Best value
The best coding intelligence at the lowest price.
The most score for the money: the point on the cost-vs-score frontier where extra dollars start buying less.
Cheapest qualifying
-
Fastest qualifying
-
Highest score
-
Score vs cost
FrontierCode, cost per task on a log scale
- Frontier
- Within noise
- Beaten
- Best value
- Top tier
- Effort curve
How to read this chart
- Large points are the frontier: nothing cheaper scores higher.
- Mid points are beaten on paper but within noise of the frontier.
- Grey points are clearly beaten by something no more expensive.
- Bars are 95% confidence intervals; the shaded band is the frontier's own.
- The ring is the best value: the frontier's knee, past which each extra dollar buys fewer points.
- The dashed ring is the top tier: the cheapest frontier point tied with the top score.
- Thin lines join one model's effort levels, low to max, in its provider's colour. Hover a name or point to isolate it.
- Click a point to pin its card, with links to its table row and its source.
Price ladder
Cheapest way into each score band
Cost to climb
Extra spend per task to reach the next model on the frontier
| Step up to | From | Score | + Points | Cost/task | + $/task | + % of top | $/pt |
|---|
Effort spread
Cost per task at each effort level on DeepSWE, , log scale · L low · M medium · H high · X xhigh · Mx max · N none
All variants
Sorted best value first. Click a row for provenance.
Each row is an upper bound across harnesses, not one configuration you can buy.
A variant's score and cost come from the best-performing harness on each benchmark, so one row can mix Claude Code on one benchmark with Codex on another. The harness behind every number is shown in its cell and in the expanded row.
| # | Variant | Effort | Provider | AA Coding | AA Intel. | FrontierCode | DeepSWE | T-Bench | SWE-Atlas | Cost/task | $/solved | Time | Pts/$ |
|---|
Each family keeps its newest generation, plus the previous one while the newest is still missing a benchmark (tagged newer: …).
Benchmarks
Excluded when two or more models score ≥ 85%, or when not updated in 30 days.
Method
How is “best value” picked?
- Every ranked figure, score and cost alike, comes from the one benchmark you select. Scores are raw percentages.
- Best value is the knee of the frontier. With cost and score scaled 0–1 from the cheapest rung to the top, it is the rung furthest above the straight line between them: below it a dollar buys more than the climb's average dollar, above it less.
- Top tier is the cheapest frontier variant statistically tied with the top: within the root-sum-square of both 95% intervals, or within 2 points where no interval is published. Best value is never dearer than it.
- Points per dollar is shown but is not the answer. It rewards whatever is cheapest, however weak.
- Time and steps are shown, never ranked on. Scores are never imputed.
What is a row, exactly?
A variant's score and cost are assembled from the best-performing harness on each benchmark. The row is the upper bound achievable across harnesses, not a single reproducible configuration.
Estimated costs and scores
- Costs inherited from another source are tagged est. and never enter a ranked figure.
- ~ scores are what an adjacent effort of the same model scored. Context only.
Why not just average the benchmarks?
The sources rarely test the same configuration and their scores sit on different scales. Averaging them would penalise the models measured most, so the page ranks on one benchmark at a time. Rows not run on it stay in the table, unranked.
Data freshness & refresh
Requires the maintainer's key.