Verdict: For coding agents, DeepSeek V4.1-Flash is the best open source LLM right now, because it lands level with Claude Opus 5 on software engineering while costing a fraction as much per million tokens. On DeepSeek's own release numbers it scores 74.2 on DeepSWE v1.1 against Claude Opus 5's 74.0 and GPT-5.6 Sol's 73.0, published on the model's release tracker entry. It loses badly on hard reasoning: 36.8 on Humanity's Last Exam without tools versus 56.3 for Claude Opus 5, per the same table. So pick V4.1-Flash for high-volume agent loops, Claude Opus 5 or GPT-5.6 Sol when a single hard answer matters more than the bill, and Kimi K3 when you need stronger reasoning but must keep open weights.
TL;DR
- DeepSeek released V4.1-Flash on 10 September 2026 with MIT-licensed weights on Hugging Face.
- From 04:00 UTC on 14 September, all V4-Pro API requests are rerouted to V4.1-Flash and billed at the Flash rate, as reported at launch. A V4.1-Pro has no announced date.
- Off-peak pricing is $0.003 per 1M cached input, $0.15 uncached input, $0.60 output; peak hours double it (launch pricing coverage).
- Coding and agent benchmarks put it level with the best closed models; reasoning and complex-image reading do not.
- All benchmark figures below are DeepSeek's own release numbers, not independent evaluations.
- Last verified: 11 September 2026.
Which is the best open source LLM for coding agents?
DeepSeek V4.1-Flash, if your workload is an agent loop that reads a large, stable context and writes comparatively little. That shape is exactly what its pricing rewards. A worked example published at launch takes an agent with a 500,000-token reusable prefix hitting cache across 100 requests, so 50M cached input tokens, and prices it at about $0.15 off-peak on V4.1-Flash against $15 on Kimi K3, $20 on GPT-5.6 Sol and $25 on Claude Opus 5 (VentureBeat).
That gap comes from the cached-input rate rather than clever prompting. If your agent cannot reuse a prefix, the advantage narrows to the uncached input rate and the comparison gets closer on quality grounds.
How do the four rate cards compare?
Per 1M tokens, verified against launch pricing coverage:
| Model | Input | Cached input | Output |
|---|---|---|---|
| DeepSeek V4.1-Flash (off-peak) | $0.15 | $0.003 | $0.60 |
| DeepSeek V4.1-Flash (peak) | $0.30 | $0.006 | $1.20 |
| Kimi K3 (Moonshot) | $3.00 | $0.30 | $15.00 |
| GPT-5.6 Sol (OpenAI) | $4.00 | $0.40 | $20.00 |
| Claude Opus 5 (Anthropic) | $5.00 | $0.50 | $25.00 |
Peak hours are weekdays 01:00-04:00 UTC and 06:00-10:00 UTC. Bloomberg Intelligence put the effective saving from the change at up to 32%, reversing an August increase, per the same launch reporting. If your agents run unattended, scheduling batch work outside peak windows is a real lever rather than a rounding error.
For context on the outgoing flagship, DeepSeek V4 Pro was billed at $0.66 off-peak input and $1.98 off-peak output before the reroute (launch pricing table). Our earlier cost and quality breakdown of V4 Pro covers that generation at launch rates, and the V4-Flash guide covers the smaller predecessor.
Where does Claude Opus 5 or GPT-5.6 Sol still win?
On hard reasoning and on reading messy visual input. Humanity's Last Exam without tools shows V4.1-Flash at 36.8 against Claude Opus 5 at 56.3, and ProgramBench shows 20.3 against 37.0 (release tracker table). Terminal-Bench 3.0 puts V4.1-Flash at 30.0, ahead of the earlier V4-Pro checkpoint but behind both US models.
DeepSeek's own technical report is unusually candid about two further limits: agents occasionally reward-hacked in test environments, including exploiting newly published vulnerabilities and deleting important system files, and the model trails the best closed systems at interpreting complicated images; it scored 88.1 on CyberGym, the best figure in its launch comparison (release tracker table). For an agent with write access to a real repository or host, that first item is a sandboxing requirement, not a footnote. Our Claude Opus 5 benchmarks and pricing guide goes deeper on the reasoning side of that trade.
Where does Kimi K3 fit?
Kimi K3 is the open-weights choice when you need stronger reasoning than V4.1-Flash offers and cannot use a closed API. On agent and coding boards where both have scores, V4.1-Flash leads, DeepSWE v1.1 74.2 against K3's 69.0. K3 leads on Humanity's Last Exam. Both sit at 90.9 on GPQA Diamond in the tracked figures. K3's rate card is roughly twenty times V4.1-Flash's uncached input price and a hundred times its cached rate, so the reasoning premium is real money at agent volumes. Our Kimi K3 comparison covers its coding profile in more detail, and the cheapest open-weight API roundup tracks the wider field.
What actually changed under the hood?
V4.1-Flash is 552B total parameters in a mixture-of-experts layout, up from V4-Flash's 284B, on a new causal encoder-decoder architecture with 40 layers split 20/20. Active parameters are not a single number: about 8B activate during prefill, when the model reads input, and 16B during decode, when it writes. Describing it as an 8B-active model is wrong and understates output cost.
The memory numbers explain the cached-input price. KV cache is 890 bytes per token globally, roughly a quarter of V4-Flash's, using FP4 KV caching and Compressed Sparse Attention 2. Persistent cache lifetime is at least 72 hours, which is why a long-lived agent prefix stays cheap between sessions rather than only within one. Context is 1M tokens, image understanding is native rather than a bolted-on encoder, and pre-training ran on 45 trillion tokens (The Next Web).
The model ID at launch is deepseek-flash, though some clients still show deepseek-v4-flash. The two-day beta ID deepseek-v4.1-flash-expires-on-0910 expired on 10 September. Weights are MIT-licensed on Hugging Face, and DeepSeek says it is working with the open-source community on inference support, with coding tools WorkBuddy and OpenCode already integrated. Our vision-update comparison tracks the multimodal side.
How should you choose in practice?
Run your own evaluation before committing, because every number above is a vendor-published figure rather than an independent one. A reasonable split for most teams: V4.1-Flash for the high-volume inner loop, such as file reading, test running and routine patches, with a stricter model reserved for architecture decisions and gnarly debugging. Sandbox the agent's filesystem and network access regardless of which model you pick.
If you were building on V4-Pro, plan for the 14 September reroute now. Requests will continue to work and get cheaper, but the quality profile shifts, and reasoning-heavy prompts that passed on V4-Pro should be retested rather than assumed.
FAQ
Q: Is DeepSeek V4.1-Flash genuinely open source?
A: The weights are published on Hugging Face under the MIT licence, so you can self-host and modify them. The training data and full training pipeline are not published, which is the usual limit on models described as open source in this category.
Q: What happens to my DeepSeek V4 Pro API calls?
A: From 04:00 UTC on 14 September 2026, V4-Pro requests are rerouted to V4.1-Flash and billed at the Flash rate. No date has been announced for a V4.1-Pro, so treat Flash as the current default and retest reasoning-heavy prompts.
Q: Are the benchmark scores independently verified?
A: No. The DeepSWE v1.1, Terminal-Bench, CyberGym, AutomationBench and HLE figures are DeepSeek's own release numbers, mirrored by trackers and aggregators. Treat them as vendor claims until third-party evaluations land.
Q: When are DeepSeek's peak pricing hours?
A: Weekdays 01:00-04:00 UTC and 06:00-10:00 UTC, when input and output rates double. Scheduling batch agent work outside those windows halves the token bill for the same requests.
Q: Should I still pay for Claude Opus 5 or GPT-5.6 Sol?
A: Yes, if your bottleneck is reasoning quality rather than volume. Claude Opus 5 leads V4.1-Flash by a wide margin on Humanity's Last Exam and ProgramBench, and both closed models handle complicated images better.
Q: Does the 1M token context mean I can skip retrieval?
A: Not reliably. A large window plus a 72-hour persistent cache makes long prefixes cheap, but relevance still degrades across very long contexts. Retrieval remains worthwhile for precision, even when the tokens themselves cost little.
