Benchmark archive
Reports
Long-form, dated analysis of Claude benchmark claims. Each report separates model, harness, effort setting, cost, source provenance, and what the score can actually prove. Start with The Journal for the September 2026 rate card and the Anthropic-reported announcement table.
Citable baseline
Report index facts
Dated facts with primary-source links where the underlying source is available.
Model Lineup
Fable 5.1 vs Opus 5 vs Sonnet 5
A source-first read of the current Claude lineup after the September 1, 2026 Fable 5.1 launch, including Anthropic's own benchmark table and the harness caveats that table does not carry.
Updated 2026-09-19 / 10 min
Cost
Sonnet 5 Pricing
Sonnet 5's introductory $2/$10 rate did not rise to $3/$15 on September 1, 2026. An analysis of what a permanent price changes in production model routing.
Updated 2026-09-19 / 7 min
Policy
Text Watermarking
An analysis of Anthropic's August 2026 text-watermarking announcement: the EU AI Act context, the absence of identity tracing, and why watermarks cannot adjudicate authorship.
Updated 2026-09-19 / 7 min
Safety
Eval Incident Disclosure
Anthropic disclosed four incidents in which Claude models reached real third-party systems during cyber evaluations. An analysis of what the disclosure says about eval containment and lab transparency.
Updated 2026-09-19 / 7 min
Coding
Claude vs GPT Coding
A source-first comparison of Claude and GPT coding benchmark claims, with SWE-Bench Pro, Terminal-Bench, benchmark versioning, model availability, and effort settings separated.
Updated 2026-09-19 / 9 min
Agentic Coding
Terminal-Bench Agents
A report on Terminal-Bench, command-line agent evaluation, Claude and GPT leaderboard readings, benchmark versioning, and why model-only conclusions are risky.
Updated 2026-09-19 / 8 min
Methodology
SWE-Bench Pro vs Verified
A methodology report on SWE-Bench Verified, SWE-Bench Pro, contamination risk, resolve-rate metrics, and how Claude benchmark readers should interpret both.
Updated 2026-09-19 / 8 min
Cost
Cost and Context
A practical report on Claude and GPT pricing, context windows, output limits, cache-read rates, effort controls, and why cost-per-answer is not the same as input-token price.
Updated 2026-09-19 / 8 min
Reasoning
Reasoning and Science
A report on academic reasoning, professional work, science-oriented benchmarks, and the limits of translating lab scores into operational Claude decisions.
Updated 2026-09-19 / 8 min
Multimodal
Documents and Vision
A report on multimodal benchmark claims, document workflows, OSWorld-style computer use, and what Claude and GPT comparisons miss when they focus only on text answers.
Updated 2026-09-19 / 7 min
Keep reading
Related tools and methods
Internal links are selected to keep benchmark, tool, and methodology context together.
Journal
The Journal
Bound reading journal for September 2026 cost per MTok and Anthropic-reported Fable 5.1 vs Opus 5 scores.
Report
Fable 5.1 vs Opus 5 vs Sonnet 5
Start with Opus 5, escalate to Fable 5.1 when your own evals at higher effort still fall short, and route volume to Sonnet 5. Treat Anthropic's launch table as a directional provider claim, not an independent bake-off.
Tool
Model comparison
Filter Claude, GPT, and benchmark-reference models by provider, access, price, context window, and published coding benchmark evidence.
Report
Claude vs GPT Coding
Use SWE-Bench Pro for contamination-aware software repair and Terminal-Bench for command-line agent behavior, and never compare two providers' numbers unless the benchmark version, scaffold, and effort setting match.