DeepSeek V4.1 Flash: Benchmarks, Prices and Pro Cutoff

DeepSeek V4.1 Flash is now available with native image understanding, much cheaper API rates and a published 19-benchmark comparison. For teams replacing V4 Pro, the immediate deadline is September 14 at 04:00 UTC: Pro requests will then switch to Flash. Use the remaining window to compare your actual work, because the release results favor Flash on many agent tasks but still favor Pro on two reasoning tests.

This September 10 update incorporates DeepSeek’s official announcement, technical report, model repository and revised API documentation. It replaces the earlier notice-only analysis with confirmed launch details, complete vendor scores and the evaluation settings needed to interpret them. Earlier security and design tests remain separately identified.

The September 10 changelog confirms the production model name deepseek-flash. V4 Flash and V4 Flash Vision Exp have been retired. Their aliases, deepseek-v4-flash and deepseek-v4-flash-vision-exp, temporarily route to V4.1 Flash at Flash prices. Keeping an old alias therefore does not preserve the old model.

V4 Pro has a separate transition: deepseek-v4-pro remains available at its existing prices until noon Beijing time on September 14, equivalent to 04:00 UTC. From that point, requests will route to V4.1 Flash and receive Flash pricing until the future V4.1 Pro release. DeepSeek has not supplied a V4.1 Pro release date.

The current specification page lists a 1M-token context window and a maximum output allowance of 384K tokens. That maximum is not the default response length: the API reference sets lower defaults, and input plus generated output must fit within the context window. Thinking is enabled by default; the default reasoning effort is high.

Save Pro prompts, settings, tool versions and accepted outputs while the original service is available. Record the request time and returned version information where provided. Our V4 Pro release coverage explains the previous checkpoint.

DeepSeek describes V4.1 Flash as a mixture-of-experts model with 552 billion backbone parameters. Its technical report, section 2.1 separately lists 196 billion Engram conditional-memory parameters. The headline backbone count therefore should not be treated as a complete inventory of every parameter in the system.

The Causal Encoder–Decoder design splits the backbone into 20 encoder and 20 decoder layers. It activates approximately 8 billion parameters while processing input and 16 billion while generating output. Most prompt tokens can bypass full decoder computation. This targets the repeated, input-heavy work of agents reading growing histories; active parameters describe computation, not the memory needed to store the whole model.

The official cache comparison reports 890 bytes of global KV cache per token, versus 3,514 for V4 Flash, about a 3.9-fold reduction. KV cache stores attention state reused during generation. Cross-layer sharing and FP4 compression reduce this global cache in high-bandwidth GPU memory; a separate sliding-window replay technique cuts persistent cache storage to roughly one-eighth of V4 Flash’s. These are cache measurements, not total GPU-memory requirements or promised reductions in every API bill.

Native vision accepts images alongside text and produces text. The vision guide confirms JPEG, PNG, GIF and WebP inputs, including screenshots and charts, through the current Flash API. Image tokens contribute to usage charges. This is visual understanding, not image generation, and the tool-assisted visual benchmarks below should not be read as unaided perception scores.

The following values reproduce the numerical comparison in DeepSeek’s launch graphic and technical report, Table 3. All seven model columns are labeled Max reasoning effort in the report. These are DeepSeek-reported results. Digital Applied has not reproduced them, and the comparison is not a test at equal dollar or token budgets.

Pass@1 measures single-attempt success; Pass@5 uses a five-attempt allowance. DeepSWE reports resolved tasks, ProgramBench uses Almost@1 and NL2Repo has its own score. Codeforces is a rating, not a percentage. Scroll horizontally to compare all seven models.

* HLE text-only subset. Flash scores 36.8 on full HLE and 39.1 on the text-only subset. Compare its 39.1 with Pro’s 42.7, rather than comparing different subsets. An em dash means no result was supplied, not zero. Terminal-Bench 2.1, 3.0 and 4.0 are different evaluations; their scores are not a progression showing one model getting worse.

The clearest Pro replacement gains are in agent work: Flash improves DeepSWE by 11.5 percentage points and Automation-Bench by 11.6 points in this table. Terminal-Bench 3.0 and 4.0 improve by 18.2 and 18.8 points. But Pro remains ahead on GPQA and text-only HLE. The vendor’s broad superiority claim should not become a claim that Flash wins every row.

Against the other listed models, the pattern is also mixed. Flash leads several agent rows, while Opus 5 and GPT 5.6 Sol remain ahead on Terminal-Bench 3.0 and 4.0, ProgramBench and ExploitGym. A 0.2-point DeepSWE lead over Opus is a reported numerical difference, not evidence of a statistically reliable advantage. The table provides no confidence intervals.

These comparisons cover the named checkpoints. They do not establish a general lead over newer models such as Astra or Fable 5.1, which are absent from this official table.

Flash’s new tariff took effect on September 10 at 04:00 UTC. DeepSeek’s pricing documentation now confirms both models’ rates below, in USD per million tokens. Pro keeps its existing tariff during the extension; it receives Flash pricing when routing changes on September 14.

Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; other hours are off-peak. The launch pricing graphic also mentions public holidays as off-peak, but the documentation does not identify a holiday calendar. Confirm holiday treatment before relying on it for scheduled workloads. Cached input means input actually billed as a cache hit.

Illustrative uncached workload: 10 million input tokens plus 1 million output tokens costs $2.10 on Flash versus $8.58 on Pro off-peak. At peak prices, the same volumes cost $4.20 and $17.16. The tariff difference is approximately 75.5%; this is arithmetic, not a measured saving from migrating an application.

If 9 million input tokens are cache hits and 1 million are misses, those off-peak totals become $0.777 and $2.838. That assumes identical output, a 90% input cache-hit share and no retries or extra fees. Longer reasoning traces can change the comparison. Our off-peak pricing analysis explains how scheduling affects the cost of completed work.

Security: better detection in a named preview

Aikido researcher Philippe Dourassov’s September 9 results identify deepseek-v4.1-flash-expires-on-0910. This is the preview endpoint he tested, not the production model name or proof that today’s released checkpoint is identical.

Recall measures how many target vulnerabilities were found; precision measures how many findings were accepted as valid. Pooled pass@3 combines discoveries from three runs. It does not give a single run an 84.4% success rate.

Aikido’s August 21 report, We burned 11.7bn tokens to find the best cyber AI model, updated September 1, describes 32 recently disclosed vulnerabilities, three passes, a 30-turn primary investigation limit and no internet. The short preview update does not republish every setting. The earlier report’s Pro result was 28 of 32 vulnerabilities across three runs, or 87.5%, above the preview’s reported 84.4%. This is a dated comparison, not a controlled retest of the two released models.

Design: a stronger average with a dashboard exception

OpenDesign Arena reports these prototype averages. Its model labels are useful external evidence, but the inspected methodology does not establish the exact final Flash checkpoint.

The overall result hides a relevant reversal: Flash scores 76.1 for dashboards versus Pro’s 83.0. Requirements contribute 30 points and design 70; non-rendering artifacts score zero without replacement retests. Costs use token counts and list prices, timing excludes queueing, and per-model sample counts and uncertainty intervals are unspecified. These are prototype results, not proof of production readiness or measured bills under today’s tariff.

The report goes beyond a leaderboard. In Table 4, DeepSeek holds the Flash checkpoint, tasks and decoding settings fixed while changing the surrounding agent framework: its prompts, tools and conversation management. The resulting spread is large enough to affect a deployment decision.

These runs use eight samples per DeepSWE task and three per Terminal-Bench task, with 1M context, temperature 1.0, top-p 0.95 and a 500-generation-round limit. Sampling counts describe the evaluation protocol; the reported Terminal-Bench metric remains Pass@1. The 8.7-point DeepSWE spread shows why swapping a model into your existing tools is a separate test from reproducing the headline score.

Effort also has a cost. In section 5.3.3 and Figure 9, increasing the experimental effort value from 25 to 100 raises DeepSWE from 66.0 to 74.2 and Terminal-Bench 2.1 from 82.4 to 90.6, with roughly 2.5 times as many output tokens. The API exposes named tiers: low, high and max map to 50, 75 and 100. The experimental value 25 is not an advertised API tier.

The official weight repository includes the checkpoints and MIT license. Its inference README describes reference code, not a production serving engine. DeepSeek’s deployment announcement says community inference support and further deployment options are being developed.

There are preliminary multi-agent results too: on a filtered 172-task ProgramBench subset, the report compares 30.04% Almost@1 for an agent team with 20.39% for a single agent at an eight-hour deadline. That experiment uses selected configurations and a different task subset. It should not replace the main table’s 20.3 or be presented as a controlled dollar-for-dollar improvement.

Before the Pro cutoff, select representative tasks with clear acceptance tests and run both released models with the same files, tools and settings. Count failed attempts, billed tokens, elapsed time and human corrections. Add a long-history retrieval task and a resumed-cache task: the report’s limitations specifically identify sparse retrieval and cache-state reconstruction as areas where untested boundary cases may degrade performance.

Our search did not establish an independent Artificial Analysis Intelligence Index score for the released V4.1 checkpoint. Do not relabel an older V4 score, or convert these specialist tests into a general intelligence ranking. The practical comparison remains the cost and reliability of an accepted result, as in our Astra and Fable comparison.

Evidence checked September 10, 2026: DeepSeek’s launch thread, technical report, model repository and API documentation; the supplied customer notices; Dourassov’s preview report and methodology background; and OpenDesign Arena. Vendor results and external preview tests are distinguished throughout. No paid model tests or independent benchmark reproduction were performed by Digital Applied.