Model picker
Your task.
The right model mix.
Set your workload and data requirements. Compare one model with a smaller first step and a final answer from your reference model, using published prices and dated evidence.
Start with your work
What do you need to do?
A few everyday details. Your comparison updates as you go. Models and prices as published by the providers on 9 Nov 2026.
Your comparison
For 1,000 documents to summarize a month: the cheapest model that meets your requirements is OpenAI: GPT-6 Luna at $1.08 a month. Your input stays under 272,000 tokens, so the standard price applies (this provider charges more above that length). A two-step pair OpenAI: GPT-6 Luna → DeepSeek: DeepSeek V4.1 Flash 1017 (peak-hour price) costs $1.94 per month instead of $2.84 per month for DeepSeek: DeepSeek V4.1 Flash 1017 alone (saves $0.90 per month).
≈7,300 input / 706 output tokens per item (10 pages × 700 + 300 instructions; includes 206 hidden reasoning tokens as measured on the route).
Estimated sizes and published prices. Meeting the requirements does not prove quality on your task; a shorter first step can lose detail.
Compared: 25 direct provider prices.
See models, prices & evidence ↓Advanced (tokens)
Start with your own words
Describe your task
Turn a description into draft parameters, then review every detail before comparing models.
Do not include personal data, confidential information or API keys. You can complete the manual form without sending a description.
The country and region filters below apply to the models you compare. They do not select where this description is processed.
Every control, in detail
Set your workflow
These are editable example inputs. Review the assumptions before comparing.
Hidden: 0 routes from providers whose terms restrict independent comparisons. Which providers ·
Cost & evidence, side by side
One model or a pair?
A shorter first step. Your chosen finish.
Step A · OpenAI: GPT-6 Luna
Reads and condenses, with brief · direct
Step A reads your whole input and writes a brief for step B. Its cost covers that brief, not your final answer, so it differs from running this model alone.
Your input stays under 272,000 tokens, so the standard price applies (this provider charges more above that length).
includes ≈206 hidden reasoning tokens per call, as measured on this route (reasoning low) on 2026-11-08
Published price source · checked 2026-11-09
↓ Pass the brief to the next step
Step B · DeepSeek: DeepSeek V4.1 Flash 1017
Final answer · direct
No long-context tier in published prices.
peak-hour price · Off-peak is cheaper. Published price source · checked 2026-11-05
Instruction following (IFEval subset): 98% strict · 2026-10-17 and 2026-10-22
Both routes passed the required filters and quality floor. Model labels do not establish task quality.
Measured in AU, CA, DE, GB, US: first answer token 3663 ms; output 162.9 tokens/s. The wait for the requested reasoning workload has not been measured. · 2026-11-08 · AU · Source · data (JSON)
Measured in AU, CA, DE, GB, US: first answer token 830 ms; output 153.1 tokens/s. Estimated call time: 4.15 seconds using the slowest measured country. Short test requests do not verify the wait for your input length or workload. · 2026-11-08 · GB · Source · data (JSON)
1,000 tasks / month · published list prices · calculated
No long-context tier in published prices.
Direct browser execution is unavailable for this endpoint.
Not available yet. Published results for this pair will appear here.
Compression can lose detail. The first model’s summary may omit details. The second model only sees that summary and your instructions. Check the pair on your own prompt before relying on it.
Calculated from the token counts, cache assumptions and published prices shown here. This is not measured savings or proof of equal output quality.
Methodology & calculation caveats (4)
1. Calculated cost, unverified task fit
The single call receives the full original input. Step B receives the brief and its own instructions. Savings compare these different workflows and do not establish equal quality.
2. Quality floor and reference model
A temporary, editorial list of current model snapshots applies until comparable quality measurements are available for each route. The data floor uses the lower of the last two eligible IFEval prompt-strict scores, at least 50%. This editorial threshold will be reviewed. IFEval and BFCL are test signals, not a guarantee for your task.
3. Token counts, reasoning and retries
Input and visible output use the confirmed token basis. Hidden reasoning uses model tokens. A zero reasoning allowance and a retry multiplier of one are editable assumptions. Unconfirmed token conversion or missing charges makes the cost unknown.
4. Cache and long-context prices
Cache savings require an eligible, unchanged prefix and a cache hit within the provider’s rules. A brief is not automatically cached. Long-context rates apply separately to each call before subtracting cached input. Sequential batch windows must fit the whole workflow.
All candidate routes
504 routes · 3 eligible pairs. The first routes are shown below; the full list is available in the comparison data.
| Model / route | Monthly single-call cost | Evidence and eligibility |
|---|---|---|
| OpenAI: GPT-6 Lunadirect · openai | $1.08 per month | Matches stated requirements Measured in AU, CA, DE, GB, US: first answer token 3663 ms; output 162.9 tokens/s. The wait for the requested reasoning workload has not been measured. · 2026-11-08 · AU · Source · data (JSON) includes ≈206 hidden reasoning tokens per call, as measured on this route (reasoning low) on 2026-11-08 Published input and output prices for this channel. · 2026-11-09 · SourcePrice observations: 2026-10-16 → 2026-11-09 |
| Qwen: Qwen3.8 Flashdirect · qwenOpen weightsRun it on your own GPUs | $1.36 per month | Matches stated requirements Measured in AU, CA, DE, GB, US: first answer token 1622 ms; output 62.2 tokens/s. Estimated call time: 9.87 seconds using the slowest measured country. Short test requests do not verify the wait for your input length or workload. · 2026-11-08 · AU · Source · data (JSON) Published input and output prices for this channel. · 2026-11-09 · SourcePrice observations: 2026-10-16 → 2026-11-09 |
| Anthropic: Claude Haiku 5.5direct · anthropic | $1.46 per month | Matches stated requirements Instruction following (IFEval subset): 82% strict · 2026-10-17 and 2026-10-22 · snapshot by vendor alias Measured in AU, CA, DE, GB, US: first answer token 653 ms; output 215.4 tokens/s. Estimated call time: 4.10 seconds using the slowest measured country. Short test requests do not verify the wait for your input length or workload. · 2026-11-08 · DE · Source · data (JSON) Published input and output prices for this channel. · 2026-11-09 · SourcePrice observations: 2026-10-19 → 2026-11-09 |
| DeepSeek: DeepSeek V4.1 Flash 1017direct · deepseek | $2.84 per monthpeak-hour price · Off-peak is cheaper. | Matches stated requirements Instruction following (IFEval subset): 98% strict · 2026-10-17 and 2026-10-22 Measured in AU, CA, DE, GB, US: first answer token 830 ms; output 153.1 tokens/s. Estimated call time: 4.15 seconds using the slowest measured country. Short test requests do not verify the wait for your input length or workload. · 2026-11-08 · GB · Source · data (JSON) Published input and output prices for this channel. · 2026-11-05 · SourcePrice observations: 2026-10-18 → 2026-11-05 |
| Google: Gemini 3.8 Flashdirect · google | $10.58 per month | Matches stated requirements Instruction following (IFEval subset): 98% strict · 2026-10-17 and 2026-10-22 · snapshot by vendor alias Measured in AU, CA, DE, GB, US: first answer token 8087 ms; output 206.4 tokens/s. The wait for the requested reasoning workload has not been measured. · 2026-11-05 · DE · Source · data (JSON) includes ≈847 hidden reasoning tokens per call, as measured on this route (reasoning low) on 2026-11-05 Published input and output prices for this channel. · 2026-11-02 · SourcePrice observations: 2026-10-16 → 2026-11-02 |
Showing 5 of 504 routes for the example inputs. Every route is included in the calculation. Enable JavaScript to compare and show all routes, or download all routes and evidence (JSON).
Prices compare the selected channel. A test location does not establish where your data is processed or stored.
Prices as shown in the United States on .