DeepSeek

Why this family matters

DeepSeek ships two parallel lines that solve different problems. The V line (V3, V4) is the chat-and-tools default. The R line (R1) is the reasoning-default, with explicit chain-of-thought as a first-class product choice. Most teams pick one line and stay there; the failure mode is treating them as interchangeable.

The structurally interesting fact in our current index is V4: every V4 variant (Pro, Pro Thinking, Flash, Flash Thinking) ships with a 1M-token context window at the same headline price as the shorter-context tiers. That moves "long context" from a premium SKU decision in other families to a free axis here. V4 Pro Thinking lands at Quality Score 92.3 (#29 of 197 models we track), which puts it on the open-weights price-quality frontier against models that cost an order of magnitude more per token.

Which variant to start with

Default to deepseek-v4-flash for chat, summarization, and tool-augmented workloads where cost dominates. At $0.094 input / $0.188 output per million tokens with a 1M context window, it is the cheapest variant in our index that combines that context size with a usable quality tier (Flash at QS 75.7, Flash Thinking at QS 91.6). For most teams shipping API-backed product features, this is the practical default.

Step up to deepseek-v4-pro ($0.435 / $0.87 per million) when the workload visibly benefits from the additional headroom: harder reasoning, more aggressive tool-use, or evals that show measurable Pro vs. Flash deltas on the work you actually run.

When to deviate:

• Reasoning-heavy workloads: consider deepseek-r1 instead of V4 Pro Thinking. R1 is the family's explicit reasoning line; the mechanism behind the answer is different, and on workloads dominated by long chain-of-thought it can route to the right answer where a chat-default model loops. Compare on the specific reasoning benchmark that matches your workload before committing.
• Self-hosting on a single GPU: check the smaller R1 distills (R1 distilled into Llama 70B, Qwen 32B, Qwen 1.5B). They are not owned by this page (their detail data is filtered out of our public dataset) but they are the realistic self-host on-ramp if you cannot run the full V4 or R1 weights.
• Long-document RAG: V4's universal 1M context makes the variant choice within V4 a quality-vs-cost question rather than a context question. Start with Flash. The "do I need Pro for this document size" question collapses on this family because both tiers have the same window.
• You already use a closed flagship and want a price-anchor fallback: start with deepseek-v4-pro-thinking. At QS 92.3 it is the variant most likely to be a drop-in for a closed-flagship workload at substantially lower per-token cost. Run a side-by-side on your eval before committing.

Where the data is weak

We aggregate benchmark scores from multiple sources but coverage is uneven across this family. Specifically:

• V3 has the most variants and the messiest naming. Several V3 sub-versions (3.1, 3.1-terminus, 3.2, 3.2-exp, 3.2-speciale) coexist with different context windows (32K to 164K) and different prices. When in doubt, the slug (deepseek-v3 vs deepseek-v4) is the unambiguous identifier; treat the V3 minor versions as variant-on-variant rather than family-on-family.
• R1 coverage is thinner than V4. R1 in our index lists two variants (deepseek-r1-0528 and the Thinking variant), with benchmark depth that lags V4. Treat R1 scores as directional, particularly outside the headline benchmarks.
• Release dates are missing upstream. We are working on backfilling these; in the meantime, variant naming and effort tier are the reliable handles, not chronology.
• R1 distills are intentionally excluded from this page's variant table. The distilled checkpoints (into Llama 70B, Qwen 32B, Qwen 1.5B) are filtered out of our public detail dataset by policy, so this surface cannot show their per-variant rows. The right move if you are evaluating a distill is to test it on your own eval rather than rely on indirect benchmark coverage.
• Pricing on this page is the published API list price. Self-host economics are the dominant cost question for open-weights families; list price is a calibration anchor, not the cost ceiling.

If you are making a procurement decision, the variant table on this page is the load-bearing artifact. Cross-check pricing against DeepSeek's own docs before you commit.

When to reach for which alternative

• Long chain-of-thought reasoning is the binding workload: DeepSeek R1 already lives on this surface, but Claude Opus and full GPT-5 are the closed-flagship anchors to compare against on the specific reasoning benchmark that matches your workload.
• Workload demands deep coding-agent reliability: check qwen-3-coder-480b-a35b and openai-gpt-5-codex against V4 Pro Thinking on coding-flavoured benchmarks. DeepSeek V4 is competitive on general benchmarks but coding-specialised variants from other families typically lead on agentic-coding throughput.
• Open-weights breadth across model sizes is the requirement: the Qwen3 family ships dense models from 0.6B to 32B plus MoE variants, which gives you a wider spread of self-host budgets than DeepSeek. Pick the family whose smallest deployable variant fits your hardware budget, not the family with the highest top-end score.

Sources worth reading

• DeepSeek API docs: authoritative pricing, context windows, and variant identifiers
• DeepSeek on Hugging Face: model cards, weights, license terms
• DeepSeek-R1 paper: primary source for the R1 reasoning approach