Every score comes from a fixed, published formula applied to independently verified benchmark results and official API prices. AI helps write editorial summaries but never sets a number. When the rules change, they change as a new numbered version with a freshly published snapshot, so the rules on this page are always the ones behind the current rankings.
The two scores that determine the public leaderboard.
Quality and Value are the only official PickAI scores. They rank the same published roster with different rules, so buyers can compare raw capability separately from buying power.
Quality Score
Capability-first benchmark pool
(2 × HLE + Terminal-Bench 4.0) / 3
66.7% Humanity's Last Exam (HLE) - Expert-level questions across mathematics, science and the humanities, answered without tools.
33.3% Terminal-Bench 4.0 - Agentic coding and system tasks completed in a real command-line environment.
Both results come from Artificial Analysis and must use the same model release, reasoning effort and fallback, taken at the highest effort for which both are published. HLE is text-only with no tools. Terminal-Bench 4.0 uses the mini-swe-agent harness, three repeats and average pass@1. Scores sit on fixed 0–100 percentage scales, so adding or removing a peer cannot move them. A missing or mismatched input means no Quality score, never a guess or a reweighting. Ranks use the unrounded total; scores display to one decimal place.
Official vs raw views
Quality and Value are scores. Intelligence and Coding are benchmark lenses.
The site publishes two official scores and two raw benchmark views. Intelligence ranks Humanity's Last Exam results in thinking mode with no tools. Coding ranks Terminal-Bench 4.0 results directly. Supporting benchmarks stay visible without becoming extra PickAI scores.
Official scores
Quality ranks capability across the shared benchmark pool. Value ranks the smartest qualified models for the lowest normal-use token cost.
Raw benchmark views
Intelligence and Coding each rank the same roster by a single benchmark. They add evidence, but they do not change the official scores.
What stays separate
Display-only signals stay visible without becoming hidden score inputs.
Subscription pricing can appear on Quality pages without changing the Quality score.
Terminal-Bench 4.0 ranks the Coding view directly. Older and alternative coding benchmarks are never substituted for it.
Score coverage
What the official scores include, and what the raw benchmark views show.
Included means the signal is part of the ranking rule. Tracked means it is shown when verified evidence exists but does not change the rank. Excluded means it is deliberately kept out of the formula.
Official PickAI scores
These cards define the published order.
Quality and Value are formula-driven. The score order is public, deterministic, and separate from the benchmark-only tabs.
Quality Score
Quality is benchmark-only: HLE and Terminal-Bench 4.0 are the only scored inputs. Price, provider speed and supporting benchmarks never change it.
HLE (text-only, no tools)
Two-thirds of the score; the broad reasoning anchor.
Terminal-Bench 4.0
One-third of the score; agentic coding evidence.
Supporting benchmark evidence
These measurements provide additional context without changing Quality, Value or the HLE ranking. Results use the published model release, reasoning effort and fallback. Missing or outdated evidence is shown as n/a.
AA-Omniscience
Factual knowledge and hallucination restraint. Index from −100 to +100: correct answers earn credit, incorrect answers lose credit, and abstentions receive zero. Higher is better. Supporting evidence only.
AA-LCR v1.1
Long-document reasoning across 100 questions, with three attempts per question. Average pass@1 percentage; higher is better. Version 1.1 is not directly comparable with older AA-LCR results. Supporting evidence only.
GDP.pdf All-pass
Artificial Analysis evaluation of professional-document reasoning: 100 tasks and five attempts per task. Percentage of attempts passing every grading criterion. Higher is better. Distinct from GDPval-AA and other GDP.pdf implementations. Supporting evidence only.
AutomationBench-AA
Business workflow automation across 657 tasks. Average percentage of objectives completed; any guardrail violation gives that task zero credit. Higher is better. This is partial-credit scoring, not the percentage of fully completed tasks. Supporting evidence only.
Supporting buyer signal
API cost per standard conversation stays visible beside the rank.
Shared basket Standard conversation size - Each model uses the same fixed input/output token basket for like-for-like API token-cost comparison.
API rates Published input and output pricing - The estimate uses verified API prices from the current snapshot. Missing prices are never guessed.
Separate plans Subscription prices and limits - Verified subscription fees are listed separately. A plan fee does not establish included conversation capacity.
Disclosure Shown without changing Quality - The Quality leaderboard can show this cost for context; it never changes the Quality rank.
This USD estimate prices a chosen 3,000-token basket. Lower is cheaper for that basket. It is not measured cost per completed task or included subscription usage. Additional reasoning tokens, tools, caching, long-context rates and taxes can change the actual bill. It does not change the published ranking formulas.
AI tools badges
Visual editorial signals for the tools directory.
Badges on the AI tools page are selected through AI editorial review. They are not scores, do not affect model rankings or tool order, and are never assigned because of affiliate payouts. The question-mark badge means the tool has not been reviewed deeply enough yet to assign a more specific signal.
Most PopularMost PopularBroad public awareness, strong adoption, or frequent discussion makes this tool especially recognizable.
Broad public awareness, strong adoption, or frequent discussion makes this tool especially recognizable.
Highly Rated
Trust & governance
Why you can trust what you are reading.
The rankings are only as good as the evidence behind them. These publication rules, review schedules and safeguards keep unverified or mismatched benchmark results out of the public rankings.
How publication works
Editorial prose is allowed. Score edits are not.
Scores are locked once a snapshot is generated.
Public pages read the published snapshot instead of fetching live benchmark pages.
Methodology changes ship as a new, numbered version with a republished snapshot, never as a silent adjustment.
Evidence & benchmark sources
Accepted reputable citations only, grouped by the evidence they support.
Public ranking rows are published only when the underlying benchmark and pricing inputs can be tied to accepted reputable sources that any buyer can fact-check. Unsupported sources are omitted rather than estimated.
Quality benchmark evidence
Independent Artificial Analysis results for HLE and Terminal-Bench 4.0. Each note records the exact model release, reasoning effort and fallback that was scored.
Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback)
FAQ
Questions buyers ask before they trust a ranking.
What is the difference between Quality and Value?
Quality ranks capability alone, from HLE and Terminal-Bench 4.0. Value asks which qualified model is smartest for the lowest normal-use token cost, using HLE and published API prices. The two scores are calculated independently and never combined.
Why can the Intelligence view rank models differently from Quality?
The Intelligence view is the Humanity's Last Exam lens in thinking mode with no tools over the same published roster. It surfaces HLE evidence directly, while Quality remains the broader deterministic ranking.
Terminal-Bench 4.0 only, measured with the mini-swe-agent harness, three repeats and average pass@1. Older Terminal-Bench versions, SWE-Bench Pro and other coding benchmarks are not substituted for it, because different versions and harnesses produce scores that are not directly comparable.
Value Score
Smartness and conversation affordability
0.40 smartness + 0.60 conversation affordability
40% Smartness - A model must score at least 20% on HLE to qualify; higher HLE scores then earn more credit.
60% Conversation affordability - Uses the same fixed conversation size and each model's published API token prices.
A model must score at least 20% on HLE to qualify. Value combines 40% linearly normalized HLE with 60% log-normalized affordability across eligible models, using 1,200 input and 1,800 output tokens per conversation.
Routing latency and throughput are excluded because they vary by provider, region, and load.
Quality leaderboard rows can show API cost per standard conversation beside capability evidence without changing the Quality rank.
AI-written prose can explain the roster, but it never edits numeric values.
Other coding benchmarks
Older Terminal-Bench versions, SWE-Bench Pro, DeepSWE and FrontierCode never substitute for Terminal-Bench 4.0.
Supporting benchmarks
AA-Omniscience, AA-LCR, GDP.pdf and AutomationBench-AA are shown for context only.
Subscription price
Shown for buyer context; not a Quality input.
Token costs
Used only by the separate Value Score.
Value Score
Value ranks qualified models by a plain buyer question: what is the smartest model for the lowest normal-use token cost?
HLE
Required floor and smartness anchor.
Standard conversation cost
Calculated using the same fixed token basket for every model.
Subscription price
Shown as context, not scored separately.
Missing pricing
Not guessed. A model needs published input and output token prices for Value.
Raw benchmark views
What the benchmark-only tabs show.
Each raw view ranks one benchmark directly over the same roster. Supporting metrics are displayed but never become hidden inputs.
Intelligence View
A direct Humanity's Last Exam ranking (thinking mode, no tools) over the same published roster. Supporting benchmarks stay visible, but they do not create a third PickAI score.
HLE (thinking, no tools)
Ranking signal.
AA-LCR v1.1
Long-document reasoning, average pass@1 across 100 questions; supporting evidence only.
GDP.pdf All-pass
Artificial Analysis professional-document reasoning. Percentage of attempts passing every grading criterion; supporting evidence only.
AutomationBench-AA
Business workflow objectives completed, with zero credit for tasks that violate guardrails. Partial-credit percentage; supporting evidence only.
AA-Omniscience
Independent factual-reliability evidence displayed when available; never a Quality or Value input.
Context window
Not part of the raw intelligence view.
Speed
Not part of the raw intelligence view.
Price
Not part of the raw intelligence view.
Coding View
A direct Terminal-Bench 4.0 ranking over the same published roster.
Terminal-Bench 4.0
Ranking signal: mini-swe-agent harness, three repeats, average pass@1.
Older or different coding benchmarks
Not substituted. Different versions and harnesses produce scores that are not directly comparable.
HLE
Not part of the raw coding view.
Price
Not part of the raw coding view.
Why the separation matters
Buyers can see more context without losing the ranking logic.
Capability stays capability
Quality and Value do the ranking work. Supporting context does not leak into the official order.
Buyer context stays visible
Pricing, app access, benchmark companions, and editorial notes stay available where they help a purchase decision.
Raw views stay raw
Intelligence and Coding show direct benchmark results rather than a blended third or fourth PickAI score. The ranked HLE result follows the site policy: thinking mode, no tools.
Trust stays explicit
The page tells you which signals rank, which signals inform, and which signals are intentionally excluded.
Highly RatedPublic review or reputation signals suggest this tool is especially well regarded.
Public review or reputation signals suggest this tool is especially well regarded.
Most InnovativeMost InnovativeThe tool stands out for a distinctive workflow, product concept, or AI implementation.
The tool stands out for a distinctive workflow, product concept, or AI implementation.
Creator FavoriteCreator FavoriteThe tool is especially useful for creative work such as design, video, images, audio, or writing.
The tool is especially useful for creative work such as design, video, images, audio, or writing.
Automation PickAutomation PickThe tool is especially useful for workflows, agents, integrations, or operational automation.
The tool is especially useful for workflows, agents, integrations, or operational automation.
Research PickResearch PickThe tool is especially useful for research, source review, search, or evidence-heavy knowledge work.
The tool is especially useful for research, source review, search, or evidence-heavy knowledge work.
Best for TeamsBest for TeamsThe tool is especially useful for team, meeting, workplace, or collaboration workflows.
The tool is especially useful for team, meeting, workplace, or collaboration workflows.
Needs ReviewNeeds ReviewWe have not reviewed this tool deeply enough yet to assign a specific badge.
We have not reviewed this tool deeply enough yet to assign a specific badge.
Refreshes are manual and editor-approved; there is no public or scheduled refresh trigger.
AI-written editorial copy can be updated after publication without touching any score.
A model can be absent from the rankings because it is not the latest generally available release in its family, because the ten-model roster is already full, or because a required result has not yet been independently verified for the exact release being ranked.
Benchmarks under review
These inputs are tracked publicly before they are allowed into the score formula.
DeepSWE and FrontierCode
Newer coding benchmarks on the watchlist. Results must keep their exact benchmark version and evaluation setup, and neither benchmark can enter Quality or Value without a new, published methodology version.
SWE-Bench Pro V2
V2 changes both the dataset and the evaluation protocol. Older SWE-Bench Pro results cannot be relabelled as V2, and neither version substitutes for Terminal-Bench 4.0.
AutomationBench-AA
Already shown as supporting evidence for business workflow automation. It does not affect Quality, Value or either raw ranking.
The rules we do not bend
01
Scores are calculated, never edited
Quality Score, Value Score and benchmark results are fixed by the published formula. AI can draft editorial prose, but it never sets or changes a number.
02
One published snapshot
Every public page reads the same reviewed, published snapshot. Pages never pull live benchmark sites or call an AI model while you browse, and evidence that fails review is withheld rather than shown.
03
Quality and Value stay independent
There is no combined score. Quality and Value are the only official PickAI scores; the raw benchmark tabs are direct views over the same roster.
04
Up to 10 models
Public rankings show up to 10 models: the latest generally available release from each tracked model family, with no hidden extra rows.
05
AI prose is labeled
Any AI-assisted verdict, pros and cons list or best-for summary carries a visible “AI-assisted, editorially reviewed” label.
When updates happen & what we refuse
Update cadence
Reviewed releases: the public snapshot changes only when an editor reviews new evidence and publishes it.
Each review checks for new or retired models, changed prices and new benchmark versions. Sources are due for re-verification after seven days and flagged as overdue after fourteen.
A model release, retirement or price change triggers a recheck sooner.
Hard exclusions
We do not let AI change numeric scores.
We do not rank more than 10 models on public pages.
We do not let anyone outside the editorial process trigger a ranking refresh.
We do not add affiliate links to leaderboard rows or this page.
AI-written editorial copy is labeled wherever it appears
Alibaba Cloud international Singapore standard pricing, 0902 release
API cost per standard conversation uses published snapshot rates. Subscription fees and limits are separate. Source dates remain visible; changed rates require review. This pricing scenario adds no new ranking score.
Quality needs both HLE and Terminal-Bench 4.0 results for the same model release, reasoning effort and fallback. If either is missing or was measured under a different setup, the model is left unscored rather than estimated.
Where do I compare price, app access, and buyer notes in more detail?
Use the models index and the model detail pages when you want pricing, app access, buyer-facing guidance, and the supporting benchmark context next to the ranking.