Pricing Comparison: OpenAI vs DeepSeek - a comparison for AI product builders

OpenAI and DeepSeek anchor the two ends of AI token pricing. On the standardized workload below, OpenAI's flagship costs about 27x more per month than DeepSeek's, and the two providers have spent the past year moving their prices in opposite directions.

OpenAI vs DeepSeek at a glance

Takeaway: DeepSeek is cheaper on every axis. What OpenAI sells above that line is the model ceiling, the ecosystem, and Western hosting.

Model overview

OpenAI's current lineup is the gpt-5.6 family, launched July 9, 2026. gpt-5.6-sol is the flagship at $5 / $30, gpt-5.6-terra the mid tier at $2.50 / $15, and gpt-5.6-luna the value pick within the family at $1 / $6. The gpt-5.5 and gpt-5.4 generations stay on the card, with gpt-5.5-pro as the premium reasoning option and gpt-5.4-nano as the cheapest model at $0.20 / $1.25.

DeepSeek runs the shortest rate card in the market: two models. deepseek-v4-pro is the flagship at $0.435 / $0.87 and deepseek-v4-flash is the value pick at $0.14 / $0.28, both with 1M context and automatic caching that drops cache-hit input to fractions of a cent. The deepseek-chat and deepseek-reasoner aliases retire July 24, 2026. The consumer app is free.

Key features

Pricing snapshot

Prices verified July 13, 2026. Providers reprice often, see the update timeline below.

OpenAI API pricing

Standard tier, per 1M tokens, prompts up to 200k. Long-context rates where noted.

Batch API: roughly 50% off. Cached input: 10x cheaper than base input. Web search tool: $10 per 1k calls.

DeepSeek API pricing

Per 1M tokens.

The deepseek-chat and deepseek-reasoner aliases retire July 24, 2026. The consumer app is free.

Chat plans

What this costs at a real workload

The standardized scenario: 1,000 requests per day at roughly 1,000 input and 1,000 output tokens each, which works out to about 30M input and 30M output tokens per month. Monthly cost = 30 x input rate + 30 x output rate.

Tokenizer caveat for the Claude rows: Fable 5, Sonnet 5, and Opus 4.7+ use a new tokenizer that produces roughly 30% more tokens for the same text. Comparing per-token rates across providers understates Anthropic's effective cost per request by about that margin. Rate cards don't tell you what the invoice will say. The tokenizer does.

The gap holds at every tier. OpenAI's cheapest listed model on this workload, gpt-5.6-luna at $210, still costs 17x DeepSeek V4 Flash. With DeepSeek's automatic caching, repetitive-prompt workloads land even lower than the table shows.

The spread across all six providers is 143x between the most and least expensive model for the same workload. The full 15-row table is on the AI Pricing Index. For how per-token rates turn into unit economics, see token economics.

Timeline of past updates

OpenAI

• July 9, 2026: gpt-5.6 family launch. Sol at $5 / $30, terra at $2.50 / $15, luna at $1 / $6

• April 24, 2026: gpt-5.5 launch at $5 / $30, doubling flagship input over gpt-5.4

• April 9, 2026: ChatGPT Pro $100 tier added below the $200 tier

• April 2, 2026: Team plan renamed Business, $25/user/mo

• March 5, 2026: gpt-5.4 family launch at $2.50 / $15, doubling over gpt-5

• August 2025: GPT-5 launch at $1.25 / $10

Trajectory: Flagship input went $1.25 to $2.50 to $5.00 in under a year. 4x.

DeepSeek

• July 24, 2026 (scheduled): deepseek-chat and deepseek-reasoner aliases retired

• April 24, 2026: V4 launch. V4 Flash at $0.14 / $0.28 with 1M context, V4 Pro at $0.435 / $0.87

• September 29, 2025: V3.2-Exp launch, API prices cut more than 50%

• September 2025: V3.1 pricing consolidated, off-peak discounts removed

Trajectory: Down. Roughly 75% cheaper than 12 months ago while adding 1M context.

The AI Pricing Index

OpenAI logged 6 price events in 12 months with flagship input up 4x. DeepSeek logged 4, with prices down roughly 75% while adding 1M context. Downward repricing is still repricing: if your unit economics assume today's DeepSeek rates, a cut by them or a hike by OpenAI moves your margins either way.

The index counts documented list-price events per provider over the trailing 12 months, recounted monthly. Full methodology and all six provider timelines are on the AI Pricing Index.

What this means for your own pricing

If you build on OpenAI, DeepSeek, or both, this pricing complexity becomes your pricing complexity. Four problems show up regardless of which side you pick.

• The margin problem. Every user interaction has variable cost. A power user generating long responses on a flagship model costs 10-50x more than a casual user on a value model. Per-seat pricing doesn't see this. You need usage-aware billing to protect margins.

• The model mix problem. Products route different requests to different models, and the table above shows the cost range that creates inside a single product. Your billing system needs to know which model served which request and price accordingly.

• The credit translation problem. Many AI products abstract provider costs with credit-based pricing. That means maintaining a conversion layer from usage to credits to dollars. Every provider price event forces you to re-derive the conversion rates.

• The visibility problem. Finance needs margin by customer, by feature, by model, by month. If your metering system lives apart from your billing system, that reconciliation happens in spreadsheets.

The index above adds the repricing angle: these two providers logged 10 price events in 12 months between them. Each one forces you to re-derive your own margins, credit conversion rates, and FX exposure if your revenue and your model bill sit in different currencies. Solvimon runs that layer: metering per provider, credits and wallets as first-class primitives, rate cards you update without an engineering sprint. See Solvimon for AI.

Related

• Anthropic vs DeepSeek pricing. The same cost gap measured against Claude.

• Gemini vs DeepSeek pricing. The two price-aggressive providers, compared on the same workload.