Step 5 Preview vs DeepSeek V4.1 Flash: Which Low-Cost Agent Model to Pick

1. Why are these two models at the top of OpenRouter?

Because both are cheap enough to run in loops, and agent frameworks are the ones burning the tokens. On the OpenRouter weekly ranking checked on October 16, 2026, DeepSeek V4.1 Flash ranked first at 37.1T tokens and Step 5 Preview ranked second at 19.5T.

Step 5 Preview is not being called by chat apps. Its top five callers on OpenRouter are Hermes Agent, Kilo Code, OpenClaw, Cline, and StepFun's own Step Code, all agent tooling (Founder Park, October 15, 2026). Step 5 Preview launched on September 20, 2026, was listed on OpenRouter on October 8, and had its weights released on October 15. Within that first week on OpenRouter it reached 18.8T weekly tokens, second place on both the weekly and daily boards, with 3.74T tokens on October 14 alone.

On the DeepSeek side, V4.1 Flash shipped on September 10, 2026. The official announcement states it leads V4 Pro on performance, cost, and speed, and that the older aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to it.

Both models are cheap, but they spend the budget differently. Step 5 Preview pays for depth: a deeper network and a model that manages its own context on long runs. DeepSeek V4.1 Flash pays for throughput and unit cost, so a loop can run without a second thought about price. The numbers behind that split, and where they come from, are below.

2. Step 5 Preview vs DeepSeek V4.1 Flash: key specs

Parameter note: the "about 600B" figure comes from the official model card. The vLLM recipe counts about 604B from checkpoint tensor shapes, which is a different counting method. The 27B figure is the activated compute path, not a memory footprint: inactive experts still have to be stored or distributed, and long contexts need KV cache on top.

If you want to try this DeepSeek V4.1 Flash model, visit coding plan.

3. Who is calling these models, and how are they promoted?

Agent tools are the distribution channel, and both vendors are acting on that. On OpenRouter's application board the largest app by volume is the open-source agent framework Hermes Agent, at 2T tokens in a single day, ahead of Claude Code (Founder Park, October 15, 2026). Being in a tool's model dropdown is now the shortest path to a user.

Step 5 Preview ran that playbook hard. It was integrated into Cline, Kilo Code, and Nous Research's Hermes, with limited-time free access in each. In China it went live on ByteDance's TREA on September 22 and had an exclusive free window on Tencent's WorkBuddy from October 12 to 25. The official line is "Keep your workflow. Switch your model."

DeepSeek did not need third-party tools to find users, but it benefits from the same shift in usage. Once token cost drops past a threshold, developers stop asking whether a run is worth the money, and agents move from single trial runs to loops. Addy Osmani named that pattern Loop Engineering in June 2026; Boris Cherny, who works on Claude Code at Anthropic, put it plainly: he no longer prompts Claude, he has loops that prompt Claude.

4. How do the two architectures differ, and where does the cost go?

Step 5 Preview: deeper and narrower, paying for long trajectories

Step 5 Preview is a change of direction for StepFun. Earlier generations, Step 3.5 Flash and Step 3.7 Flash, activated only about 11B parameters each and focused on speed and tool calling. Step 5 Preview scales to 600B total and 27B active and raises context from 256K to 1M, with the stated goal of extending cheap Flash-style serving to harder, longer tasks.

It stacks 92 transformer blocks, close in depth to the trillion-parameter Kimi K3 while using under a quarter of its parameters. The reasoning given: agents spend most of their time processing tool output rather than writing out step-by-step reasoning, which makes implicit reasoning during prefill more important. A deeper network provides a longer compute path for multi-hop information, while a narrower one keeps cost down. On attention it uses Sparse GQA with Block-wise Token Merging to optimize the sparse index, cutting index and Top-K selection cost to one eighth with almost no loss of coverage. Cache reads are priced at one tenth of the standard rate, and repeated system prompts and tool definitions in an agent loop hit that cache naturally: Founder Park reported that 96% of Step 5 Preview input tokens were cache hits.

Context management is the most distinctive part. The window is 1M, but real tasks often run longer. Rather than truncating passively, Step 5 Preview summarizes its working state into a resumable summary when the window fills and keeps going. That behavior was trained end to end inside reinforcement learning with the task reward, not bolted on as an external trimming rule. The companion agent harness, Step Code, was open-sourced on September 23 under the slogan "Same smarts, fewer tokens," splitting long tasks across sub-agents with independent contexts. In StepFun's own comparison of six harnesses running the same model, Step Code came first on both long-horizon completion (73.3%) and average tokens per task (5.09M). These are vendor-reported Step 5 Preview results. The supplied source does not specify full harness versions or evaluation settings; we have not reproduced the comparison.

StepFun also showed a single long run: 24 hours on one H100 optimizing an MLA GPU kernel from scratch, writing code, compiling, measuring, and falling back to the best earlier version when a change regressed. After more than 180 iterations it peaked at about 508 TFLOPS.

DeepSeek V4.1 Flash: throughput and unit cost

DeepSeek V4.1 Flash puts its budget on the other axis. The public numbers are speed and price: 214.1 tokens/s in max reasoning and 216 tokens/s in non-reasoning mode, $0.30 per 1M input tokens and $1.20 per 1M output at peak, $0.006 per 1M cached input tokens, and half price off-peak. It carries a 1M context window, 384K maximum output, and native multimodal input, and the official announcement claims it beats V4 Pro on performance, cost, and speed.

That combination maps directly onto looping. Prompt prefixes repeat heavily from step to step, so cache hits are the norm, and at $0.006 per 1M cached tokens each extra iteration is close to free. Throughput sets how many steps a loop can advance per minute. What it does not publish is a mechanism aimed specifically at very long trajectories; context handling relies on the standard window and compression approach.

The core trade-off

Step 5 Preview buys depth per task: the deeper network, the RL-trained context self-summarization, and the paired harness all serve getting one long trajectory to the end. DeepSeek V4.1 Flash buys throughput per dollar: high decoding speed, near-zero cached input cost, and time-of-day pricing serve running a large, repetitive loop at all. One optimizes capability depth, the other optimizes cost and speed.

What a month of agent loops costs

Illustrative arithmetic from published per-token prices, not a benchmark. Cache hit rates depend entirely on your workload and prompt structure; the 90% row is a sensitivity check, not a claim about typical traffic.

5. Which model fits which loop?

Loop Engineering describes a way of working where you no longer write the prompt; you design a loop that writes prompts, with the agent discovering tasks, spawning sub-agents, verifying results, and remembering progress. That pattern splits model requirements into two kinds, and each of these models answers one of them.

• Step 5 Preview fits work with a long single trajectory, multi-hop information passing, and frequent self-correction inside one task. Sweeping an entire codebase file by file, long kernel or algorithm optimization runs, and research tasks that must hold working state across many steps all land here. Its context self-summarization and deeper network target exactly that shape of work, and the 1M window plus self-compression means task length is not capped by the window alone.
• DeepSeek V4.1 Flash fits loops where individual steps are fairly independent but there are a great many of them and prefixes repeat: bulk file-level review and rewriting, batch classification and extraction, wide fan-out of sub-tasks. In these, throughput and cached-input price matter more than per-task depth.

Neither model is strictly better. They sit on different directions of the same cost-and-capability frontier. If you would rather spend more per task to raise completion on hard work, you move toward Step 5 Preview. If you would rather trade per-task depth for a much larger loop at a much lower unit cost, you move toward V4.1 Flash. Task shape decides, not a leaderboard rank.

6. What are the drawbacks of each model?

Step 5 Preview

• The license is not Apache 2.0. The weights ship under the custom StepFun Model License 1.0, tagged "other" on Hugging Face. A framework license does not substitute for the model license, so read the version-specific text before any commercial deployment.
• The hardware bar is high. The published recipes are all server-class NVIDIA: FP8 on one node with 4×B300 (TP4, TEP4, or DEP4), FP8 on one node with 8×H200, or NVFP4 on 4×B300 with the example context lowered to 320,000. There is no official single-GPU or Apple Silicon path, and 27B active parameters do not mean 27B of weights is enough.
• Output is verbose. Developer feedback collected by Founder Park centers on the model producing more output than needed during task execution, with room to tighten single-step replies.
• RL fine-tuning is not ready to use yet. As of the October 15, 2026 check, the verl adaptation pull request (#8246) was open, draft, and unmerged. A working inference image does not imply the training dependencies are in place.

Self-hosting reference (as published)

From the deployment write-up checked on October 15, 2026. That write-up did not launch the server itself, and the RL numbers come from an unmerged draft PR. Treat these as starting points for validation, not as validated configurations.

# FP8, TP4 example (StepFun/vLLM recipe, trimmed)
vllm serve stepfun-ai/Step-5-Preview \
  --tensor-parallel-size 4 --max-model-len 1048576 \
  --max-num-batched-tokens 16384 --max-num-seqs 64 \
  --gpu-memory-utilization 0.85 --enable-prefix-caching \
  --spec-method mtp --spec-tokens 3 \
  --reasoning-parser step5 --enable-auto-tool-choice --tool-call-parser step5

If VRAM is short, lower --max-model-len or concurrency first. Also note that three MTP candidate tokens do not mean three times the throughput; the gain depends on acceptance rate and load.

DeepSeek V4.1 Flash

• There is a ceiling on very long, complex agent tasks. Throughput and unit cost are the optimization targets, and no mechanism aimed at very long trajectories is published. The longer the chain and the more multi-hop reasoning it needs inside one trajectory, the more this shows.
• Self-hosting adds operational work. DeepSeek now publishes V4.1 Flash weights under MIT; serving, capacity planning, and runtime validation remain your responsibility. API prices do not describe self-hosting costs.
• Old scores do not carry over. Benchmark scores from the earlier V4 Flash Preview do not transfer to V4.1 Flash, so do not reuse older numbers when evaluating it.

7. Which one should you pick?

If your loop is high volume, repetitive, and made of fairly independent steps, and you would rather not run any infrastructure, take DeepSeek V4.1 Flash. The cached input price and the off-peak discount make extra iterations cheap, and 214 tokens/s keeps the loop moving. Price and throughput are its strongest, most defensible parts.

If a single trajectory has to run for a long time, needs repeated self-correction inside the task, and you want the weights on your own machines next to your own data and systems, take Step 5 Preview. The deep network and the context self-summarization trained with RL exist for that shape of work. The price is the hardware bar, a non-Apache-2.0 license, and a higher cost per token.

If you need both, split the work rather than looking for one model that is optimal on cost and depth at the same time: send the high-volume cheap loops to V4.1 Flash and the few genuinely long, hard tasks to Step 5 Preview.

FAQ

Is Step 5 Preview Apache 2.0?

No. The weights use a custom StepFun Model License 1.0 and Hugging Face tags the repository as "other". Read the license text for the specific version before commercial use.

Can Step 5 Preview run on a single GPU or Apple Silicon?

There is no official recipe for either. The published paths are server-class NVIDIA configurations starting at four B300 or eight H200 GPUs. The 27B active parameter figure describes compute, not the memory needed to hold the weights.

Can I fine-tune Step 5 Preview with RL today?

Not from the mainstream path yet. The verl adaptation pull request was still open, draft, and unmerged as of October 15, 2026. A runnable inference image does not mean the training dependencies are ready.

Is DeepSeek V4.1 Flash available as open weights?

Yes. The official deepseek-ai/DeepSeek-V4.1-Flash repository on Hugging Face provides weights under the MIT License, checked October 17, 2026.

Do older DeepSeek V4 Flash Preview scores apply to V4.1 Flash?

No. The Preview benchmark scores are not transferred to V4.1 Flash, and the old API aliases now route to the newer model. Re-evaluate it rather than reusing older numbers.

Which is cheaper for a high-volume agent loop?

DeepSeek V4.1 Flash, by a wide margin once cache hits are involved. At published prices, $0.006 per 1M cached input tokens against Step 5 Preview's reported $1.00 per 1M input tokens is a difference of roughly two orders of magnitude on the input side.