Independent benchmark desk / July 2026 baseline

Claude Reports

Source-first analysis of Claude, GPT, and agentic AI benchmarks. We preserve the part most leaderboard screenshots omit: model access, harness, effort, cost, contamination risk, and what the score is allowed to prove.

Read reports How we read scores
Claude Reports benchmark desk graphic with bar chart and source-first labels
Benchmark reporting for coding, terminal agents, cost, context, reasoning, and multimodal work.
Fable / Opus / Sonnet Anthropic's current comparison set includes Claude Fable 5, Opus 4.8, Sonnet 5, and Haiku 4.5. Primary source: Anthropic docs
SWE-Bench Pro Designed around contamination, task diversity, underspecified work, and reproducible environments. Primary source: Scale Labs
Preview matters GPT-5.6 Sol is a limited preview, not a universal ChatGPT replacement at this baseline date. Primary source: OpenAI Help

Featured analysis

Claude vs GPT is not one scoreboard.

The right comparison changes when the benchmark changes. SWE-Bench Pro, Terminal-Bench, provider launch tables, and private task suites answer different questions.

Coding

Claude vs GPT Coding Benchmarks: What the 2026 Scores Actually Say

A source-first comparison of Claude and GPT coding benchmark claims, with SWE-Bench Pro, Terminal-Bench, model availability, and effort settings separated.

Read the report

Methodology

SWE-Bench Pro vs Verified

Use Verified for continuity and Pro for current frontier signal; when scores conflict, inspect dataset, scaffold, cost caps, and whether the result is model-only or agent-plus-model.

Agentic Coding

Terminal-Bench Agents

Terminal-Bench is strongest when you care about the shipped agent experience, not when you want to isolate a base-model capability.

Cost

Cost and Context

Compare total task cost: input, cache writes, cache reads, output, effort, retries, and whether the model finishes without human intervention.

Primary-source posture

"SWE-Bench Verified is increasingly contaminated."

OpenAI's warning is one reason this site reports benchmark provenance, not just winner labels. Read the source.