ModelRank

Methodology

ModelRank is an editorial reading of public coding evidence. It is a decision aid, not a claim that one model will win every repository, language, agent or task.

Scope and evidence

We use published or publicly inspectable measurements and retain the source URL, model identity, configuration and snapshot date where available. The current coding snapshot was updated on ; source fetch timestamps are retained in the JSON metadata. It includes Artificial Analysis measurements, LiveBench checks and Terminal-Bench 4.0 agent results. ModelRank does not run or present private first-party tests of its own; the results shown here are transcriptions and calculations from the cited public sources.

A benchmark result belongs to the tested system. Agent, harness, tool permissions, reasoning effort, model version, budget and task set can all change the result. We keep those configuration labels instead of presenting an agent score as an isolated model property.

Coding Index

When a model has both required Artificial Analysis measurements, the displayed index is calculated from two equally weighted ratios against the fixed deepseek-v4-flash-0731 reference (100). The implementation is:

100 × √[(Terminal-Bench 2.1model / 0.786516853932584) × (SciCodemodel / 0.503472222222222)]

The two inputs are stored as completion-rate fractions (for example, 0.88 represents 88%), and the result is rounded to two decimal places in the data layer. The baseline is fixed to 100; it is not a claim that the reference is average, and an index of 120 is not “20% more capable” in general. LiveBench and Terminal-Bench 4.0 are supplementary checks and are not silently mixed into this two-test index.

If either required measurement is missing, there is no comparable Coding Index. A reported number copied from another source is labelled provisional and is not treated as independent verification.

Editorial tiers

Tiers are editorial assessments based on the available public evidence, configuration comparability, source quality and coverage. The assignments are versioned under coding-tiers-2026-09-10; they are not automatic score cut-offs or statistically proven classes. Price, speed, context window, sponsorship or affiliate availability do not determine a tier.

Within a tier, tables use the Coding Index when one is available and use a stable name fallback otherwise; that display order is a reference sort, not a proven capability ranking. A close index, or a higher index inside a tier, does not establish a reliable winner. Different tiers indicate a different editorial assessment of the evidence, not a guarantee of performance on a particular project.

Dates, versions and unknowns

“Fetched”, “released”, “leaderboard updated” and “evaluated” are different dates. We show the date a source or snapshot supplies and do not turn a publication or retrieval date into an evaluation date. In the current snapshots, some task dates, evaluation dates and task ranges are explicitly unknown; unknown remains unknown. Source version, configuration and retrieval date should be read together.

Terminal-Bench 4.0 is a model-plus-agent result set. Its current snapshot is version 4.0.0, released on 2026-08-28, and the source reports confidence intervals. Overlapping intervals or a small rank gap should not be read as a conclusive order.

Pricing and provisional records

Prices are separate purchasing information, quoted in USD per million input or output tokens for the displayed channel and time band. Estimates exclude items documented in the interface as excluded, such as cache effects, tool fees or additional uncounted tokens. A missing or pending quote is not a zero price and does not lower a coding tier.

Provisional records are kept visible so readers can see the catalog boundary. They are not promoted to comparable scores until the required source evidence and identity/configuration checks are available.

Updates, corrections and affiliations

Data updates replace or extend a versioned snapshot and retain the cited source URL and retrieval metadata. Article publication dates are not backfilled: an article has an updated date only when its content was materially revised after publication. Corrections should be made at the source data, content or code that caused the error; a current page alone is not evidence that an earlier measurement was rerun.

ModelRank is the editorial label for this site. The site distinguishes its assessments from the cited publishers, benchmark projects, model vendors and other linked organizations; a link identifies a source and does not by itself indicate endorsement or certification. External data, names and trademarks remain subject to their owners’ terms.

Primary references

Read the editorial policy and site scope →