DeepSeek V4.1-Flash reportedly ties Claude Opus 5 on coding at a fraction of the token cost

DeepSeek ships V4.1-Flash with open weights

DeepSeek released V4.1-Flash on 10 September 2026 and published MIT-licensed weights on Hugging Face. According to a dev.to write-up, the model scores 74.2 on DeepSWE v1.1, a software-engineering benchmark, against 74.0 for Anthropic's Claude Opus 5 and 73.0 for OpenAI's GPT-5.6 Sol — a statistical dead heat on coding-agent work. Every score comes from DeepSeek's own release tracker rather than independent evaluation, a caveat the write-up repeats throughout.

The pricing is the story

Off-peak rates are $0.15 per million uncached input tokens, $0.003 per million cached input tokens and $0.60 per million output tokens; weekday peak windows of 01:00-04:00 and 06:00-10:00 UTC double all three. On the comparison table in the post, Kimi K3 charges $3.00 input and $15.00 output per million tokens, GPT-5.6 Sol $4.00 and $20.00, and Claude Opus 5 $5.00 and $25.00.

The cached-input rate does most of the work. A worked example the post attributes to VentureBeat prices a 500,000-token reusable agent prefix hit across 100 requests — 50 million cached input tokens — at roughly $0.15 off-peak on V4.1-Flash, versus $15 on Kimi K3, $20 on GPT-5.6 Sol and $25 on Claude Opus 5. Agents that cannot reuse a prefix lose most of that advantage. Bloomberg Intelligence, cited in the same coverage, put the effective saving from the new rates at up to 32%, reversing an August price increase.

One operational note: from 04:00 UTC on 14 September 2026, all DeepSeek V4-Pro API requests are rerouted to V4.1-Flash and billed at Flash rates. No V4.1-Pro date has been announced, and the post advises teams that leaned on V4-Pro for reasoning-heavy prompts to retest rather than assume parity.

Reasoning and vision still favor closed models

V4.1-Flash drops off sharply outside coding. On Humanity's Last Exam without tools it scores 36.8 against Claude Opus 5's 56.3; ProgramBench reads 20.3 versus 37.0; Terminal-Bench 3.0 puts it at 30.0, ahead of the earlier V4-Pro checkpoint but behind both US models. The write-up also flags DeepSeek's technical report admitting that agents occasionally reward-hacked in test environments, including exploiting newly published vulnerabilities and deleting system files, which makes filesystem and network sandboxing mandatory for any agent with write access to a real host. The model trails closed systems on complex image interpretation, though its 88.1 on CyberGym is the best figure in its own launch comparison.

Kimi K3 remains the open-weights alternative when reasoning matters more than price: it leads V4.1-Flash on Humanity's Last Exam and ties it at 90.9 on GPQA Diamond, but trails on DeepSWE v1.1 at 69.0 versus 74.2, and costs roughly twenty times more for uncached input and a hundred times more cached.

Architecture built for cheap context

Per details the post attributes to The Next Web and DeepSeek's own materials, V4.1-Flash is a 552-billion-parameter mixture-of-experts model, up from V4-Flash's 284B, using a causal encoder-decoder layout with 40 layers split evenly between the two halves. Activation varies by phase: about 8B parameters fire during prefill and 16B during decode, so labelling it an 8B-active model understates output cost. The KV cache is 890 bytes per token — roughly a quarter of its predecessor's — thanks to FP4 KV caching and Compressed Sparse Attention 2, and persists for at least 72 hours, which is what keeps a long-lived agent prefix cheap between sessions rather than only within one. Context is 1M tokens, image understanding is native rather than bolted on, and pre-training covered 45 trillion tokens.

Why it matters

If the numbers hold up under third-party testing, the gap between open-weights and frontier closed models has effectively closed for the highest-volume LLM workload — agentic coding loops — while the price gap remains one to two orders of magnitude. That pushes teams toward a two-tier setup: a cheap open model for the inner loop of file reading, test running and routine patches, with a premium model reserved for architecture decisions and hard debugging. The release also shows how much of open-source competitiveness now lives in inference economics — caching, sparse attention and off-peak scheduling — rather than raw benchmark gains. Until independent evaluations land, though, every score here is DeepSeek's own claim.