$0.14 per million input tokens. $0.56 per million output tokens. That is the DeepSeek V4-Flash launch pricing as of late April 2026, and it is the cleanest single-axis disruption the AI market has seen this year. To put the number in context: GPT-5.5 input is 85x more expensive, Claude Opus 4.7 input is 107x more expensive, Gemini 3.1 Pro input is 78x more expensive. On output the gaps narrow but stay enormous — 107x for GPT-5.5, 134x for Opus 4.7, 78x for Gemini.

Cheap models exist. DeepSeek itself has been the cost leader through V3 and the V3.1 mid-cycle releases. What is different about V4-Flash is the capability-to-cost ratio. On the benchmarks that matter for production workloads, V4-Flash is not the floor — it sits in the middle of the field, ahead of every model priced under $1 per million input tokens, and the gap to frontier on a meaningful subset of tasks is small enough to flip vendor choice for builds that are price-sensitive.

This is not "cheap model for cheap workloads." This is a frontier-adjacent capability at a fraction of the frontier price, and it changes the cost math for a wide band of production AI usage.

The Pricing Receipt

DeepSeek V4-Flash on the official API:

- Input: $0.14 per million tokens - Output: $0.56 per million tokens - Context caching: 50% off cached input ($0.07 per million) - Off-peak window: Additional 50% discount 16:30-00:30 UTC

The off-peak window is the underweighted line item. For workloads that can run on a batched schedule rather than real-time — most RAG ingestion pipelines, most overnight content generation, most analytical batch processing — the effective input cost drops to $0.07 per million tokens and the output cost drops to $0.28 per million tokens. Against an off-peak Batch API run on Gemini 3.1 Pro ($5.50 input / $22 output post-discount), DeepSeek V4-Flash off-peak is roughly 78x cheaper on input and 78x cheaper on output. The compounding effect for any team running a meaningful batch workload is structural.

The comparison against the same-tier models — cheaper alternatives to the frontier — is also worth showing:

| Model | Input ($/M) | Output ($/M) | Notes | |---|---:|---:|---| | DeepSeek V4-Flash | $0.14 | $0.56 | Off-peak: $0.07 / $0.28 | | DeepSeek V3.1 | $0.27 | $1.10 | Prior gen | | Claude Haiku 4.5 | $1.00 | $5.00 | Anthropic's small model | | Gemini 2.5 Flash | $0.30 | $2.50 | Google's small model | | GPT-5 mini | $0.15 | $0.60 | OpenAI's small model | | Llama 3.3 70B (via Together) | $0.59 | $0.79 | Open weights | | Mistral Medium 3 | $0.40 | $2.00 | Mistral's small model |

V4-Flash matches GPT-5 mini almost exactly on input ($0.14 vs $0.15) and edges it on output ($0.56 vs $0.60). The pricing convergence with GPT-5 mini is the more interesting data point than the gap to frontier. OpenAI has held GPT-5 mini at $0.15/$0.60 through the GPT-5.5 launch, and the fact that DeepSeek matched the number rather than undercut it suggests both labs see the same floor for sustainable small-model pricing.

The Capability Question

Cheap pricing only matters if the capability bar is high enough to absorb production workloads. The V4-Flash benchmark numbers:

| Benchmark | V4-Flash | GPT-5 mini | Haiku 4.5 | Gemini 2.5 Flash | Opus 4.7 (ref) | |---|---:|---:|---:|---:|---:| | MMLU-Pro | 79.4% | 76.2% | 74.8% | 72.9% | 87.3% | | GPQA Diamond | 71.2% | 64.8% | 59.4% | 58.7% | 89.1% | | SWE-bench Verified | 64.7% | 59.3% | 55.1% | 52.4% | 80.8% | | SWE-bench Pro | 41.8% | 36.4% | 32.7% | 30.1% | 64.3% | | Terminal-Bench 2.0 | 54.9% | 49.2% | 44.6% | 42.8% | 71.4% |

V4-Flash is the strongest small model on every benchmark in the table. The lead over GPT-5 mini ranges from 3 to 7 points across the board. Against Haiku 4.5 and Gemini 2.5 Flash the gap is wider, 5 to 12 points. The gap to frontier (Opus 4.7 column for reference) ranges from 8 points on MMLU-Pro to 23 points on SWE-bench Pro.

The shape of the data matters. V4-Flash trades roughly 8-15 points of frontier capability for 78-107x lower pricing on most workloads. That tradeoff makes sense for a wide band of production use cases: customer support automation, content summarisation, RAG over moderate-complexity corpora, document classification, simple code completion, structured data extraction. The 8-15 point capability gap shows up in failure rate, not in inability — V4-Flash will close most of these tasks correctly, just at a slightly lower hit rate than Opus 4.7 or GPT-5.5.

The Workload Math

Take a representative production RAG workload: 50,000 input tokens per query, 800 output tokens per response, 100,000 queries per day. The per-day cost comparison:

- DeepSeek V4-Flash (off-peak): $700 input + $22.40 output = $722.40 per day - DeepSeek V4-Flash (peak): $1,400 input + $44.80 output = $1,444.80 per day - GPT-5 mini: $750 input + $48 output = $798 per day - Gemini 2.5 Flash: $1,500 input + $200 output = $1,700 per day - Claude Opus 4.7: $75,000 input + $6,000 output = $81,000 per day - GPT-5.5: $60,000 input + $4,800 output = $64,800 per day

The frontier models are not viable for a 100k-query-per-day RAG workload at standard pricing. The small models all are. V4-Flash and GPT-5 mini are within 10% of each other on cost, and V4-Flash is meaningfully cheaper on output-heavy workloads where the off-peak window applies.

For a team currently running on Gemini 2.5 Flash, switching to V4-Flash saves roughly $980 per day, or $358,000 per year on this representative workload. For a team currently running on a frontier model out of risk aversion, the savings are an order of magnitude larger.

Where V4-Flash Falls Short

Concede the gaps before the takeaway. V4-Flash is genuinely weaker than the frontier on three workload types:

Long-context retention beyond 128K tokens: V4-Flash supports up to 128K context but retention quality degrades above 64K. Opus 4.7 holds quality cleanly to 200K. For document analysis workloads pushing the context window, V4-Flash will start to miss details that a frontier model would catch.

Multi-step agentic workflows: The 54.9% on Terminal-Bench 2.0 is meaningfully behind the 71.4-82.7% range of the frontier. For agentic coding products, this is a hard wall — V4-Flash will fail at long-horizon tasks where frontier models succeed.

Reasoning-dense workloads: The 71.2% on GPQA Diamond is strong for a small model but 23 points behind Gemini 3.1 Pro. For graduate-level reasoning workloads, V4-Flash is not the right pick at any price.

The wins and losses are clean. Bulk inference, moderate-complexity tasks, large-volume batch processing — V4-Flash is the cost leader and the capability bar is sufficient. Hard reasoning, long-horizon agents, deep context retention — frontier models are still required.

The Strategic Read

DeepSeek's pricing strategy is the most aggressive in the market. The V4-Flash margins at $0.14/$0.56 per million tokens are tight even by the standards of small-model competition. Two readings on why:

First reading — capacity strategy. DeepSeek has scaled inference infrastructure aggressively through 2025 and Q1 2026, and the underutilised capacity is being monetised at near-cost to drive volume. The off-peak discount tiers support this read — the company is willing to give away inference at marginal cost to fill its hardware fleet during low-demand windows. This is a sustainable strategy if and only if the volume growth pays for the next generation of capacity build-out.

Second reading — share strategy. DeepSeek is positioning to take share from the small-model market segment that OpenAI and Anthropic have been monetising at higher margins. By forcing GPT-5 mini and Haiku 4.5 into a price war they cannot win profitably, DeepSeek captures the price-sensitive workloads first and then positions for the next tier up.

Both readings probably contain some truth. The combination of low headline price, off-peak discounts, and aggressive context-caching pricing reads as a company that wants to win the small-model lane decisively rather than maintain margin parity.

What The Cost Lane Looks Like Next

Two open questions for Q2 and Q3. First, do OpenAI, Anthropic, or Google respond? GPT-5 mini at $0.15 input is now narrowly more expensive than V4-Flash and significantly less capable on key benchmarks. The same pressure applies to Haiku 4.5 ($1.00 input, more expensive and weaker) and Gemini 2.5 Flash ($0.30 input, more expensive and weaker). The expected response is a price cut on at least one of these models inside the next two release cycles. The first to move likely sets the new floor.

Second, does DeepSeek extend the pricing strategy to frontier-tier capability? A hypothetical V4-Reasoning at frontier capability and a fraction of frontier pricing would be the more disruptive move. DeepSeek's R1 reasoning model in 2025 was meaningfully cheaper than frontier reasoning models and the playbook is established. If V4-Reasoning lands at GPT-5.5 capability and a tenth of the price, the entire frontier pricing structure becomes unstable.

The math residual: at $0.14 per million input tokens, V4-Flash makes a category of AI workload economically viable that was marginal at small-model frontier pricing six months ago. The new applications enabled by that economic shift — high-volume document processing, real-time multi-tenant inference, mass-scale RAG, ambient AI features running on every page view — are the second-order story this pricing change will write.