Anthropic vs DeepSeek Pricing
DeepSeek is cheaper than Anthropic at the entry level: DeepSeek R1 0528 Qwen3 8B costs $0.100 per 1M output tokens against $1.25 for Claude 3 Haiku — 12.5× the price. On current flagships, Anthropic's Claude Fable 5.1 costs $50.00/1M output against $2.19/1M for DeepSeek's DeepSeek R1. Prices are USD, current as of September 21, 2026.
Per-million-token pricing for Anthropic and DeepSeek, with side-by-side flagship models, cheapest tiers, and context windows. Pricing data syncs weekly from a continuously-updated model catalog — last updated September 21, 2026.
Who wins on what
Cheapest input tokens
$0.02/1MDeepSeek
DeepSeek R1 0528 Qwen3 8B — $0.02/1M input
Cheapest output tokens
$0.10/1MDeepSeek
DeepSeek R1 0528 Qwen3 8B — $0.10/1M output
Longest context window
1.0MAnthropic
Claude 4 Sonnet (2025-05-14) — 1.0M input tokens
Lowest average output cost
$0.82/1MDeepSeek
Provider-wide average across 24 models
Largest model catalog
40 modelsAnthropic
More options to match cost vs capability
Cheapest cached input
$0.00/1MDeepSeek
DeepSeek V4 Flash — $0.00/1M cached read
Most reasoning models
5 modelsAnthropic
Models with dedicated reasoning / thinking support
Most vision models
5 modelsAnthropic
Models that accept image input
Open-weights available
YesDeepSeek
Offers open-weight models you can self-host
Side-by-side
Anthropic
Full Anthropic pricing →Cheapest input
$0.250
Claude 3 Haiku
Cheapest output
$1.25
Claude 3 Haiku
Longest context
1.0M
Claude 4 Sonnet (2025-05-14)
Avg output / 1M
$28.50
Across catalog
Cheapest cached input
$0.200
Claude Sonnet 5
| Model | In/1M | Out/1M | Ctx |
|---|---|---|---|
| Claude Fable 5.1 VisionReasoningToolsCache | $10.00 | $50.00 | 1.0M |
| Claude Opus 5 VisionReasoningToolsCache | $5.00 | $25.00 | 1.0M |
| Claude Sonnet 5 VisionReasoningToolsCache | $2.00 | $10.00 | 1.0M |
| Claude Fable 5 VisionReasoningToolsCache | $10.00 | $50.00 | 1.0M |
| Claude Opus 4.8 VisionReasoningToolsCache | $5.00 | $25.00 | 1.0M |
| Claude 3 Haiku | $0.250 | $1.25 | 200K |
DeepSeek
Full DeepSeek pricing →Cheapest input
$0.020
DeepSeek R1 0528 Qwen3 8B
Cheapest output
$0.100
DeepSeek R1 0528 Qwen3 8B
Longest context
1.0M
DeepSeek Flash
Avg output / 1M
$0.817
Across catalog
Cheapest cached input
$0.0030
DeepSeek V4 Flash
| Model | In/1M | Out/1M | Ctx |
|---|---|---|---|
| DeepSeek V4 Flash VisionReasoningToolsCache | $0.150 | $0.600 | 1.0M |
| DeepSeek V4 Flash Vision Exp VisionReasoningToolsCache | $0.150 | $0.600 | 1.0M |
| DeepSeek V4.1 Flash VisionReasoningToolsCache | $0.150 | $0.600 | 1.0M |
| DeepSeek V4 Pro ReasoningToolsCache | $0.435 | $0.870 | 1.0M |
| DeepSeek R1 | $0.550 | $2.19 | 66K |
| DeepSeek R1 0528 Qwen3 8B | $0.020 | $0.100 | 33K |
All prices in USD per 1 million tokens. Showing top 6 models per provider, sorted by output cost.
Frequently asked questions
Is Anthropic or DeepSeek cheaper?
DeepSeek has the cheaper entry point at $0.10/1M output (DeepSeek R1 0528 Qwen3 8B — $0.10/1M output). Provider-wide, Anthropic averages $28.50/1M output against $0.817/1M for DeepSeek. Which is cheaper for you depends on which model tier your workload actually needs.
How much do Anthropic and DeepSeek cost per 1M tokens?
Anthropic starts at $1.25 per 1M output tokens (Claude 3 Haiku), with its current flagship Claude Fable 5.1 at $50.00. DeepSeek starts at $0.100 (DeepSeek R1 0528 Qwen3 8B), with DeepSeek R1 at $2.19. Input tokens cost less than output on both.
Which has the larger context window, Anthropic or DeepSeek?
Anthropic — Claude 4 Sonnet (2025-05-14) — 1.0M input tokens. For comparison, Anthropic's largest is 1.0M tokens (Claude 4 Sonnet (2025-05-14)) and DeepSeek's is 1.0M tokens (DeepSeek Flash).
Which has more reasoning models, Anthropic or DeepSeek?
Anthropic lists 5 reasoning models and DeepSeek lists 4. Reasoning models bill their internal thinking as output tokens, so a reasoning call costs several times a standard completion of the same visible length — compare them on total tokens billed, not headline rate.
Can I self-host Anthropic or DeepSeek models?
DeepSeek publishes open-weight models you can run on your own hardware; Anthropic does not. Self-hosting swaps per-token pricing for GPU-hour cost, which usually only wins above sustained high utilisation — below that, hosted inference is cheaper.
Should I switch from Anthropic to DeepSeek to save money?
Only if the cheaper model still meets your quality bar. Token price is one input; the ones that decide your bill are prompt size, response length, retries and how much conversation history you resend each turn. Model the switch against your real traffic before committing. Prices here are current as of September 21, 2026.
Related comparisons
Run the numbers for your workload
Calcaas multiplies per-token costs by your real usage patterns — inputs, outputs, retries, and conversation history — across both providers in one model.