DeepSeek
The cheapest option here by a wide margin and the only one that open-weights its actual frontier model — with the caveat that the hosted API stores data in China.
How to prompt it
Understand that these are inferred, not documented. DeepSeek publishes no prompting guidance, so what follows comes from its capability pages rather than from advice it gives.
Set thinking deliberately. It is on by default at high effort, and turning it off is a parameter rather than a prompt instruction.
Do not bother setting temperature while thinking is on. It is accepted and silently ignored, which means a prompt that seems not to respond to sampling changes is behaving as documented.
For JSON output, include the word json in your prompt and show an example of the shape you want. DeepSeek documents this as a requirement, not a suggestion.
Use assistant prefill to pin the start of a reply. It is one of the few strong format levers the platform offers.
Keep the front of your prompt stable across turns. Caching matches on exact prefixes, so a stable opening is worth a large discount as well as being good practice.
Strengths
- Dramatically the cheapest, with two structural discounts others do not offer: off-peak hours at half price, and cached input at a small fraction of full price.
- The only provider here publishing a real maximum output figure across its whole line, on a very large context window. For long generations that is the clearest specification available.
- Genuinely open weights under MIT — the most permissive licence of any provider on this site, and applied to the frontier model rather than a cut-down one.
- Three API surfaces, matching the shapes used by two other major vendors, so prompts written elsewhere port with little rework.
Weaknesses
- No prompting documentation at all. There is no prompt-engineering page and no structural guidance anywhere, so nothing here is guidance from DeepSeek about how to write for its models.
- Several failure modes are silent. Temperature and related sampling settings are accepted and ignored while thinking is on, an unrecognised model name reroutes rather than erroring, and the vendor acknowledges that structured output sometimes returns empty.
- DeepSeek’s own privacy policy states that personal data is collected, processed and stored in the People’s Republic of China. There is no EU residency option and no published processor agreement, which rules the hosted API out for a great deal of European commercial use. The open weights are the clean escape hatch.
- The model documentation gives no positioning, no benchmarks and no limitations section, so there is little to go on when choosing between its models.
Models
DeepSeek V4 Pro
1.6T total parameters with 49B active, the more capable tier.
- Context
- 1M
- Max output
- 384K
Good for: Hard work at a fraction of competitors’ prices, and very long generations.
Watch out: Hosted data is stored in China. Self-host the MIT weights if that is a problem.
DeepSeek V4 Flash
284B total parameters with 13B active, the fast tier.
- Context
- 1M
- Max output
- 384K
Good for: High volume at the lowest price of anything on this site, with the same context and output limits as the larger model.
Watch out: Considerably higher allowed concurrency than the Pro tier, which matters more than the price for batch work.
Sources and grounding
- 1Thinking modeDeepSeekThinking defaults, effort levels, and the sampling parameters that are silently ignored.
- 2JSON outputDeepSeekThe requirement to include the word json and an example, and the acknowledged empty-response issue.
- 3PricingDeepSeekContext, maximum output, and the off-peak and cached-input discounts.
- 4Privacy policyDeepSeekThe statement that personal data is processed and stored in the People’s Republic of China.