DeepSeek V3.1 (open weights) folded reasoning back into the base model: one checkpoint, two modes, switched by the chat template rather than by swapping endpoints. On DeepInfra it runs $0.25 per Mtok input and $0.95 per Mtok output with a 128K tokens context, making it the cheapest way to get a togglable-thinking open model into Atlas. The catch is that thinking traces are billed as output tokens.
What does togglable thinking mean in DeepSeek V3.1?
DeepSeek V3.1 (open weights) has hybrid thinking and non-thinking modes in a single 671B/37B-active checkpoint, selected via the `thinking` flag in the chat template. One model id covers both, so you do not maintain two model configs inside Atlas.
DeepSeek V3.1 (open weights) collapsed a configuration problem. Before V3.1, running a reasoning model and a fast model meant two endpoints, two model ids, and two sets of pricing to track. V3.1 has hybrid thinking and non-thinking modes in a single 671B/37B-active checkpoint, selected via the `thinking` flag in the chat template, so you do not maintain two model configs. Inside Atlas that maps cleanly onto the agent split. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which means the phase where you want deliberation and the phase where you want speed are already separated by the tool. One checkpoint, two behaviors, one line in atlas.json.
How much does DeepSeek V3.1 cost on DeepInfra?
DeepSeek V3.1 (open weights) costs $0.25 per Mtok input and $0.95 per Mtok output on DeepInfra hosting in 2026. Output priced at $0.95/Mtok is under a tenth of most frontier reasoning models, so long agent traces with visible chain of thought stay affordable.
Pricing on DeepSeek V3.1 (open weights) is $0.25 per Mtok input and $0.95 per Mtok output through DeepInfra hosting. The first-party DeepSeek API only exposes the current V4 line, so DeepInfra is the route. What matters about $0.95 per Mtok output is what it enables: it is under a tenth of most frontier reasoning models, so long agent traces with visible chain of thought stay affordable. In Atlas, an agent run that plans, edits, verifies, and re-plans generates a lot of output tokens, and reasoning multiplies that. At a tenth of frontier output pricing, you can leave thinking on for the phases that need it instead of rationing it.
What does thinking mode actually cost you in an Atlas agent loop?
DeepSeek V3.1's reasoning traces cost output tokens, so an unbounded agent loop in thinking mode can spend more than the sticker price suggests. At $0.95 per Mtok output, the rate is low, but a verbose chain of thought repeated across every turn multiplies the token count, not the rate.
The cheapest reasoning model in the world still bills you for reasoning. On DeepSeek V3.1 (open weights), the reasoning traces cost output tokens, which means an Atlas agent loop left in thinking mode across dozens of tool-call round trips will run up a bill that the $0.95 per Mtok headline does not intuitively predict. The fix is structural rather than financial. Drive it with Atlas's plan agent first, then the build agent, so the thinking mode is spent on the plan and the mechanical edits run cheap. Deliberate once about the architecture of the change, then execute it without deliberating about each rename. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so you retain the review gate even in the fast phase.
Does DeepSeek V3.1's 128K context limit Atlas on a big repo?
DeepSeek V3.1 (open weights) has a 128K context, so Atlas's tree-sitter AST chunking and RRF code search matter more here than they do on the 1M-context V4 models. Retrieval quality, not window size, is what makes V3.1 workable on a large repository.
128K tokens is a real constraint on DeepSeek V3.1 (open weights), and the way you live with it is retrieval discipline. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so each chunk that lands in the prompt is a whole function or type rather than an arbitrary slice. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so the ranking blends meaning and exact symbol matches. Together those mean the 128K window holds a much higher density of relevant code than a naive grep-and-paste would produce. On the 1M-context V4 models you can be careless. On V3.1 you cannot, which is exactly why Atlas's index design matters more here.
Why does the MIT license on DeepSeek V3.1 matter?
The MIT license on the deepseek-ai/DeepSeek-V3.1 weights permits commercial use and private self-hosting with no per-seat licensing. For teams running Atlas across an engineering org in 2026, that removes the per-developer cost that hosted frontier models impose.
The MIT license on the deepseek-ai/DeepSeek-V3.1 weights is a commercial fact, not a legal footnote. It permits commercial use and private self-hosting with no per-seat licensing, which means the same togglable-thinking model can serve an entire team from your own infrastructure once the volume justifies the hardware. Atlas pairs with that well: Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so a fully self-hosted Atlas plus V3.1 setup keeps both the index and the inference in your own environment. Until you cross that threshold, DeepInfra hosting at $0.25 per Mtok input and $0.95 per Mtok output is the cheap default, and the exit path stays open.
Setup
- 01Export DEEPINFRA_API_KEY, or run `atlas login` and choose DeepInfra.
- 02Confirm the id with `atlas models deepinfra | grep V3.1`.
- 03Pin `"model": "deepinfra/deepseek-ai/DeepSeek-V3.1"` in atlas.json.
- 04Drive it with Atlas's plan agent first, then the build agent, so the thinking mode is spent on the plan and the mechanical edits run cheap.
- 05Keep retrieval tight, since the 128K context means Atlas's tree-sitter AST chunking and RRF code search do more of the work than on a 1M-context model.
Frequently asked questions
- How much does DeepSeek V3.1 cost per million tokens?
- On DeepInfra hosting, DeepSeek V3.1 (open weights) is $0.25 per Mtok input and $0.95 per Mtok output. The first-party DeepSeek API only exposes the current V4 line.
- How do I turn on thinking mode in DeepSeek V3.1?
- DeepSeek V3.1 selects between hybrid thinking and non-thinking modes via the `thinking` flag in the chat template. It is one checkpoint with two modes, so you do not maintain two model configs.
- How do I run Atlas on DeepSeek V3.1?
- Export DEEPINFRA_API_KEY or run `atlas login` and choose DeepInfra, confirm the id with `atlas models deepinfra | grep V3.1`, then pin `"model": "deepinfra/deepseek-ai/DeepSeek-V3.1"` in atlas.json.
- Is DeepSeek V3.1 free to self-host commercially?
- The MIT license on the deepseek-ai/DeepSeek-V3.1 weights permits commercial use and private self-hosting with no per-seat licensing. You still pay for the hardware to serve a 671B/37B-active model.
- Why is my DeepSeek V3.1 bill higher than expected?
- Reasoning traces cost output tokens on DeepSeek V3.1, so an unbounded agent loop in thinking mode can spend more than the $0.95 per Mtok sticker price suggests. Spend thinking on the plan and run the edits without it.
- What is DeepSeek V3.1's context window?
- DeepSeek V3.1 (open weights) has a 128K context. That is why Atlas's tree-sitter AST chunking and reciprocal rank fusion code search matter more here than on the 1M-context V4 models.
- Should I use DeepSeek V3.1's plan agent or build agent mode?
- Both, in sequence. Drive DeepSeek V3.1 with Atlas's plan agent first so the thinking mode is spent on the plan, then hand off to the build agent so the mechanical edits run cheap.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Gin in 2026
Atlas is a terminal-native AI coding agent for Gin in 2026. It reads router groups and binding tags, then runs go test ./... -race behind a permission prompt.
Atlas for Nim: A Terminal-Native AI Coding Agent for Nimble Packages and Macros in 2026
Atlas is a terminal-native AI coding agent for Nim in 2026. It reads .nimble requires and asterisk-exported symbols, adds std/unittest suites, runs nimble test, formats with nph.
Atlas for Nuxt: Auto-Imports, useAsyncData, and Nitro Handlers in 2026
Atlas is a terminal-native AI coding agent for Nuxt in 2026. It reads nuxt.config.ts, pages/ routes, composables/ auto-imports, and server/api/ Nitro handlers, and tests with @nuxt/test-utils.
Atlas vs Zed: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Zed, the GPU-accelerated collaborative editor, for developers in 2026. Evaluate their AI autonomy, privacy, and workflow.
Atlas vs Windsurf: Terminal AI Coding Agents in 2026
Atlas and Windsurf offer distinct AI coding experiences in 2026. Atlas provides a terminal-native TUI with permission-gated tools, while Windsurf features an in-editor live preview and a dedicated GUI.
Atlas vs Amazon Q Developer: Terminal AI Coding Agents in 2026
Comparing Atlas and Amazon Q Developer in 2026. Atlas offers terminal-native AI with permission-gated tools and local code indexing. Amazon Q Developer excels in AWS services, providing Java upgrades and security
Atlas vs Mistral Vibe for Code: Terminal AI Coding Agents in 2026
Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, explicit diffs, and BYO models. Mistral Vibe for Code provides a four-model stack, multi-platform access, and EU data sovereignty.
Atlas vs Gemini CLI: A Developer's Guide to Terminal AI Agents in 2026
Atlas vs Gemini CLI in 2026: Compare terminal AI coding agents. Atlas offers permission-gated tool calls and diff-based approvals. Gemini CLI provides a million-plus token context and a free tier.