DeepSeek V4.1 Flash: API Pricing, Benchmarks, Vision Support and V4 Pro Migration

By AiCybr Editorial · Published August 5, 2026 · Updated September 10, 2026

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026. The hosted model adds native visual understanding, keeps a 1 million-token context window and 384K maximum output, and is available through the new API model ID deepseek-flash. DeepSeek has also retired the previous V4 Flash and V4 Flash Vision Exp hosted models; their legacy API names temporarily route to V4.1 Flash.

The release changes pricing and production routing as well. DeepSeek lists V4.1 Flash at $0.15 per million uncached input tokens and $0.60 per million output tokens during off-peak periods, with cache-hit input at $0.003 per million tokens. Peak prices are twice those figures. DeepSeek plans to route requests sent to deepseek-v4-pro to V4.1 Flash after 12:00 Beijing Time on September 14, 2026 until V4.1 Pro is released, with those requests billed at V4.1 Flash rates.

DeepSeek's launch benchmark table reports 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, 65.4 on NL2Repo-Bench, 88.1 on CyberGym, 90.9 on GPQA Diamond and 63.9 on Humanity's Last Exam with tools. These are vendor-reported release results; production teams should validate the model on their own agent harness, tools and repositories.

DeepSeek V4.1 Flash at a glance

Sources: DeepSeek API changelog and DeepSeek Models & Pricing.

What changed from V4 Flash 0731

The September release changes the hosted architecture family, modality support, API identity and pricing compared with the July 31 V4 Flash 0731 update. DeepSeek describes V4.1 Flash as the smallest model in a new architecture family and gives it native multimodal visual understanding. The company has moved the primary hosted identifier to deepseek-flash and retired the previous Flash API generations.

The benchmark rows above reproduce DeepSeek's published results for each release. Some benchmark names or versions changed between releases, so direct deltas are only comparable where the benchmark version and evaluation setup match.

Official V4.1 Flash benchmark results

DeepSeek published the following launch results for V4.1 Flash:

The table spans general reasoning, coding, terminal, tool-use, cybersecurity and visual evaluations. The results remain vendor-reported release measurements, so production evaluation should use the exact harness, tool schemas, context management, execution policy, timeout budgets and retry logic used by the target workload.

API migration: use deepseek-flash

New integrations should use:

model = "deepseek-flash"

DeepSeek currently accepts the older identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp, with requests served by V4.1 Flash and billed at the Flash price. This compatibility routing gives existing applications time to migrate while keeping the hosted backend on the new generation.

For teams that pin behavior for evaluation or regulated workflows, the alias change deserves a regression test. A stable API string can now resolve to a different model generation, which can alter response style, tool-selection behavior, visual capability, latency and benchmark performance.

V4 Pro users have a September 14 migration event

DeepSeek says that after 12:00 Beijing Time on September 14, 2026, requests to deepseek-v4-pro will route to V4.1 Flash until V4.1 Pro becomes available. Those requests will be charged at V4.1 Flash pricing.

Production teams using V4 Pro should complete four checks before that cutover:

1. Run a representative regression set against V4.1 Flash.
2. Re-test tool calling, structured output and long agent loops.
3. Recalculate token budgets using the new Flash pricing.
4. Record the model-routing change in observability and audit metadata so behavior changes can be traced to the serving transition.

DeepSeek states that its internal and external testing found V4.1 Flash ahead of V4 Pro across performance, cost, speed and total task-completion time. That statement is a vendor claim; task-level acceptance rates and wall-clock completion should determine production migration decisions.

DeepSeek V4.1 Flash pricing

DeepSeek uses peak and off-peak API pricing. Current V4.1 Flash rates are:

For a workload with 1 million uncached input tokens and 100,000 output tokens, the listed token charge is approximately $0.21 off-peak or $0.42 at peak pricing. Cache reuse can reduce input cost sharply: the same request with a full one-million-token cache hit would cost approximately $0.063 off-peak, including 100,000 output tokens.

Cost planning should use cost per accepted task in addition to token price. Agent systems may spend more tokens on retries, reasoning and tool loops, while a higher first-pass success rate can reduce the number of calls needed to finish the job.

Native vision changes the Flash deployment model

V4.1 Flash supports visual input through the same current Flash model. This consolidates the previous hosted split between V4 Flash for text and V4 Flash Vision Exp for image understanding.

Developer workflows can use the same model identity for source code, screenshots, rendered UIs, diagrams, charts and other visual debugging evidence instead of routing visual requests to the retired experimental Vision endpoint.

DeepSeek's pricing table marks Vision support for deepseek-flash and no Vision support for the current deepseek-v4-pro entry. Teams moving from Pro to V4.1 Flash therefore gain a multimodal input path during the September transition.

Context, output and API features

DeepSeek lists a 1M-token context window and 384K maximum output for V4.1 Flash. The current API table also lists:

• JSON output;
• tool calls;
• Responses API support;
• an Anthropic-compatible API;
• chat prefix completion in beta;
• FIM completion in non-thinking mode;
• thinking and non-thinking operation;
• a 2,500-request concurrency limit for the Flash tier.

The one-million-token ceiling can accommodate large repositories, document collections and long agent trajectories. Actual retrieval quality and task success should be measured at the context sizes used in production. Large prompts also increase prefill work and can reduce the economic benefit of sending irrelevant material simply because the context window allows it.

Coding and agent implications

V4.1 Flash's launch results emphasize coding and tool-using workloads. Terminal-Bench 2.1 rises from DeepSeek's published 82.7 for V4 Flash 0731 to 90.6 for V4.1 Flash, while the launch table lists 74.2 on DeepSWE v1.1 and 65.4 on NL2Repo-Bench.

Useful production tests include:

• multi-file bug fixes with executable tests;
• repository-wide dependency migrations;
• shell and container troubleshooting;
• structured tool calls with invalid-input recovery;
• long-context repository navigation;
• screenshot-assisted frontend debugging;
• reviewer-agent passes over patches generated by another model;
• failure recovery after an initial tool or command error.

Measure final task success, first-pass success, wall-clock time, tokens consumed, valid tool calls, unrelated edits and human intervention. These metrics are more actionable than a single public benchmark score.

Self-hosting status and the old 284B/13B figures

The previous DeepSeek V4 Flash 0731 checkpoint remains important for self-hosting because DeepSeek published downloadable weights for that generation under the MIT license. Its documented architecture uses 284B total parameters with about 13B active per token, and local deployments require access to the full expert set even though only a fraction is activated for each token.

Those 284B/13B figures describe the previous V4 Flash generation. DeepSeek's September 10 V4.1 release page and current pricing documentation do not provide a V4.1 Flash parameter count or identify a new downloadable V4.1 checkpoint. Self-hosting plans should therefore keep the 0731 hardware guidance separate from the hosted V4.1 API until DeepSeek publishes corresponding V4.1 weights and deployment documentation.

For the older 0731 checkpoint, documented heterogeneous deployments have used very large system-memory pools and GPU acceleration. A 13B active-parameter count reduces per-token compute; the full 284B checkpoint still governs weight-storage capacity.

Deployment recommendations

Existing V4 Flash API users

Move application configuration to deepseek-flash, then run a compact regression suite. Legacy Flash aliases currently route to V4.1, so migration can be staged without an immediate endpoint redesign.

Existing V4 Flash Vision Exp users

Move to deepseek-flash and validate image tokenization, image transport, tool use after visual inputs and any application-side modality routing. The old Vision Exp model is retired at the hosted-service level.

Existing V4 Pro users

Treat September 14 at 12:00 Beijing Time as a serving-change checkpoint. Validate workloads before the routing change and track model identity in telemetry.

New coding-agent deployments

V4.1 Flash is a candidate when API cost, long context, tool support and visual input matter. Start with representative repositories and production tool schemas, then compare accepted-task cost and completion time against the current premium alternative.

Private or offline deployments

Use the published V4 Flash 0731 weights when that generation meets the requirement, or wait for DeepSeek to publish V4.1-specific local deployment artifacts before sizing hardware around the new model.

Bottom line

DeepSeek V4.1 Flash is a material platform change for existing DeepSeek users. It combines native visual understanding, a new API identity, a new peak/off-peak price schedule, a one-million-token context window and substantially higher vendor-reported agent benchmarks. The September 14 V4 Pro routing change also makes the release relevant to teams that never selected the Flash tier directly.

The immediate operational actions are to adopt deepseek-flash for new integrations, regression-test applications currently using legacy Flash aliases, and validate V4 Pro workloads before the scheduled routing change. Self-hosters should keep the older 0731 checkpoint specifications separate from V4.1 until DeepSeek publishes V4.1 weights and hardware guidance.

References

• DeepSeek API changelog — V4.1 Flash release
• DeepSeek Models & Pricing
• DeepSeek V4 Flash 0731 model card
• DeepSeek V4 technical report
• Reuters — DeepSeek launches V4.1 Flash, September 10, 2026

Updated September 10, 2026 with the V4.1 Flash release, current API identifiers, pricing, benchmark results and V4 Pro routing schedule.