DeepSeek V4.1 Flash: Opus 5-tier agentic coding at 1/33 the price

On September 10, 2026, DeepSeek did something it has never done before: it released a Flash model and then announced it would retire the Pro tier because the cheaper model beat it. DeepSeek-V4.1-Flash is a 552B-parameter mixture-of-experts model with native vision, a 1M-token context window, and a new Causal Encoder-Decoder architecture that activates just 8B parameters per input token and 16B per output token. On DeepSeek's own max-effort benchmark table it edges Claude Opus 5 and GPT-5.6 Sol on Terminal-Bench 2.1 (90.6 vs. 89.1 and 88.8) and DeepSWE v1.1 (74.2 vs. 74.0 and 73.0) — while costing $0.15 input and $0.60 output per million tokens off-peak.

The question everyone is asking: which Claude and ChatGPT models does this actually compete with? The short answer, backed by the numbers below: V4.1 Flash claims the Claude Opus 5 / GPT-5.6 Sol tier on agentic coding — Terminal-Bench 2.1, DeepSWE v1.1, CyberGym, AutomationBench — while staying behind both on deep reasoning (GPQA Diamond, HLE, Terminal-Bench 3.0/4.0). And it is priced like nothing in that tier: $0.15/$0.60 per million tokens, which is 33x cheaper on input and 42x cheaper on output than Claude Opus 5, and 33x/50x cheaper than GPT-5.6 Sol. If you have been running coding agents on Opus 5 or Sol, this release deserves a slot in your eval matrix today.

Everything below cites its source and names the exact benchmark variant (Terminal-Bench 2.1 is not Terminal-Bench 3.0, DeepSWE v1.1 is not SWE-bench Verified). One honest caveat up front: the Opus 5 and Sol comparisons are DeepSeek-reported — as of launch day, Artificial Analysis and vals.ai have not yet published independent V4.1 Flash scores, so the vendor table is all we have.

What actually shipped

DeepSeek-V4.1-Flash is the smallest model in a new DeepSeek architecture family, and “smallest” is relative: the backbone is a 552B-parameter MoE with 1 shared expert and 384 routed experts per layer, of which 6 are activated per token. The headline innovation is the Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the final encoder hidden states instead of being derived from each decoder layer. The result is asymmetric compute — 8B active parameters per input token, 16B per output token — which is exactly the shape of agent workloads, where tens of thousands of context tokens are read for every few hundred generated. The technical report also lists a 196B-parameter Engram conditional memory module that is sparsely accessed via token-based lookup, so the 552B backbone count is not the whole inventory.

It is DeepSeek's first Flash model with native vision: a from-scratch DeepSeek-ViT encoder (2D-RoPE, 3×3 pixel-unshuffle downsampling) plus a two-layer MLP projector feeds image embeddings into the model from the start of pre-training. The API accepts JPEG, PNG, GIF, and WebP images — screenshots, charts, documents — alongside text. It was pre-trained from scratch on a 45T-token multimodal corpus, with sparse attention trained at 64K sequence length and the context extended to 1M tokens at 34T tokens. Post-training follows the standard SFT → RL → on-policy distillation recipe, with the changes concentrated in the data pipeline: large-scale automated synthesis of agent tasks and environments. Reasoning effort is continuously controllable from 1 to 100, and the API defaults to thinking mode.

The spec sheet: 1M-token context, 384K max output, tool calls, JSON output, Responses API, and an Anthropic-format endpoint. The weights are on Hugging Face under MIT — the same license as V4 Flash and V4 Pro — with a technical report and inference code included. The API model name is deepseek-flash; the retired deepseek-v4-flash and deepseek-v4-flash-vision-exp names temporarily route to it. WorkBuddy (including CodeBuddy) and OpenCode announced full support on day one.

The KV-cache story: why this model is cheap to serve

The architecture's second act is cache compression. V4.1 Flash's global KV cache is 890 bytes per token — roughly one-quarter of V4 Flash's 3,514 bytes, and about 437× smaller than DeepSeek-V1's. DeepSeek gets there with Compressed Sparse Attention 2 (CSA2), which assigns each attention layer one of three static modes — Full, Reindex, or Reuse — to share main KV and indexer keys across layers and reuse Top-K sparse-attention indices, plus FP4 KV caching (E2M1 format, one E4M3 scale per 16 channels). A separate SWA Bounded Replay trick reconstructs missing sliding-window states by replaying only the most recent tokens, so the persistent cache footprint on SSD drops to roughly one-eighth of V4 Flash's. DeepSeek's announcement frames it as 1/4 the HBM and 1/8 the SSD of the previous generation.

Why this matters: cache-hit charges often dominate agent bills, because an agent loop re-sends the same growing context on every turn. A smaller cache means more concurrent users per GPU and cheaper cache reads — and DeepSeek prices those reads at $0.003 per million tokens off-peak, a ~98% discount off the cache-miss input rate. That is the lever that makes the economics below work.

The benchmark table: Opus 5-tier agentic coding, with gaps

DeepSeek published a 19-benchmark comparison of V4.1 Flash against Claude Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.3, V4 Pro 0813, and V4 Flash 0731, all at max reasoning effort. The agentic column is where the Flash model lands its punches: Terminal-Bench 2.1 at 90.6 beats Opus 5 (89.1), Sol (88.8), and V4 Pro (87.9); DeepSWE v1.1 at 74.2 edges Opus 5 (74.0) and Sol (73.0) and crushes V4 Pro (62.7); CyberGym at 88.1 tops Sol (84.5) and V4 Pro (83.3); AutomationBench at 54.8 beats Opus 5 (50.3) and Sol (45.8); HLE with tools at 63.9 edges Opus 5 (63.6); and Agents' Last Exam at 31.8 leads the whole table. On Codeforces, V4.1 Flash posts a 3471 rating versus 3348 for V4 Pro and 3289 for V4 Flash.

The reasoning column is the honest counterweight. GPQA Diamond at 90.9 trails Opus 5 (93.4) and Sol (94.1) — and even V4 Pro (92.4). HLE at 36.8 (39.1 on the text-only subset) is far behind Opus 5's 56.3 and Sol's 44.5. Terminal-Bench 3.0 at 30.0 and Terminal-Bench 4.0 at 31.2 sit well below Opus 5's 43.3 and 51.8 and Sol's 34.4 and 39.9 — though both are huge jumps over V4 Pro's 11.8 and 12.4. ProgramBench (20.3 vs. Opus 5's 37.0) and NL2Repo-Bench (65.4 vs. Opus 5's 75.3) also stay red. The pattern is consistent: V4.1 Flash is an agentic-coding specialist that trades deep-reasoning depth for cost.

The pricing math

DeepSeek cut Flash-series prices with this release and kept its peak/off-peak structure. Per million tokens: $0.15 input and $0.60 output off-peak, doubling to $0.30 and $1.20 during peak hours (01:00–04:00 and 06:00–10:00 UTC, Monday–Friday), with cache hits at $0.003 off-peak / $0.006 peak. The API concurrency limit is 2,500. For contrast, the V4 Pro endpoint it is replacing costs $0.66 input and $1.98 output off-peak — 4.4x and 3.3x more — with a 500-request concurrency cap.

A worked example makes the gap concrete. Take a coding-agent task that reads 10M input tokens (90% cache hits) and writes 50K output tokens. On V4.1 Flash off-peak: 9M cache reads × $0.003 = $0.027, 1M new input × $0.15 = $0.15, and 50K output × $0.60 = $0.03 — about $0.21 total. The same task on Claude Opus 5 (cache reads at $0.50): $4.50 + $5.00 + $1.25 = $10.75, and on GPT-5.6 Sol (cache reads at $0.50): $4.50 + $5.00 + $1.50 = $11.00. That is a 51x and 52x gap on one task, and it compounds across the hundreds of calls a long-horizon agent makes.

Benchmark figures are DeepSeek's published max-effort comparison (Terminal-Bench 2.1, DeepSWE v1.1, GPQA Diamond). DeepSeek prices are off-peak list rates from the Models & Pricing page; Claude and OpenAI prices are list rates from their pricing pages (Sonnet 5 at its September 1, 2026 rate). V4 Pro 0813 parameters are as reported in our August articles.

The V4 Pro retirement

The most unusual part of this launch is what happens to the model above it. DeepSeek says tests by multiple parties put V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime, so it is phasing V4 Pro out: starting 04:00 UTC on September 14, 2026, every deepseek-v4-pro request will be served by V4.1 Flash and billed at Flash rates, until V4.1 Pro launches at an unannounced date. V4 Flash and V4 Flash Vision Exp are already retired, with their model names temporarily routing to V4.1 Flash. In effect, DeepSeek's entire API lineup collapsed to one model overnight — the consumer app's fast, expert, and image-recognition modes were merged into a single mode served by V4.1 Flash too.

For Pro users the math is unambiguous: the same workload gets a model that DeepSeek claims is better on agentic tasks at 4.4x cheaper input and 3.3x cheaper output. The two places V4 Pro still wins on DeepSeek's own table are GPQA Diamond (92.4 vs. 90.9) and the text-only HLE subset (42.7 vs. 39.1) — if your workload is pure-reasoning rather than agentic, capture a Pro baseline before September 14 while the endpoint still serves the real model.

Where the closed frontier still wins

The honest framing: everything above is DeepSeek-reported. As of launch day, Artificial Analysis has no V4.1 Flash page and vals.ai has no V4.1 Flash scores — their latest snapshots predate the release. The one early third-party data point comes from OpenDesign's design arena, which scored V4.1 Flash at 81.2 out of 100, second overall behind GPT-6 Astra (82.7) and ahead of GPT-5.6 Sol (77.6), with a reported $0.023 cost per run. Community speed tests during the two-day beta reported roughly 300–500 tokens/second, with speed holding up as context grows — a direct consequence of the compressed KV cache. Those are encouraging signals, not a substitute for independent SWE-bench Verified or Vibe Code Bench runs, which have not happened yet.

And the closed frontier's lead is real where it matters most to some buyers. On DeepSeek's own table, Claude Opus 5 keeps a commanding edge on HLE (56.3 vs. 36.8), Terminal-Bench 3.0 (43.3 vs. 30.0), Terminal-Bench 4.0 (51.8 vs. 31.2), ProgramBench (37.0 vs. 20.3), and NL2Repo-Bench (75.3 vs. 65.4); GPT-5.6 Sol keeps GPQA Diamond (94.1 vs. 90.9) and SEC-Bench Pro (74.3 vs. 62.8). On the vision benchmarks, Opus 5 also leads Chartography (84.0 vs. 78.9), BabyVision (94.1 vs. 89.6), and ZeroBench (52.0 vs. 49.0). If your workload is frontier math, deep research, or long-horizon terminal autonomy, the expensive models still earn their price.

What this means for your stack

The tier mapping in one line: V4.1 Flash is priced below Claude Haiku 4.5 and claims Claude Opus 5-tier agentic coding. At $0.15/$0.60 it undercuts Haiku's $1/$5, yet on DeepSeek's table it beats Opus 5 — a $5/$25 model — on Terminal-Bench 2.1, DeepSWE v1.1, CyberGym, AutomationBench, HLE with tools, and Agents' Last Exam. The practical playbook for high-volume coding agents is two-tier: default to V4.1 Flash for the loop — the cache-hit rate of $0.003 makes long-context agent runs nearly free — and escalate to Opus 5 or Fable 5 only for the deep-reasoning steps where the benchmark gaps above bite. With the weights under MIT, you can also self-host and pay nothing per token.

The bigger story is the direction of travel. In April, DeepSeek V4 Flash was a 284B/13B-active model that trailed the frontier on agentic work; the July 0731 refresh closed much of that gap at the same price. Now V4.1 Flash doubles the backbone to 552B, adds vision, and claims the Opus 5 tier on agentic coding at the same Flash rate card — and DeepSeek is confident enough to delete its own Pro tier over it. If the independent harnesses confirm even most of the vendor table, the cost of Opus 5-class agentic coding just fell by roughly 98%, and the closed frontier's defensible ground shrank to deep reasoning and from-scratch app building. That is the September 10, 2026 headline, and it will be tested within days.