How much did DeepSeek V4 Pro pricing drop?

Output tokens dropped from $3.48 per million to $0.87 per million, a 75 percent reduction. Input (cache miss) dropped from $1.74 to $0.435. Cache hits dropped from $0.0145 to $0.003625. The cut was made permanent on May 22, 2026 after starting as a promotion that was scheduled to expire May 31.

How does new DeepSeek V4 Pro pricing compare to Claude Opus 4.8 and GPT-5.5?

DeepSeek V4 Pro output at $0.87 per million is roughly 11 times cheaper than GPT-5 output ($10 per million) and 28 times cheaper than Claude Opus 4.8 retail ($25 per million). Gemini 3.5 Flash remains cheaper than DeepSeek on input ($0.15 vs $0.435) but closer on output ($0.60 vs $0.87). DeepSeek's 1.6 trillion parameter frontier model is now the cheapest large-model option on the market.

Should I switch from Claude or GPT to DeepSeek?

For high-volume bulk work like content drafts, page generation, and SEO research, yes, the cost savings justify the swap. For factcheck-sensitive AI news, security-critical code, or anything that gets shipped to a client without a second review, no, you still want Claude Opus or GPT-5.5 as the final pass. The smart play is a tiered stack: DeepSeek does the bulk, premium models do the verification.

Does this make Ollama Pro obsolete?

No. Ollama Pro at $20 per month is still cheaper for any workload over roughly 23 million output tokens of V4 Pro per month, and it covers 12 other cloud models in the same flat fee. For variable workloads or solo developers, direct API at the new price is competitive. For agencies running multiple parallel agents, the subscription still wins on predictability.

DeepSeek frames it as an efficiency gain, not a competitive response. They cite a hybrid attention architecture that reduces V4 Pro compute per inference to 27 percent of the prior generation and memory to one tenth, making the price cut commercially viable. Industry analysts at InfoWorld read the timing as an escalation of margin pressure on OpenAI and Anthropic regardless of how DeepSeek positions it.

AIPricingSEOMay 28, 2026by Greg9 min read

DeepSeek V4 Pro Just Cut Output Pricing 75 Percent. What It Means for the AI Stack.

DeepSeek V4 Pro pricing cut 75 percent, May 2026

Bottom line

DeepSeek made its V4 Pro 75 percent price cut permanent on May 22, 2026, four months earlier than expected. Output tokens dropped from $3.48 to $0.87 per million. The 1.6 trillion parameter frontier model is now the cheapest large model on the public API market. For SEO and content workloads, that changes the economics. For security-sensitive code review or AI news writing, it does not. The right answer is a tiered stack, not a wholesale switch.

Two news cycles collided this week. On May 22, DeepSeek made its V4 Pro promotional pricing permanent, four months earlier than the discount was set to expire. On May 28, Anthropic shipped Claude Opus 4.8. The two releases tell the same story from opposite ends. AI inference is getting cheaper faster than most agencies have repriced their work, and the gap between premium and budget models matters less every quarter.

Here is what actually changed at DeepSeek, what it does to the cost stack you are probably running, and the parts that get easier (and harder) for SEO, content, and developer workflows.

What DeepSeek Actually Did

According to Engadget's coverage and the official DeepSeek pricing page, V4 Pro pricing now sits at:

$0.87 / M

DeepSeek V4 Pro output token pricing as of May 22, 2026. Down 75 percent from the prior $3.48 per million. The same cut applies to input cache misses ($1.74 to $0.435) and cache hits ($0.0145 to $0.003625, which is one quarter of the original discounted rate).

The drop started in April as a promotional rate scheduled to revert on June 1, 2026. (We covered the original promo in Sub-as-API: The Quiet Pricing Shift in AI Agents.) Then on May 22, DeepSeek announced the new rate is permanent.

The justification, per InfoWorld, is architectural. DeepSeek says V4 Pro's hybrid attention mechanism reduces compute per inference to 27 percent of the prior generation and cuts memory use to one tenth. That is the technical narrative. The strategic narrative is harder to ignore: this happened the same week Anthropic shipped a flagship at the same price point as the prior Opus, and within four months of OpenAI shipping GPT-5.5. The Next Web's pricing comparison puts V4 Pro's new range ($0.003625 to $0.87 per million tokens) below all three major US frontier providers at their current published rates.

The New Cost Stack, Side by Side

This is the comparison that actually matters. All prices per million tokens, USD, retail API rates.

DeepSeek V4 Pro (new)

Input (cache miss): $0.435
Input (cache hit): $0.003625
Output: $0.87
Context: 1M tokens

DeepSeek V4 Flash

Input (cache miss): $0.14
Input (cache hit): $0.0028
Output: $0.28
Context: 1M tokens

Claude Opus 4.8

Pricing held at same level as Opus 4.7 per TechCrunch. Fast mode is 2.5x faster and 3x cheaper than the prior generation.

Gemini 3.5 Flash

Input: $0.15
Output: $0.60
Cheaper than V4 Pro on input. Closer on output. Smaller context window.

The headline read: DeepSeek V4 Pro is now the cheapest large model on the public market. The full read: Gemini Flash is still cheaper for small-context retrieval and quick text tasks, V4 Flash undercuts everyone on bulk, and Claude and GPT remain the premium picks where the cost per token matters less than the reliability per output.

What Changes for SEO and Content Work

This is where the price drop bites first. Most SEO workflows are token-heavy on the input side (long-context page audits, competitor analysis, SERP scraping ingestion) and token-medium on the output side (briefs, drafts, schema generation). At the old DeepSeek rate, a 50-page programmatic SEO build came out around $40 in API cost. At the new rate it comes out around $10. (We documented the underlying multi-model workflow in Multi-Model AI Workflow for Real Estate SEO.)

Three concrete shifts I am making in client work this week:

1. Bulk content drafting moves to V4 Pro by default. First drafts of service pages, location pages, and FAQ sections were already being routed through DeepSeek under the promo. The permanent rate makes this stick. Claude still gets the last pass on anything that publishes under a client byline, because the honesty improvements in Opus 4.7 and now 4.8 have an actual measurable impact on factcheck rubric scores. But the bulk gets cheaper.

2. Competitor SERP analysis goes wide instead of deep. When the cost of running a competitor SERP through a frontier model drops 75 percent, the right move is not to bank the savings. The right move is to run more analyses. Look at 50 SERPs instead of 10. Get distribution data instead of point estimates.

3. Programmatic SEO becomes viable at smaller scale. Programmatic only made economic sense at scale because API costs were the variable expense holding it back. At $0.87 per million output tokens, a 200-page programmatic build is closer to $25 in compute. Service businesses that could not justify programmatic before can now.

What Changes for Developer Workflows

For coding work, the story is more nuanced. The MacRumors SWE-Bench Pro number on Claude Opus 4.8 is 69.2 percent, which beats GPT-5.5 and Gemini 3.1 Pro on agentic coding. DeepSeek V4 Pro does not publish a current SWE-Bench Pro number that beats those. So the right swap is workload-specific, not blanket.

Use DeepSeek V4 Pro For

Bulk codegen, scaffolding, documentation writing, test case generation, first-pass refactors, large-context code review of legacy systems.

Use Claude / GPT-5.5 For

Security-sensitive code review, production diffs, architectural decisions, anything that ships without a second human review.

Verify Across Vendors

When DeepSeek writes the code, Claude or Codex should review. Same-family verification (DeepSeek API checking DeepSeek output) does not catch the failure modes you actually need to catch.

The cross-vendor rule matters more at the new price. When DeepSeek is cheap enough to run on every commit, you have to actively prevent it from also being the reviewer. The whole point of multi-model verification is the architectural diversity, not the model count.

Does This Kill Subscription Pricing?

Short answer: not for the workloads that justified subscription pricing in the first place.

I run on Ollama Pro at $20 per month, which gives unmetered access to V4 Pro, V4 Flash, qwen3-coder 480B, and ten other cloud models. The break-even at the new direct-API rate is roughly 23 million output tokens of V4 Pro per month. Below that, the API is cheaper. Above that, the subscription wins. For solo developers and small projects, direct API at the new rate is genuinely competitive now. For agencies running multiple parallel agents across multiple clients, the subscription still pencils out because the cost is flat and predictable.

The cost-model takeaway: Direct API rates are now cheap enough that a single-developer workload is fine on metered pricing. Agencies and high-volume workloads still want the flat-fee subscription, both for predictability and because the rate-limit headroom on Ollama Pro is closer to "unlimited" than the API gives you. The right answer depends on token volume, not model preference.

Where DeepSeek Still Loses

Three areas. Be honest about these before you swap.

Factcheck reliability. Our internal factcheck rubric catches AI-news posts written on DeepSeek V4 Pro at a higher rate than the same posts written on Claude Opus 4.8. Specifically, DeepSeek is more likely to extrapolate from a real source (claim 10,000 vulnerabilities when the source says "thousands") and more likely to merge two adjacent entities into a fictional one. This shows up most on time-sensitive content. The 4x improvement in code-flaw oversight Anthropic just shipped is the equivalent gain on the coding side. Pay the premium where it matters.

Tool-use orchestration. Claude Code's new dynamic workflows feature for running hundreds of parallel sub-agents is not yet matched on the DeepSeek side. If your agent topology is wider than one main and a couple sub-agents, the Opus 4.8 stack is meaningfully easier to coordinate.

Hosting and data residency. DeepSeek's hosting situation is what it is. For US client work with data-residency or contractual concerns, the right comparison is not just DeepSeek API vs Anthropic API. It is DeepSeek API vs an Anthropic-hosted Opus model with the AWS / GCP residency guarantees baked in. The per-token price gap shrinks once those are part of the equation.

What I Am Doing This Week

Three operational changes at MySEODesk this week, all driven by the new rate:

Re-routing bulk SEO content drafts to DeepSeek V4 Pro via the multi-model workflow. Claude Opus 4.8 stays as the final-pass verifier and factchecker. Saves roughly 70 percent on the per-page cost of programmatic builds.
Reopening programmatic SEO conversations with two service-business clients whose math did not pencil at the prior rate. At $0.87 per million output, a 150-page location-page build is no longer a five-figure project.
Patching the sub-as-API post to remove the stale "promo reverts June 1" claim. Permanent rate, permanent line edit.

What South Florida and Lynchburg Business Owners Should Do

If you are not running these workflows yourself, here is the short version of how this affects your invoice.

The agencies you hire just got their inference costs cut 75 percent on the bulk-content side. The ones that internalize that drop will pass it through to your line items. The ones that do not will pocket it. Ask the question directly when you renew or sign a new contract: which models is the agency using, and how has its cost basis changed in May 2026. A vendor that cannot answer that is operating on a 2024 cost model and pricing it like 2025.

If you are doing your own SEO, the practical move is to run your bulk-content tasks (initial briefs, FAQ drafts, first-pass schema markup) on DeepSeek V4 Pro via Ollama Pro or the direct API. Keep Claude or ChatGPT for the parts you actually ship under your business name.

The AI cost curve keeps bending. The agencies that quote for the new curve win the next two quarters of work. Get in touch if you want to know what your specific stack looks like at the new rate.

Back to Blog