Incorrect token count and pricing calculation for DeepSeek models #27458

Problem

When using DeepSeek models (deepseek-v4-flash / deepseek-v4-pro) via OpenCode, both token counting and cost estimation are inaccurate.

Root Cause Analysis

1. Tokenizer mismatch — OpenCode uses a different tokenizer (likely tiktoken for OpenAI) to count tokens locally, which produces different counts than DeepSeek's official byte-level BPE tokenizer. The DeepSeek API response includes usage.prompt_tokens / usage.completion_tokens in its own encoding, but OpenCode displays its own locally-computed counts instead of (or in addition to) the API-reported values.

2. Stale hardcoded model info — The context window, pricing per token (input/output), and tokenizer selection are compiled into the binary. For DeepSeek V4:

   • Context window shows 262K instead of the correct 1M (related: issue System prompt reports wrong context window for DeepSeek-V4-Pro (262K instead of 1M) #24119)
   • Input/output pricing per token uses outdated rates
   • The deepseek-chat and deepseek-reasoner model names are being deprecated on 2026/07/24 but may still be listed

Expected Behavior

• Token counts should match what DeepSeek's API reports (or at minimum use the correct tokenizer)
• Pricing should reflect current DeepSeek API pricing: https://api-docs.deepseek.com/quick_start/pricing
• Context window should be 1M for V4 models (not 262K)

Environment

• OpenCode version: 1.14.50
• Model(s): opencode/deepseek-v4-flash-free, deepseek-v4-pro
• Provider: DeepSeek (via OpenCode Zen / direct API)
• OS: Windows

Suggested Fix

• Update the internal model registry for DeepSeek V4 models with correct tokenizer, context window (1M), and pricing
• Consider using tiktoken with the cl100k_base encoding as a closer approximation, or better yet, respect the usage field returned by DeepSeek's API response
• Add a configuration override mechanism so users can supply custom pricing/tokenizer info for provider models (future-proofing)