Use 1.5x Speed when waiting matters
Start with Normal Speed for background work that can wait. Time savings do not equal quota savings, and USD prices cannot stand in for subscription quota percentages.
CLI message
Speed comparison
Normal Speed and 1.5x Speed on the same prompt. Choose effort to compare models.
Balanced. Ready to use.
Start right away, or add optional packs to suit your work.
Balanced base
ActiveEnough capability, a reasonable wait, and value considered together
Ready to use. No extras needed.Add-on packs Optional extras
Using 0 / 2 slotsHow recommendations work
Start with a capability group suited to the task, then compare completion time and cost. There is no weighted score or fixed ranking by model name. Close choices, or choices with different trade-offs, can share a rank.
AA is a rough reference: the target is 47 for everyday work, 54 for hard work, 38 for cost-first work, and 35 for fast, affordable work. A score up to 2 points below the target is allowed. These are advisory groups, not our own coding-capability measurements.
Time differences within about 10% or 2 seconds, and cost differences within about 15% or $0.0002, are grouped as similar. Speed puts time first; lower cost puts price first. Selecting both shows trade-offs where neither choice wins on both dimensions. Each group is anchored to its lowest value.
The balanced default favors responses within 60 seconds and API-equivalent cost up to $0.06 per job. If a matched 1.5x Speed run saves less than 10% of completion time at higher cost, start with Normal Speed. Consider lower effort when its AA reference is within 2 points and it is materially faster and cheaper. This does not establish equal coding quality.
Prices are API equivalents, not subscription charges or quota percentages. Recommendations are recalculated from the selected data period and update as new rounds arrive.
Every variant, for your priorities · Normal Speed and 1.5x Speed rank separately · Select a column header to sort · Open “Why” under a model for its trade-offs
Scroll the table sideways to see time and cost ranks
Time and cost ranks compare every measured profile in the selected period before effort or mode filtering. Completion time includes CLI startup, waiting, reasoning, and output. Costs are API-equivalent estimates, not subscription charges. Advisor Max and Expert Low/High/Max each have one Mac run per effort. Advisor uses the supplied Qwen3.8 Flash rates; Expert uses the supplied LARNAPI off-peak rates. AA scores of Advisor Max 40 and Expert Low 40, High 43, Max 45 were supplied by the user. Missing rows do not borrow measurements from other efforts. Effort is the value requested through the CLI, not a backend verification unless reported.
Choosing for real programming work
Every model, across all efforts and modes
How to read these results
Every run uses the same prose prompt, but answer lengths differ. Consider completion time and characters/s alongside tokens/s, which can include reasoning and depend on each model's tokenizer. These measurements do not test code correctness, tool use, or revision count.
The a5090 Codex/Claude results appear alongside one Mac run per effort for LARNAPI/Advisor Max and Expert Low/High/Max. The a5090 feed remains separate from older Mac/VPS/API results. One sample does not establish a lasting trend. Requested Codex 1.5x Speed does not prove a backend verified 1.5x tier.
The nine base models have 35 AA Index references checked against Artificial Analysis v4.3.2 on 18 Oct 2026. Advisor Max 40 and Expert Low 40, High 43, Max 45 were supplied by the user. These are reference scores, not outcomes of this benchmark. Sonnet/Opus/Fable use AA’s “with fallback” labels, which do not prove identical pure effort settings in our CLI.
The analysis covers 39 model/effort profiles and 58 separate variants when Codex Normal Speed and 1.5x Speed are split and four LARNAPI variants are added. Each uses its own result; another effort’s result is never substituted. Lower effort is not guaranteed to finish faster on every task.
References: Artificial Analysis · OpenAI · Anthropic · Guidance reviewed 18 Oct 2026 · Ranks follow results and your priorities
Speed history
One chart per effort · Normal Speed color follows provider · 1.5x Speed is orange
Normal Speed1.5x Speed
Individual measurements
Select a timestamp to inspect waiting, streaming, and sample details
| Started | Effort | Mode | tokens/s | chars/s | Completion | Status |
|---|
Models shownChoose a model
Efforts shownChoose effort
Combined chart
Color / model
Normal Speed / Effort
1.5x Speed / marker
★ Star
Models are never joined together. Lines connect only the same model, effort, and mode. Select a point to inspect that run.
Selected series
Mini charts use their own scales to show each series' trend. Compare numeric values across series.