DeepSeek vs OpenAI vs Anthropic API Pricing 2026
The Short Answer
DeepSeek wins on price. Anthropic wins on the top end. OpenAI wins on range. The gaps are much smaller than they were six months ago.
| Tier | DeepSeek | OpenAI | Anthropic |
|---|---|---|---|
| Budget | V4-Flash $0.22/$0.66 (off-peak) | Luna $0.20/$1.20 | Haiku 4.5 $1/$5 |
| Mid | V4-Pro $0.66/$1.98 (off-peak) | Terra $2/$12 | Sonnet 5 $2/$10 |
| Frontier | V4-Pro $1.32/$3.96 (peak) | Sol $5/$30 | Opus 5 $5/$25 |
| Premium | — | Sol Ultra $12.50/$75 | Fable 5 / Mythos 5 $10/$50 |
Per million tokens, verified August 17, 2026.
Cost Per Real Task
Headline rates mislead because workloads have different input/output ratios. At a representative 30K input / 5K output:
| Model | Provider | Cost | Notes |
|---|---|---|---|
| DeepSeek V4-Flash (off-peak) | DeepSeek | $0.0099 | Cheapest available |
| GPT-5.6 Luna | OpenAI | $0.0120 | Flat rate, no clock |
| DeepSeek V4-Flash (peak) | DeepSeek | $0.0198 | Loses to Luna |
| DeepSeek V4-Pro (off-peak) | DeepSeek | $0.0297 | Cheapest frontier |
| Claude Haiku 4.5 | Anthropic | $0.0550 | |
| DeepSeek V4-Pro (peak) | DeepSeek | $0.0594 | |
| Claude Sonnet 5 | Anthropic | $0.1100 | Increase cancelled |
| GPT-5.6 Terra | OpenAI | $0.1200 | Cut 20% Jul 30 |
| Claude Opus 5 | Anthropic | $0.2750 | 1M ctx, 128K output |
| GPT-5.6 Sol | OpenAI | $0.3000 | |
| Claude Fable 5 | Anthropic | $0.5500 | |
| GPT-5.6 Sol Ultra | OpenAI | $0.7500 |
The spread from cheapest to most expensive is 76x. Choosing the right tier matters far more than choosing the right vendor.
The Cache Layer
For workloads with a large reusable prefix, cache pricing dominates and reorders everything:
| Model | Base input | Cache hit | Discount |
|---|---|---|---|
| DeepSeek V4-Flash (off-peak) | $0.22 | $0.007 | ~97% |
| DeepSeek V4-Pro (off-peak) | $0.66 | $0.022 | ~97% |
| Claude Sonnet 5 | $2 | $0.20 | 90% |
| Claude Opus 5 | $5 | $0.50 | 90% |
Anthropic’s structure is a clean multiplier: 0.1x for reads, 1.25x for a 5-minute cache write, 2x for a 1-hour write. It’s predictable and it stacks with the Batch API discount. DeepSeek’s discount is deeper but was repriced up to 12x on August 16, 2026 — proof that a deep cache discount is a vendor decision, not a law. See prompt caching explained.
Where Each Lab Actually Wins
DeepSeek wins on cost per token, conditionally. Since August 16, 2026 the V4 family runs peak/off-peak: peak is 01:00-04:00 and 06:00-10:00 UTC, and the other 17 hours are half price. For a North American team the entire business day is off-peak, so the advertised “price increase” barely lands. Both models carry a 1M context and a 384K max output — the largest output ceiling of the three labs. Flexible reasoning effort (low/high/max) is a genuine cost lever. DeepSeek ships both OpenAI Responses API and Anthropic-format endpoints, which makes it unusually easy to trial. The blocker for many organisations is not price but hosting jurisdiction.
OpenAI wins on range and rate stability. Five tiers from Luna at $0.20/$1.20 to Sol Ultra at $12.50/$75 mean you can route within one vendor and one integration. The July 30, 2026 cuts — Luna down 80%, Terra down 20% — reset the budget tier, and flat pricing with no peak window is worth real engineering time. GPT-5.6 Sol’s Ultrafast tier (previewed August 13, 2026, up to 750 tokens/sec on Cerebras hardware) addresses latency, which is a dimension price tables ignore.
Anthropic wins on the top of the range and predictability. Opus 5 at $5/$25 undercuts GPT-5.6 Sol’s $5/$30 on output while offering a 1M context and 128K max output — roughly double most rivals’ output ceiling, which matters when the deliverable is a large refactor rather than an answer. Anthropic also just did something rare: it cancelled the scheduled September 1, 2026 rise of Sonnet 5 to $3/$15, making $2/$10 permanent.
Three Modifiers That Change Your Bill
1. Tokenizers differ. Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text than Sonnet 4.6 and earlier. A $2/MTok model with a 30% heavier tokenizer bills like $2.60 against a lighter one. Never compare headline rates across tokenizer generations without adjusting.
2. Data residency costs 10%. Anthropic applies a 1.1x multiplier to every token category when you request US-only inference via inference_geo: "us". Bedrock and Google Cloud regional endpoints carry a similar 10% premium. If compliance pins your region, add 10% to every Anthropic number above.
3. Peak hours are timezone-dependent. DeepSeek’s schedule is a discount for Western buyers and a tax for Asian ones. Same rate card, opposite outcomes. See off-peak pricing explained.
The Decision Framework
- Cheapest possible tokens, Western hours, no residency constraint → DeepSeek V4-Flash off-peak.
- Cheap, flat, forecastable, no clock management → GPT-5.6 Luna.
- Best mid-tier value → Claude Sonnet 5 at $2/$10, now permanent, or GPT-5.6 Terra at $2/$12 if you want one vendor across all tiers.
- Long-horizon agent work where failure is expensive → Claude Opus 5: 1M context, 128K output, 90% cache discount.
- Need many tiers behind one integration → OpenAI. Five price points, one API.
- Enterprise with residency or jurisdiction constraints → Anthropic or OpenAI; price is not your binding constraint.
- Whatever you choose → measure cost per completed task, not per token. A cheap model needing three attempts is expensive.
Last verified: August 17, 2026. Prices from official vendor pricing pages.