Model intelligence
Inspect every assumption.
Feature heuristics, measured adoption, and workload cost projections answer different questions. None establishes quality on your actual production workload.
Capability index · alm-capability-0.7
context_points = min(18, 3 × log10(max(context_tokens, 1))) capability_points = 7 × count(reasoning, tools, structured_output, vision) index = min(99, round(42 + context_points + capability_points + (open_weights ? 5 : 0), 1))
Unconfirmed capabilities contribute zero points. The base and weights are published editorial preferences, not measured quality or a percentage of tasks passed. Only currently listed general text or chat models with a finite positive context limit, known creator, and at least one observed provider offering marked current are ranked. Realtime and other specialized model types are excluded. Provider labels with a shared canonical source ID or matching stable serving ID are scored once, even when provider release dates differ. Conflicting source IDs and unverified matches stay separate. Records containing more than one canonical source ID are held out of rankings until their identity is reviewed. Moving aliases such as “latest,” “auto,” and “agentic chat” are not ranked. Open weights is a preference, not evidence of better answers. Equal scores use canonical slug ascending; payment and sponsorship never enter the formula.
Directory identity review v5 joins source-backed provider aliases even when their marketing names differ. It gives each unresolved fixed-model record an unambiguous source or serving ID label. Different pinned versions remain separate; a matching name alone never joins them. Moving provider routes and empty unverified creator records are omitted from model listings while their stored source history is retained. Each daily refresh checks new name collisions and mixed source IDs before ranking or public publication; an unreviewed case stops the update.
Value index and reference offering
Choose a single current offering with both prices observed within seven days and no unresolved price review, minimizing input + 0.25 × output. When paid offerings exist, exclude zero-rate routes from this reference because their usage limits may not be comparable. Free offerings remain visible and are used when they are the only eligible rates. Ties use the stable offering ID. Never mix prices across providers. Rates are USD per million tokens. Missing, invalid, stale, or disputed rates produce no value score. Explicit zero rates remain eligible. This fixed 4:1 input/output token basket excludes cache, request fees, reasoning surcharges, tiers, and other billing conditions; it does not measure total cost or quality per dollar. Inspect the provider offering before using a price. Price eligibility uses alm-price-eligibility-1.2. Offerings explicitly labeled as plans, subscriptions, or Copilot are conservatively excluded from token estimates and reference prices until their billing scope is represented. This label filter does not establish that every other offering is unconditional; provider terms still apply. Source prices remain visible. Version 0.3 removes invented fallback prices and aligns current-model eligibility and tie-breaking across surfaces. Version 0.4 also excludes models with no observed offering. Version 0.5 resolves source-backed provider delivery labels before scoring. Version 0.6 also consolidates source-backed date aliases and excludes moving provider routes. Version 0.7 excludes mixed-source records. Earlier ranking snapshots retain their original methodology.
price_score = max(0, 100 − 22 × ln(1 + input + 0.25 × output)) value_index = min(99, round(0.72 × capability_index + 0.28 × price_score, 1))
Interactive ranking explorer · alm-explorer-0.6
The explorer separately weights feature coverage (45), price score (30), context score min(100, 15 × log10(max(context, 1))) (15), and openness (100 for confirmed open weights, zero otherwise) (10). Feature coverage is 25 × the count of confirmed reasoning, tools, structured output, and vision flags; it excludes context and openness so each preference is counted once. Unknown openness remains labeled unknown and earns no confirmed-openness points. User weights divide by their sum; all-zero weights produce no ranking. With any positive price weight, models lacking an eligible paired price are excluded; set price to zero to compare their metadata. Ties use canonical slug ascending. The ingestion ranking job retains default capability and value snapshots with input hashes; arbitrary custom explorer weights are local calculations. Movement stays unknown without a comparable history.
Cost formula · alm-cost-2.0
Sum uncached input × input rate, cached input × sourced cache rate, visible output × output rate, and additional reasoning × reasoning rate (or output rate), divide token costs by one million, then add any sourced per-request fee. Multiply by calls including retries and monthly runs. Annual cost is 12 × monthly; sensitivity is ±20% token/call volume. The multi-call calculator sums the estimates for the selected calls. Projections are not verified savings.
Historical coding runs · aider-run-explorer-v1
The coding benchmark explorer shows published Aider Polyglot results for 225 exercises. Each configuration remains a separate run. The default view shows single-model runs ordered by reported percentage correct, highest first. Multi-model architect/editor setups can be viewed separately. Date order is newest first; model-name order is ascending; all ties use source run ID ascending. Sponsorship and referral economics never affect ordering.
Scores are copied from the licensed source without alteration. Run dates, settings, harness versions, capture hashes and attribution are retained. Different configurations are not controlled head-to-head comparisons. A catalog link requires a reviewed exact provider/model identity and a unique canonical match. Unmatched runs remain visible under their source label; mixed-model setups do not inherit a single model’s identity.
Model pages and model comparisons use a different selection rule: latest test date, then run ID ascending (aider-polyglot-latest-run-v1), rather than the highest score. These historical results do not change capability/value index formulas or establish quality on a current production workload.
History and market coverage
Vercel daily percentages remain shares within Vercel AI Gateway. OpenRouter tokens remain public OpenRouter traffic, preserving 64-bit counts and source as-of timestamps. They are never combined into a whole-market share. The matched price index (alm-matched-price-1.0) compares a fixed 1M-input/250K-output basket of offerings observed on both endpoint dates; entrants and zero-cost baselines are excluded. No historical values or proprietary benchmark scores are invented.
Canonical adoption matching (alm-adoption-match-1.0) uses exact, case-insensitive model names or Vercel offering IDs. Ambiguous labels remain unmatched. Directory adoption uses the latest Vercel text request share; momentum subtracts the observation exactly seven days earlier, in percentage points. Missing baselines stay unavailable, stale models sort last, and no fuzzy identity guesses are made.