DeepSeek's Official Agent Integrations Guide: The [1m] Context Specifier and Model Mapping Table Every AI Gateway Must Adopt

DeepSeek has quietly expanded its official API documentation with a dedicated Agent Integrations section covering 15 tools — Claude Code, GitHub Copilot, GitHub Copilot CLI, Kilo Code, WorkBuddy/CodeBuddy, OpenCode, Oh My Pi, OpenClaw, AstrBot, Deep Code, Hermes, nanobot, Crush, Pi, Reasonix, and Langcli. For AI engineering teams, the surface-level story is ecosystem breadth. The routing story is something narrower and more urgent: an undocumented model specifier notation and an official model-mapping table that determines which DeepSeek tier your Claude requests actually land on.

What the official docs now specify

Two routing-critical facts are now spelled out for the first time in official DeepSeek API docs.

The [1m] context-length suffix. The recommended Claude Code configuration is:

export ANTHROPIC_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max

The [1m] suffix is a bracket notation that explicitly pins the 1M-token context window on the V4 Pro model. Without it, the API defaults to a shorter context tier. For agentic workloads — long codebase context, multi-file refactors, extended reasoning chains — the difference between default and [1m] is the difference between hitting context limits mid-session and completing the task.

The automatic model-mapping table. When Claude Code or the Claude Desktop app sends requests using Claude model names through the DeepSeek Anthropic API endpoint, DeepSeek automatically maps them:

This mapping is applied server-side at https://api.deepseek.com/anthropic. It means that any team routing Claude Code through an AI gateway to DeepSeek's Anthropic endpoint does not control model selection at the model-name level — the upstream mapping table overrides it.

Why this matters for AI engineering teams

Context cost is now a routing dimension. The [1m] suffix is not just a convenience — it is an opt-in to the higher-cost 1M-context tier. DeepSeek's pricing structure distinguishes context window sizes; the [1m] flag requests the full-context variant. Teams that route all requests with this flag active will see per-token costs reflect the 1M-context pricing even for short exchanges. The correct pattern is to pin [1m] only on the Opus-tier model (for orchestrator-level tasks) and route Haiku-equivalent requests to deepseek-v4-flash without the suffix — exactly the config in the official docs.

The Anthropic endpoint model mapping table breaks gateway-level model pinning. If your AI gateway is translating Claude model names and forwarding them to https://api.deepseek.com/anthropic, you do not get DeepSeek model name granularity — the mapping table applies. Teams that need to pin deepseek-v4-pro explicitly must use the OpenAI-compatible endpoint at https://api.deepseek.com and set the ANTHROPIC_MODEL env var to the explicit DeepSeek model ID rather than routing through the Anthropic format. These are two separate routing surfaces with different behavior.

Web search adds a second billing event. The DeepSeek API natively supports Claude Code's web search tool. When a request triggers web search, the API issues a second LLM call to summarize retrieved content. This means web-search-capable agentic sessions generate two billing events per tool invocation. Teams building cost dashboards or per-session attribution need to account for this hidden second call in their usage accounting.

The ecosystem now includes 15 officially documented integrations. DeepSeek has moved beyond Claude Code and OpenCode to a much broader agent catalog. Tools like Kilo Code, Oh My Pi, AstrBot, Hermes, and OpenClaw all have dedicated per-tool docs. Each has its own model configuration pattern. Teams operating a gateway that fronts multiple agent tools should treat this as a signal that DeepSeek intends its Anthropic-compatible endpoint to be a primary backend for agent workloads — and plan their routing policies accordingly.

The router/operator angle

For teams running an AI gateway in front of DeepSeek's Anthropic-compatible endpoint, the key decision is whether to route at the gateway level or delegate model selection to the environment variable layer.

Gateway-level routing to DeepSeek: If your gateway translates Claude model names and forwards to the DeepSeek Anthropic endpoint, the server-side mapping table controls final model assignment. You lose the ability to distinguish V4 Pro vs V4 Flash at the gateway — the mapping is determined by the prefix of the requested Claude model name (opus → pro, sonnet/haiku → flash). This is fine if that coarse split aligns with your cost and quality targets.

Env-var-level pinning (bypassing gateway abstraction): The official config explicitly sets ANTHROPIC_MODEL=deepseek-v4-pro[1m], which overrides what Claude Code would normally select from the mapping table. This means the env var layer takes precedence over the Anthropic-format model name for the primary model slot — but sub-agent calls (via CLAUDE_CODE_SUBAGENT_MODEL) still route to flash separately.

Recommended routing policy for mixed-agent teams:

1. Set the orchestrator model to deepseek-v4-pro[1m] for Claude Code and any Opus-tier tool
2. Route Sonnet/Haiku-tier sub-agent calls to deepseek-v4-flash without the [1m] suffix
3. Account for web search second-call billing in cost attribution
4. Review each of the 15 official integration guides if you front multiple agent tools — each may have different model config patterns

What TheRouter users should watch

DeepSeek's Anthropic-compatible endpoint behavior — including the model mapping table and [1m] specifier — is relevant to any team routing Claude Code or other Anthropic-SDK tools through TheRouter to DeepSeek as an upstream. Verify that your provider routing configuration correctly handles the Anthropic endpoint format and that your cost accounting captures both primary and web-search-triggered billing events. The provider setup docs cover the base URL and auth configuration; the model-specifier details above are the operator-level config that must sit above the gateway layer.