DeepSeek's V4.1 Flash 1017 Boosts Agent Scores 150 Elo at No Extra Cost

DeepSeek has released V4.1 Flash 1017, a new checkpoint that improves benchmark performance while retaining the previous Flash pricing and API name. Requests to deepseek-flash now use the updated model, and older Flash aliases also resolve to it, so existing integrations require no request changes.

Artificial Analysis gives the checkpoint an Intelligence Index score of 45, six points above the previous release. That score ties GLM-5.3 for second place among the tested open-weight models and trails MiMo-V2.6-Pro by one point. The service measured 224.7 output tokens per second and an evaluation cost of about $0.26 per task.

Because DeepSeek repointed existing aliases, production applications may receive different answers without a deployment change. Teams should rerun regression, latency, output-length, and safety tests. Applications that require reproducible behavior should also check whether their provider offers an immutable checkpoint identifier.

Agent work drives the gain

Most of the composite improvement comes from agentic work, knowledge tasks, hallucination handling, and physics reasoning:

AA-Omniscience rewards correct answers, penalizes fabricated answers, and applies no penalty when a model declines to answer. Its scale runs from -100 to 100. The modest accuracy increase and much larger non-hallucination gain suggest that abstention accounts for a substantial share of the improvement. That behavior can limit error propagation in agent workflows, provided the application detects refusals and routes them to retrieval, another model, or human review.

Flat token prices, shorter claimed outputs

First-party pricing remains unchanged during peak hours, with input tokens at $0.30 per million and output tokens at $1.20 per million. Off-peak input and output rates are half the peak prices.

The cached-input rate represents a 98% discount from the peak input price. Applications with stable system prompts, repeated documents, or long agent histories can benefit, although cache eligibility and retention rules should be confirmed before estimating production costs.

DeepSeek says the checkpoint completes tasks faster and produces shorter answers than the September V4.1 Flash release. Artificial Analysis still classifies its output as verbose relative to peers, and its Intelligence Index evaluation consumed about 89,000 output tokens per task. Output limits and response budgets therefore remain important controls for latency and cost.

A large backbone activates a fraction

The checkpoint uses a mixture-of-experts architecture that selects a small subset of parameters for each token. This routing allows the model to draw on a large parameter pool without applying the full network during every inference step.

• Backbone: 552 billion parameters.
• Engram parameters: 196 billion additional parameters for the model’s memory system.
• Active during prefill: 8 billion parameters per token while processing the prompt.
• Active during decoding: 16 billion parameters per token while generating the response.
• Context window: up to 1 million tokens.
• Maximum output: up to 384,000 tokens.
• Inputs: text and images.

Provider implementations may expose smaller context, output, or multimodal limits. Developers should verify each host’s limits, quantization, structured-output support, and serving configuration before comparing latency or quality.

DeepSeek publishes the weights under the MIT license on Hugging Face. Hosted access is available from DeepSeek, Fireworks, Baseten, DeepInfra, Wafer, and Cloudflare. NVIDIA also lists the checkpoint in its NIM catalog for deployment through NVIDIA inference microservices.

Two rivals lead on score per dollar

Artificial Analysis places V4.1 Flash 1017 in a competitive cost range, though MiMo-V2.6-Pro and GPT-6.1 Sol medium post higher Intelligence Index scores at lower estimated task costs.

Among these nearby options, DeepSeek combines the highest listed score with an MIT license. MiMo offers a one-point advantage at roughly half the benchmark task cost, while GPT-6.1 Sol medium leads by three points and costs five cents less per task. Actual spending will depend on prompt size, generated output, cache utilization, and provider pricing.

Search and Fetch join the stack

DeepSeek also opened its internal retrieval services through Search and Fetch APIs. Each service costs $0.75 per 1,000 calls, and free access through selected agent tools is scheduled to end October 31.

A workflow that performs one search and one fetch for every request adds $1.50 per 1,000 completed workflows before model-token charges. First-party retrieval can simplify billing and integration for web-connected agents, although teams still need to evaluate source coverage, freshness, ranking quality, citation handling, and failure behavior.

Roll out the alias carefully

Existing Flash users can adopt the checkpoint without changing request code, but the automatic alias update calls for a normal model migration process:

• Rerun representative prompts and compare answer quality, refusals, tool calls, and structured outputs.
• Measure generated-token counts, time to first token, total latency, and cache-hit rates.
• Add explicit handling for abstentions so agents can retrieve more evidence or escalate the task.
• Confirm each provider’s context window, image support, output cap, and immutable-version options.
• Review the MIT license alongside any provider terms, model dependencies, and deployment requirements.

V4.1 Flash 1017 best fits teams that need open weights, permissive licensing, multimodal input, long context, and a stable API alias. MiMo-V2.6-Pro retains the benchmark advantage on score per dollar, while DeepSeek offers a stronger deployment package for organizations prioritizing self-hosting and broad provider access.