Claude Code – Claude Haiku 4.5 (thinking 32k) vs. Codex – GPT-5.6 Luna (high): geospatial agent comparison
Comparison between Claude Code – Claude Haiku 4.5 (thinking 32k) and Codex – GPT-5.6 Luna (high) on the Geospatial Agent Index, each benchmark, cost, time and token use, from the same tasks, agent and limits. Methodology.
Side by side
| Metric | Claude Code – Claude Haiku 4.5 (thinking 32k) | Codex – GPT-5.6 Luna (high) |
|---|---|---|
| Geospatial Agent Index | 58* | 84* (partial: attempt 1 on 46 of 488 tasks) |
| Earth-Bench V2 score | 51 | No data |
| GeoBenchX score | 53 | No data |
| GS-QA2 score | 69 | 84 |
| GeoAgentBench score (guided GIS workflow; outside the index) | 78 | No data |
| GeoAnalystBench score (guided GIS workflow; outside the index) | 74 | No data |
| Raster and remote sensing score | 65 | 82 |
| Vector and overlay score | 67 | 100 |
| Climate and time series score | 49 | No data |
| Spatial statistics and interpolation score | 23 | No data |
| Mapping and cartography score | 56 | No data |
| Recognising infeasible requests score | 53 | No data |
| Cost per task | $0.092 | $0.0083 |
| Time per task | 1.5 min | 1.7 min |
| Tokens per attempt | 313k | 126k |
| Turns per attempt | 10.6 | 9.4 |
| Attempts decided | 488 | 46 |
| Input price per 1M tokens | $1.00 | $0.20 |
| Output price per 1M tokens | $5.00 | $1.20 |
| Context window | 200,000 | No data |
| Image input | Yes | No data |
| Developer | Anthropic | OpenAI |
Charts
Partial coverage. Codex – GPT-5.6 Luna (high): Earth-Bench V2 incomplete; GeoBenchX incomplete; 46 of 221 GS-QA2 tasks. Grey placeholders identify these entries in the index, cost and output token charts; their height does not represent a value. Comparable bars appear when first-attempt coverage is complete. Their available results remain in the side-by-side table.
Geospatial Agent Index
Equal-weight average of the evaluation scores, 0 to 100 · Higher is better
First eight selected entries in metric order. Full results in table view.
Data table: Geospatial Agent Index
| Model | Creator | Geospatial Agent Index | Range while attempts are pending | Benchmarks covered | Coverage (attempts) |
|---|---|---|---|---|---|
| Claude Code – Claude Haiku 4.5 (thinking 32k) | Anthropic | 58* | No pending judgements; coverage incomplete; 153 results awaiting review | 3 of 3 | 488 of 722 scheduled attempts |
Cost per task
Average cost per task (USD) · Lower is better
First eight selected entries in metric order. Full results in table view.
How to read this chart
Not shown: partial coverage. 1 entries without a first attempt on every index task. What is still running is on the models page.
Data table: Cost per task
| Model | Creator | Cost per task | Coverage (attempts) |
|---|---|---|---|
| Claude Code – Claude Haiku 4.5 (thinking 32k) | Anthropic | $0.092 | 488 of 722 scheduled attempts |
Output tokens per task
Average output tokens per attempt, reasoning included · Lower is better
First eight selected entries in metric order. Full results in table view.
How to read this chart
Not shown: partial coverage. 1 entries without a first attempt on every index task. What is still running is on the models page.
Data table: Output tokens per task
| Model | Creator | Output tokens per task | Coverage (attempts) |
|---|---|---|---|
| Claude Code – Claude Haiku 4.5 (thinking 32k) | Anthropic | 7k | 488 of 722 scheduled attempts |
Frequently asked questions
Which scores higher on the Geospatial Agent Index, Claude Code – Claude Haiku 4.5 (thinking 32k) or Codex – GPT-5.6 Luna (high)?
Which is cheaper per task, Claude Code – Claude Haiku 4.5 (thinking 32k) or Codex – GPT-5.6 Luna (high)?
Which writes fewer output tokens per task, Claude Code – Claude Haiku 4.5 (thinking 32k) or Codex – GPT-5.6 Luna (high)?
Which has the larger context window, Claude Code – Claude Haiku 4.5 (thinking 32k) or Codex – GPT-5.6 Luna (high)?
More: Claude Code – Claude Haiku 4.5 (thinking 32k) · Codex – GPT-5.6 Luna (high) · all comparisons