Model Selection

Source scores and fixed anchors

Nine tasks use six Arena text boards and Code Arena. Finance and strategy share the business board; coding and systems share the software board. Writing, legal, medical and engineering use writing, legal, medicine and mathematical boards. Visual uses Code Arena (webdev). Duplicate names with no registered binding are removed only when each board retains its original central anchors. On each board, lower_anchor is the minimum central rating and upper_anchor is the maximum central rating across every published model setting. The formula is arena_score = 100 * (rating - lower_anchor) / (upper_anchor - lower_anchor). It preserves rating-gap ratios within that board. Equal ratings have equal scores. A zero-range board has no normalised score and leaves its tasks unmeasured. Its raw evidence remains available.

Each raw interval endpoint uses those same fixed central anchors, without clipping to 0–100. The transformed range is conditional on fixed central anchors. It has no joint confidence level and does not include uncertainty in the anchors. Task quality uses central scores. Raw endpoints, votes, publication date, exact model setting and the complete cohort remain recorded.

Artificial Analysis supplement

Artificial Analysis supplies measured Intelligence Index for orchestrator, AA-Briefcase presentation Elo for visual, Terminal-Bench 4.0 for coding and systems, and the published Finance & Accounting, Strategy & Ops, Legal, Healthcare & Medical and Engineering capability indexes for the matching tasks. Strategy uses Strategy & Ops alone. Estimated or unknown Intelligence Index status is excluded from its scores and cohorts. Each measure uses every eligible snapshot record and aa_score = 100 * (cohort_size - midrank) / (cohort_size - 1). Ties use their mean rank. Fewer than two eligible records leaves the measure unavailable. Explicit per-category bindings can use a different tested effort. AA intervals and benchmark publication dates are unavailable; observation dates remain visible. Conditional interval scenarios hold AA central scores fixed. MMMU-Pro remains visible as reading evidence only. It measures reading images, charts and slides. AA-Briefcase presentation judges produced deliverables in its Stirrup harness; Code Arena measures preference on generated web apps. These measures have limited coverage of native presentation tools. Medical and engineering can bind another published effort of the same model when the default effort lacks the index. The page names that effort. Surveying, geospatial and quantity work have no direct measure.

Arena raw-gap scores and AA percentiles use source-specific units, tasks and cohorts. Scaling both to 0–100 permits the chosen weighted index; it does not calibrate equal ability or accuracy. Raw rating gaps retain meaning within each board only. Changed anchors or cohorts can change the combined result.

Professions and task priorities

The seven profession presets populate the existing task ratings. The page marks these values From profession; an edited task shows Chosen. Weekly task inclusion, volume and importance are separate. An included task at importance 1 keeps its positive volume. Legacy requests treat 1 as Not needed. Basic uses Low 3, Important 5 and Critical 7 in both paths. Medical and engineering start excluded. Choosing No profession restores the neutral public defaults. Budget, usage and interfaces keep their values. The URL stores the profession, effective task ratings and edit markers. Existing URLs retain their ratings.

Task quality

A model's quality on a task combines both sources and the user feedback: task_quality = mean(aa_score, arena_score) * (1 + praise_complaint). When one source has no measure, the other stands alone. praise_complaint is the model's Agent Arena praise-versus-complaint score on the Overall board, the only bucket captured. Opus 5.5 on most tasks is about 129. Classification and housekeeping have no measure. They add demand and never change quality.

Each usage level holds a quality floor. A model serves a task only when its quality is at least the floor share of the best model on that task: 0.5 for Light and Regular, 0.8 for Heavy and 0.9 for Multi-agent. "Drop to lower tiers" relaxes every floor to 0.5. Top tier is the share of work on models within 3% of the best on that task. Other work is the lower-tier share.

Each task contributes task_contribution = share * served_quality / (demand * margin). The contribution uses that task's importance share and allocated quality. Weekly demand is the task's session count. Legacy demand is its monthly Opus-equivalent dollar amount. The score sums measured-task contributions.

Weekly sessions and task volume

The empty public page starts a weekly scenario with seven active days. A session is one declared group of work. Sessions per day set frequency. Arena task equivalents per session set size, with 1 as the default assumption. Size changes usage per session and leaves the session count unchanged. An empty sessions-per-day field does not copy the level table. Fleet sessions per day are 2 for Light, 10 for Medium, 20.4 for Heavy, and 61.2 for Multi-agent at 3 agents. The level table still stores legacy counts of 2, 10, and 3. Light and Medium legacy sessions use the $1.35 Opus "10-49" cost. Heavy and Multi-agent legacy sessions use the $9.18 Opus "50+" cost. A weekly session is one Arena task, so the empty field scales each count by that Opus cost divided by $1.35. That is 2, 10, and 20.4 sessions per agent per day. Multi-agent then multiplies by the agent count. An entered sessions-per-day value is used as written. These presets describe scenarios. They are not validated personal burn estimates.

Per-agent scope multiplies frequency by the selected agent count once. Whole-fleet scope uses frequency as the fleet total. Ordinary usage levels use one agent. Multi-agent uses the selected count. The equations are:

weekly_sessions = active_days * sessions_per_day * agent_multiplier.

task_sessions = weekly_sessions * task_volume / included_volume_total.

Only explicitly included tasks with positive volume enter the total. Importance 1 to 7 weights measured-task quality independently. Increasing importance leaves session volume unchanged. Seven days produces 40% more sessions than five days at the same daily frequency and scope. Changing plans, budget or allowances leaves this ordinary session forecast unchanged.

Classification and housekeeping divide the initial total with measured tasks. They enter demand once and contribute no quality score. Each combination uses one background model. It needs positive exact-category costs for every selected background task and positive weekly and monthly allowances at the selected effort. The retained quality rule requires at least 85% of the best usable candidate's task-weighted mean. Among those candidates, the cheapest selected background volume mix wins. Missing usable background tuples make that forecast unavailable.

Exact Arena cost and allowance evidence

Agent-task usage uses Arena Agent Pareto P50 USD per task. Overall, Code, Chat and Work remain separate categories. Code maps coding, systems and orchestrator. Work maps knowledge, presentation and specialist tasks as declared intent proxies. Overall maps the two unmeasured background tasks. An explicit category setting can change that mapping. Agent Chat supplies no established app-chat session cost or allowance transfer. Selecting app chat therefore exposes unavailability.

Each cost row records the exact contender, model, effort, category, workload, statistic, sample counts, capture hash, source URL and date. The retained cost capture is dated 11 October 2026. Named effort bindings include Opus High, Sol Max and Astra Max. Another effort needs its own cost and allowance evidence. Missing costs stay unavailable. P50 describes a typical-task scenario. It does not establish expected mean workload cost or a validated aggregate calibration.

units_per_session = exact_category_P50 * Arena_tasks_per_session.

forecast_usage = sum(task_sessions * assigned_model_share * units_per_session).

Supply names the plan, exact model, effort and workload. Its weekly and monthly API-equivalent USD carry source, basis, date, status and transfer provenance. Accepted model allowances are transferred to the named Arena effort as inferred scenarios because the source quotas are not effort-specific. Missing model allowances use the nearest usable same-plan agentic anchor and the exact Overall P50 ratio, labelled inferred. Missing anchors remain unavailable. Fable on Claude plans requires its own retained tuple.

Claude Opus retains Wokey's real-mix Max 20x anchor of USD 7,482 per month, dated 8 October 2026. Pro and Max 5x scale by the retained SemiAnalysis estimate. Fable and OpenAI retain their recorded model allowances. Claude chat and Claude Code consume one shared pool. Their alternative allowances are never added. The retained monthly Opus estimates are USD 748.20 on Pro and USD 3,741 on Max 5x. Work on a model spends its tuple. Missing underlying measurements stay unavailable. A monthly edit changes only the recorded model, effort and workload's monthly amount and labels it a guess. Weekly amounts and reset-window evidence remain fixed. Invalid monthly amounts are rejected.

Sources include [Arena Agent Pareto](https://arena.ai/leaderboard/agent/pareto), [Arena cost method](https://arena.ai/blog/agent-categories-and-cost), [Wokey](https://wokey.ai/claude-max-limits) and [SemiAnalysis](https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x).

Joint weekly allocation

The weekly engine compares eligible combinations of up to four plans within budget, with one plan per provider. Recorded included access (R30) and required capabilities restrict those combinations. Tools and dictation costs reduce the plan budget. Local hardware contributes no allowance. Free plans can join sets at Light and Regular only. Heavy and Multi-agent sets exclude them. A free-only set qualifies at Light only. A free-only Regular set retains the sustained-use exclusion.

For each combination, one linear program allocates exact model and effort pairs to task sessions while applying every known monthly pool, weekly pool, model cap and reset window. A common served fraction preserves the forecast's task mix. Monthly constraints extrapolate weekly sessions by 52/12. A tuple's pool fraction governs shared-pool debit. A recorded separate model cap also constrains that model. An evidenced model window restricts its named effort. A shared window restricts every model using that pool. A plan that does no work is dropped from its set.

required_usage = forecast_usage * safety_multiplier.

Safety applies once to task targets. The safety input is separate from output headroom. Peak sessions describe fleet activity in the busiest recorded reset window. Its fraction of total sessions prices that window at the same task mix. The public boundary first validates the visitor's usage and agent count. Each comparison column then bounds that peak at its own weekly total. A workload edit that reduces the visitor total also reduces an excessive peak and reports the adjustment. This compatibility rule keeps one saved request valid across columns.

Known reset metadata without an amount stays visible as an unknown limit. Separate unmeasured model caps also remain visible. An unknown limit supplies no guaranteed capacity. App-chat and missing exact pairs appear with their exclusion reasons. Retained task scores, quality floors and lower-tier rules remain independent of these costs and volume settings.

Limiting headroom and days

Coverage is the fraction of the safety target allocated. Actual fit needs that allocation to serve the full forecast before reserve. Safety fit needs coverage of at least 0.999 with usable task tuples. Reported constraint debits distinguish the allocated forecast from the full forecast at the same served model mix. For coverage c, safety multiplier s and allocated constraint debit b:

full_forecast_debit = b / (s * c).

remaining_fraction = 1 - full_forecast_debit.

headroom = c * s / largest_allocated_constraint_debit.

Incomplete coverage extrapolates the served mix to the full forecast. Remaining fractions can be negative. A zero allocation leaves headroom unavailable. Per-model raw days use the largest relevant constraint on that model's serving plans: active_days * c * s / relevant_debit. Other models' separate caps and windows do not limit those days. A shared pool does. Displayed days cap at the declared active days and round to one decimal place. Window-limited days retain the declared burst assumption.

The covering rank gate requires safety coverage and available weekly evidence. The retained comparison orders covering rows first, then quality, with the cheaper set winning inside 0.5 quality points. The separate display marks headroom below 2 as tight. The later headroom tie-break policy remains held. The display labels headroom under 1 as runs out, 1 to under 2 as tight, and 2 or more as covers when the safety gate also passes.

Legacy requests and W5 sensitivity

Saved links without valid weekly settings retain the original monthly path. The explicit Use weekly forecast control copies legacy ratings above 1 into initial included volume and preserves exclusions at 1. Later volume and importance edits are independent. Invalid weekly links display a warning and offer that migration. Original defaults and legacy URL contracts remain intact.

Weekly links retain the legacy t ratings for compatibility. A valid t rating for the same task takes precedence over wk.task_mix importance when a link is restored. A weekly link without t uses its saved weekly importance. Restoring an older conflicting link and editing a priority synchronise both fields. Inclusion and volume retain their separate saved values. The direct JavaScript and Python public requests use the same task-rating precedence.

Legacy demand uses a 65/3-day working month and Opus-equivalent dollars. Its session buckets, AA allocation-cost ratios and monthly solver remain unchanged. Its ordinary background share adds rated unmeasured work once. The monthly solver checks its allocation against the separate five-day weekly gate.

The matrix's W5 column remains an explicitly labelled legacy dollar sensitivity. Its measured-task anchor depends on retained subscription allowances or an explicit W5 edit. Legacy background ratings scale that anchor. W5 supplies no ordinary weekly session count. A retained private volume sensitivity remains provisional and supplies no accepted universal transfer. Public weekly defaults contain no private calibration factor or private measured rows.

Value lens

The model comparison orders each task by quality, by Artificial Analysis cost per task, or by value: value = 0.8 * 100 * quality / best_quality + 0.2 * cost_score, where cost_score = 100 * (ln max_cost - ln cost) / (ln max_cost - ln min_cost) across the compared models with a cost per task.