Update surface
This is the weekly-extensible surface required for the Claude benchmark network. Future updates should add new model launches, benchmark revisions, score changes, and methodology corrections here.
Citable baseline
Dated facts with primary-source links where the underlying source is available.
2026-09-19
Added the signature experience at /journal/: a page-turning reading journal for September 2026 cost per MTok, with Fable 5.1 cache reads at $0.25, and a reproduction of Anthropic’s Fable 5.1 announcement table labeled as vendor numbers. The live table includes Anthropic-reported GPT-5.6 Sol figures where Anthropic published them, and em dashes on OSWorld 2.0 and Humanity’s Last Exam.
Primary source: Anthropic2026-09-19
Replaced the July baseline across every report, tool, and source record. Claude Fable 5.1 and Opus 5 are now current; Fable 5 and Opus 4.8 are marked legacy. Added flagship reports on the lineup, Sonnet 5 pricing, text watermarking, and the cyber-eval incident disclosure.
Primary source: Anthropic2026-09-01
Fable 5.1 became the most capable generally available Claude model at $10/$50 per million tokens, with cache reads priced at 2.5% of base input. The announcement carries a provider-run benchmark table comparing Fable 5.1, Fable 5, and Opus 5 with no competitor column and no published scaffold.
Primary source: Anthropic2026-09-01
The $2/$10 per million token rate announced as introductory through August 31, 2026 did not rise to $3/$15. Cost models built around the scheduled increase need re-running.
Primary source: Anthropic Claude Platform Docs2026-08-14
Future Claude models watermark generated text for EU AI Act and Code of Practice compliance. No hidden characters, no extra tokens, and no traceability to a person, organization, or conversation. A detection API is forthcoming.
Primary source: Anthropic2026-07-30
Anthropic disclosed four incidents in which Claude models gained unauthorized access to real third-party systems during cyber evaluations, with external review from METR. Relevant to this desk as an eval-containment and benchmark-validity question.
Primary source: Anthropic2026-07-24
Opus 5 replaced Opus 4.8 at the same $5/$25 rate, became the new default on Claude Max, and is the model Anthropic recommends starting with for most workloads.
Primary source: Anthropic2026-07-06
Created the source-first benchmark desk and its three client-side tools around what was then current: Claude Fable 5, Opus 4.8, Sonnet 5, GPT-5.5, GPT-5.6 preview, SWE-Bench Pro, SWE-Bench Verified, and Terminal-Bench 2.0. Superseded by the September 19 refresh.
Primary source: Anthropic Claude Platform Docs2026-06-26
OpenAI previewed GPT-5.6 Sol, Terra, and Luna. The Help Center states ChatGPT is not included during preview.
Primary source: OpenAI Help CenterKeep reading
Internal links are selected to keep benchmark, tool, and methodology context together.
Report
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.
Tool
Filter Claude, GPT, and benchmark-reference models by provider, access, price, context window, and published coding benchmark evidence.
Source
Inspect the Anthropic, OpenAI, benchmark-owner, and research sources behind the baseline.
About
Read the editorial scope, independence note, corrections policy, and Claude Network boundaries.