the price per token, but the Total Cost of Task.
65%VerifiedThe Claude 3 series uses Haiku, Sonnet, and Opus to name different capability tiers
65%VerifiedWithin Anthropic's product line, the Opus series focuses on the strongest reasoning capability, the Haiku series focuses on being lightweight and fast, and the Sonnet series strikes a balance between the two.
65%UnverifiedSonnet并非Anthropic的顶级模型，Opus才是Anthropic的顶级模型
90%UnverifiedClaude Code通过npm安装，基于Claude Sonnet/Opus模型运行
90%UnverifiedFor Claude Opus, input pricing is $15 per million Tokens and output is $75 per million Tokens
85%UnverifiedThe Claude Opus series has built a strong reputation in reasoning ability and code generation
75%Partially VerifiedAnthropic's Claude Opus is the flagship version of its Claude model family, positioned to handle the most complex reasoning, analysis, and creative tasks
70%UnverifiedUsing top-tier models like Claude Opus or GPT-4 for simple tasks is the most common source of Token waste in Cursor
70%Partially VerifiedOn the Terminal Bench terminal task test, Gemini 4 Argon scored 57.4%, below Claude Opus 5.5's 66.4%.
65%UnverifiedWithin the Claude series, Opus-tier models excel at complex architectural reasoning and handling ambiguous requirements but have a higher unit price, while Sonnet-tier models are suited for everyday tasks and workflow execution and offer the best cost-effectiveness.
60%UnverifiedA benchmark test conducted by Databricks on their own codebase of one million lines showed that Pi outperformed Claude Code and Codex in most scenarios, with the Pi + Claude Opus combination achieving the highest code quality.
60%Unverified采用分层模型使用策略（高频任务用轻量模型、关键任务用高级模型）可以将额度使用效率提升3-5倍
60%UnverifiedWhen Claude Opus is used with Max mode, the token consumption of a single request can be several times or even more than ten times that of normal mode.
60%UnverifiedAnthropic's model product line forms a three-tier structure of Haiku → Sonnet → Opus
60%UnverifiedThe experiment selected four frontier coding agents—Claude Opus, Claude Sonnet, Codex, and Gemini CLI—each allocated eight H200 GPUs and 12 hours of runtime.
50%