A clearer way
to choose.

Compare the tools you work with and the models they use.
Keep the facts, assumptions, and measurements separate.

What you're comparing

An agent manages context and tools. A model generates output. A billing route sets the rates. Look summarizes the models available through each coding tool. Compare lets you pin exact model and effort configurations. Aider and OpenCode are coding applications that run your chosen model; they are not models themselves.

The atlas includes models with documented API price scenarios. Coding tools and environments filter their documented routes. Subscription credits and enterprise compute units remain in agent profiles; they have no verified conversion to the token workload used here.

Provider integrations are documented at the agent level; exact model and effort support can still vary by version, account, and route. Unknown compatibility stays unknown.

From tokens to dollars

The default scenario is 10,000 uncached input and 2,000 billable output tokens. It's a pricing workload, not a measured coding task. Different tokenizers mean equal token counts may represent different text.

Cost = (input × input rate + cached reads × read rate
        + cache writes × write rate + output × output rate)
       / 1,000,000 + storage + paid operations

Rates are USD per million tokens. Cache reads and writes are disjoint; storage uses token-hours. Reasoning counts as output when billed that way—never add it twice. Example: 10k input at $2/1M and 2k output at $10/1M costs $0.04. Those are illustrative rates.

The ledger combines the calculator and comparison: edit your inputs and every row updates locally. Current coverage is standard text requests up to 100,000 input tokens, excluding tax and regional premiums. Missing rates, expired prices, required source-review dates, or uncovered tiers return N/A. Qwen Coder Plus/Flash base rates cover fewer than 32k total input tokens. DeepSeek peak and off-peak prices are explicit selections; this calculator does not choose a time window or holiday calendar automatically.

The small color bars show each value's position on its scale. Workload costs use a shared logarithmic $0.001–$3+ scale; input/output rates and relative indexes have separate axes. Tooltips identify units and limits. Bar positions are visual guides; sorting uses exact values.

Base index = model cost ÷ reference cost. A positive-cost reference is 1×. The default reference is the lowest valid workload cost in this catalog; you can choose another. Bands are Economical <5×, Moderate 5–<25×, Expensive 25–<100×, and Very expensive ≥100×; classification uses unrounded values.

Effort is a setting, not a score.

The atlas uses a cool-to-warm scale for estimated dollar cost, separate from reasoning settings. Supported levels come from model documentation; they are not a certification of every agent configuration. Native controls without a verified mapping stay separate.

The optional preset assumes None 0.8×, Low 1×, Medium 1.5×, High 2.5×, XHigh 4×, and Max 6× whole-workload consumption. These multipliers are illustrative, not measured.

Estimated index = (model cost × model effort multiplier)
                  / (reference cost × reference effort multiplier)

Compare separates identical-token pricing from an illustrative effort scenario. Its default Opus 4.8 High versus Opus 5.5 Medium example costs $0.10 versus $0.08 for the same 10k uncached input and 2k billable output tokens. The planning multipliers yield $0.25 versus $0.12; those are assumptions, not measured task costs. Effort names do not imply equivalent token consumption across model versions.

My token usage replaces the multiplier with your separate billable output counts for each configuration. Those counts include billable reasoning; it is not added again. The input workload is shared. A lower token quote does not establish a lower cost per verified task without completion and retry measurements.

Both sides are normalized, so the selected reference remains 1×. Higher effort does not establish better quality, success, or speed. An unsupported level is a dash; unknown support is labeled Unknown.

A catalog average, not an agent bill.

Look's default cost signal uses the arithmetic mean of compatible, currently priced catalog rows for 10,000 uncached input and 2,000 billable output tokens. Each row gets equal weight. Provider and billing-route filters recalculate that set. Separate routes, billing windows and older versions are separate price rows; this is not a traffic-weighted average or a prediction of which models you will use.

Catalog average = sum of included workload quotes / quote count

The optional models × effort view gives each documented configuration equal weight and applies the illustrative multipliers. A row with no mapped effort contributes its base token quote only. Unknown, expired or unverified-required prices are excluded from the average, never counted as zero. If nothing is priced, the average is unavailable.

Each spectrum uses the same logarithmic dollar scale, from $0.001 to $3+. Higher average costs produce taller signals. Colors identify fixed price bands; the signal profile emphasizes bands containing more quotes. This is a price visualization, not a frequency analysis. The detail histogram shows the share of quotes in each band. The six-segment strip shows how many included model price rows document each effort level, with stronger color for broader coverage. Missing mappings are not treated as unsupported controls. These graphics measure neither quality nor subscription usage. Tools with the same supported catalog can have the same average; their actual context growth, tool calls and retries still differ. Use Compare for an exact model-and-effort question, then open price details for raw rates and sources.

What the evidence can say

Published rates link to official sources with verification dates. API costs, subscription allowances, local running costs, estimates, and measured workflows are separate quantities. No token-to-subscription conversion is assumed.

There are no original benchmark results yet. Future comparisons will include objective task verification, complete failure/retry costs, mean time, task distribution, agent/model versions, sample size, and uncertainty.

Future efficiency formulas
AEI = 100 × (P/P₀) × (C₀/C)ʷ × (T₀/T)¹⁻ʷ
VTE = 1,000 × verified completions / total workflow cost
Cost per success = total workflow cost / verified completions

P is completion probability, C mean USD cost per workflow, T mean seconds, and 0 the matched reference. Default cost weight w is 0.75. AEI 100 equals the reference; VTE is verified tasks per $1,000. Positive denominators and comparable cohorts are required. No successful completions means measured cost/success is undefined.

Include failed runs and retries. First-attempt success is separate. Report Wilson success intervals and task-stratified bootstrap intervals for cost/time; a small repeated task set is not broad evidence. Historical spend remains separate from current-price simulations.

Reading the graphics

The home matrix bins model/billing-route scenarios by API cost for 10k input + 2k output tokens. Documented model efforts use the illustrative multipliers above; models without a mapped effort contribute one base-price scenario. Every quoted offer/effort has equal weight. Gateway prices and direct prices are kept distinct; peak/off-peak windows are separate examples, not a calendar- or traffic-weighted forecast. Taller bars and stronger arcs mean a greater share of catalogued scenarios in that cost bin, not more users or more real requests.

The same fixed dollar bins apply to every agent: under $0.003, then $0.003–0.01, $0.01–0.03, $0.03–0.10, $0.10–0.30, $0.30–1, $1–3, and $3+. Exact provider/model support still needs agent-version verification. Multi-provider agents cover a broader range; these are API examples, not subscription prices or agent operating costs. Explore's cost strip uses a logarithmic $0.001–$3 scale, with values outside it capped visually. Numeric prices keep full coverage.

Agent environment counts can overlap. Repository stars come from GitHub's public repository API , with source links and a snapshot date. They measure repository interest, not installed users, active users, task success, or token usage. A repository can cover several product surfaces. Bars share a linear scale; unavailable metrics remain unknown. Interest snapshots older than seven days are marked stale.

Sources & updates

Rates are checked against provider documentation or the billing route's own API. Failed updates keep the previous accepted data and date. Prices older than 72 hours are marked stale; freshness doesn't certify compatibility or performance.

Browsing and calculation need no provider API calls. Production analytics may send page views; Contact sends a form request only when submitted. See Privacy.

Inspect the source notes ↗ · Contact