Key findings
- The frontier runs from Qwen3.8 Flash on OpenCode ($0.049 a task, 832 points) up to Claude Opus 5.5 on Claude Code ($2.60, 1335 points).
- GPT-6.1 Sol on Codex reaches 1187 for $0.47 a task. Going from there to the top costs 5.5 times as much for 148 more points.
- DeepSeek Flash on Codex is the cheapest pairing above 1000, at $0.11 a task.
- The two most expensive pairings on the board, both Claude Fable 5.1, are off the frontier: Claude Opus 5.5 on Claude Code scores higher for less.
The pairings on the frontier
Sorted from cheapest to most expensive, each pairing below scores higher than every pairing that costs less.
| Model | Harness | Cost per task | Quality | Overall place |
|---|---|---|---|---|
| Qwen3.8 Flash | OpenCode | $0.049 | 832 | 42 |
| DeepSeek Flash | Codex | $0.113 | 1031 | 20 |
| DeepSeek Flash | Claude Code | $0.183 | 1039 | 19 |
| GPT-5.6 Terra | OpenCode | $0.425 | 1105 | 15 |
| GPT-6.1 Sol | Codex | $0.472 | 1187 | 9 |
| GPT-6 Sol | OpenCode | $0.503 | 1190 | 8 |
| GPT-6.1 Sol | OpenCode | $0.514 | 1196 | 7 |
| GPT-6 Astra | Codex | $1.282 | 1235 | 5 |
| Claude Opus 5.5 | Claude Code | $2.600 | 1335 | 1 |
For every other pairing, at least one pairing in this table scores as high or higher for less money.
What each step up costs
The frontier falls into three bands, and the price of extra quality rises sharply between them.
- Under $0.20. DeepSeek Flash on Codex scores 1031 for $0.11, roughly 200 points above Qwen3.8 Flash for a little over twice the money. For bulk work where a decent first draft is enough, this is the cheapest place to start.
- Around $0.50. GPT-5.6 Terra, GPT-6.1 Sol and GPT-6 Sol sit between $0.42 and $0.51. GPT-6.1 Sol on Codex adds 156 points over DeepSeek Flash on Codex for 4.2 times the cost, and is also among the fastest pairings at 500.7 seconds a run.
- Over $1. GPT-6 Astra on Codex adds 48 points over GPT-6.1 Sol on Codex for 2.7 times the cost. Claude Opus 5.5 on Claude Code adds another 100 points for twice the cost again.
In win-rate terms, Claude Opus 5.5 on Claude Code is expected to beat GPT-6.1 Sol on Codex in about 70% of same-task comparisons. Whether that is worth 5.5 times the spend depends on what a weaker result costs you to fix.
Expensive pairings that are not on it
Several pairings cost as much as the leaders without matching them:
- Claude Fable 5.1 on Claude Code is the most expensive pairing at $3.66 a task, and scores 1320, below Claude Opus 5.5 on the same harness at $2.60.
- Claude Opus 5.5 on OpenCode costs $3.17 for 1298, more money and fewer points than on Claude Code.
- GLM-5.3 Prime on Claude Code costs $2.51 for 1021. GLM-5.3 on OpenCode scores 993 for $0.47.
- Kimi K3 on Claude Code costs $2.46 for 1095. The same model on OpenCode costs half as much and scores 1147.
- Claude Sonnet 5 on Claude Code costs $1.62 for 906, while GPT-6.1 Sol on Codex scores 281 points more for less than a third of the price.
In two of these five, Claude Opus 5.5 and Kimi K3, the same model does better on another harness. The harness analysis looks at that in detail.
The time frontier looks similar
Ranking by time instead of cost gives almost the same list. Claude Haiku 4.5 on Claude Code is the fastest pairing at 387.6 seconds a run but scores 681. From GPT-5.6 Terra on Codex (457.4 seconds, 1098) up to Claude Opus 5.5 on Claude Code (746.5 seconds, 1335), every pairing on the time frontier is also on the cost frontier or a harness away from it.
The fastest pairings in the top ten are GPT-6.1 Sol on Codex at 500.7 seconds and GPT-6 Sol on OpenCode at 544.0. Claude Opus 5.5 on Claude Code takes about 50% longer than GPT-6.1 Sol on Codex.
Costs differ a lot by field
The overall board averages very different kinds of tasks. Per-task cost depends heavily on the field:
| Field | Top pairing | Its cost | Cheapest pairing | Its cost |
|---|---|---|---|---|
| 3D | Claude Opus 5.5 · Claude Code | $17.73 | DeepSeek Flash · Claude Code | $0.157 |
| Legal | Claude Fable 5.1 · Claude Code | $4.76 | Qwen3.8 Flash · OpenCode | $0.047 |
| Presentations | Claude Fable 5.1 · OpenCode | $4.41 | MiniMax M2.5 · Codex | $0.094 |
| Website | Claude Opus 5.5 · Claude Code | $3.41 | Qwen3.8 Flash · OpenCode | $0.043 |
| Medical | Claude Opus 5.5 · Claude Code | $0.57 | DeepSeek Flash · Codex | $0.004 |
| Data analysis | Claude Fable 5.1 · Claude Code | $0.22 | Qwen3.8 Flash · OpenCode | $0.005 |
In 3D work the top pairing spends $17.73 a task, so the gap to a cheaper option matters more: GPT-6.1 Sol on Codex places fifth in 3D at $3.70. In data analysis even the top pairing costs $0.22, and the choice can rest on quality alone.
For field-by-field picks, see the best AI agent for each field.
Questions
What is the cheapest AI agent that is still good
On the November 2026 data, DeepSeek Flash on Codex is the cheapest pairing scoring above the 1000 reference, at $0.11 a task and 1031 points. GPT-6.1 Sol on Codex is the cheapest pairing in the overall top ten, at $0.47 and 1187 points.
Is the most expensive AI agent the best
No. The most expensive pairing, Claude Fable 5.1 on Claude Code at $3.66 a task, places second overall. Claude Opus 5.5 on Claude Code places first for $2.60.
What does cost per task include
Only the model usage of the pairing itself, priced at each vendor's official token prices including cache rates. It leaves out grading, subscriptions and infrastructure. The methodology has a worked example.
Figures in this article come from the leaderboard data of 2 November 2026. The leaderboard shows any changes since.